A data center refrigeration and energy saving method and system

Abstract

The application relates to the technical field of intelligent regulation and control of data center refrigeration energy saving, and discloses a data center refrigeration and energy saving method and system; the method comprises the following steps: constructing a three-dimensional coordinate system of a machine room and marking three-dimensional coordinates and the air supply direction of each precision air conditioner; based on the three-dimensional coordinates and the air supply direction, the spatial directed distance from each IT equipment or temperature sensor in a cold aisle to the precision air conditioner is calculated through a direction correction factor; a reference operation parameter library is constructed, and multi-source operation condition data are acquired; a light-weighted weighted fusion decision model is constructed based on the spatial directed distance, the reference operation parameter library and the multi-source operation condition data, and a regulation and control coefficient and an operation temperature setting value of each precision air conditioner are generated; each precision air conditioner is regulated and controlled based on the regulation and control coefficient and the operation temperature setting value; and the application realizes long-term accurate, efficient, safe and reliable collaborative optimization of refrigeration regulation and control under complex conditions.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for data center cooling and energy saving, specifically to a data center cooling and energy-saving method and system. Background Technology

[0002] With the rapid development of cloud computing, big data, and artificial intelligence technologies, data centers, as the core infrastructure of the digital economy, are experiencing continuous growth in scale and energy consumption. Cooling systems, as a major component of data center energy consumption, directly impact the power usage efficiency (PUE) and operating costs of data centers.

[0003] Currently, most mainstream solutions for data center cooling control operate in a fixed temperature threshold mode, where precision air conditioners start, stop, or adjust their frequency based on manually preset upper and lower limits for return or supply air temperatures. While this mode is simple to implement and highly reliable, it struggles to adapt to complex operating conditions such as dynamic changes in IT load, seasonal variations, and localized hotspots in the data center, commonly resulting in problems like overcooling, low energy efficiency, and lag. Therefore, achieving refined and intelligent control of data center cooling systems to minimize cooling energy consumption while ensuring the safe operation of IT equipment has become a critical technical challenge that the industry urgently needs to address.

[0004] For example, Chinese patent CN120129220A discloses an energy consumption optimization method for a cooling system in a computing center. It uses a pre-trained cooling control decision model for regulation decisions. However, the parameters of this model are usually static and lack an online adaptive update mechanism. When the data center faces seasonal changes, load structure changes, or equipment aging, the model performance will significantly decline, requiring frequent manual retraining and making it difficult to achieve long-term stable prediction and control. For example, patent CN109800066B discloses a data center energy-saving scheduling method and system. It uses a prediction model to evaluate the total power consumption and air conditioning parameters after deploying tasks. However, such models usually treat the data center as a whole system and often use a simple splicing method when integrating multi-source heterogeneous data, which cannot characterize the key influence of different data sources on energy consumption prediction.

[0005] The existing technologies mentioned above all suffer from the problems proposed in this background: the decision model parameters are statically fixed and the fusion of multi-source data is coarse, making it difficult for cooling control to achieve long-term accurate, efficient, energy-saving and safe and reliable synergistic optimization under complex dynamic operating conditions.

[0006] The information disclosed in this background section is intended only to enhance the understanding of the overall background of this application and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0007] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a data center cooling and energy-saving method and system, which solves the problems mentioned in the background technology to a certain extent and ensures that cooling control achieves long-term accurate, efficient energy saving and safe and reliable synergistic optimization under complex dynamic operating conditions.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] In a first aspect, the present invention provides a data center cooling and energy-saving method, comprising the following steps:

[0010] Construct a three-dimensional coordinate system for the computer room and mark the three-dimensional coordinates of each cabinet, IT equipment, precision air conditioner and temperature sensor, as well as the air supply direction of each precision air conditioner;

[0011] Based on the three-dimensional coordinates and air supply direction, a spatial directional distance calculation model including a direction correction factor is used to calculate the spatial directional distance from each IT device or temperature sensor in the cold aisle to the precision air conditioner.

[0012] A baseline operating parameter library is constructed based on season, day / night, and time period, and multi-source operating condition data from the data center is acquired in real time; a lightweight weighted fusion decision model is constructed based on the spatial directional distance, the baseline operating parameter library, and the multi-source operating condition data.

[0013] The lightweight weighted fusion decision model generates the control coefficient and operating temperature setpoint for each precision air conditioner, and controls each precision air conditioner based on the control coefficient and operating temperature setpoint.

[0014] As a preferred embodiment of the data center cooling and energy-saving method described in this application, the construction method of the three-dimensional coordinate system of the computer room is as follows:

[0015] Using a corner of the computer room as the origin of the coordinate system, the horizontal arrangement of the equipment is defined as the X-axis, the vertical arrangement as the Y-axis, and the vertical height as the Z-axis.

[0016] Based on the CAD drawings of each server room in the data center and the size data of its internal equipment, determine the three-dimensional coordinates of each rack center, IT equipment air inlet, precision air conditioner air outlet, cold aisle temperature sensor and hot aisle temperature sensor in the coordinate system.

[0017] The airflow direction of each precision air conditioner outlet is represented as a unit direction vector in the three-dimensional coordinate system of the computer room.

[0018] As a preferred embodiment of the data center cooling and energy-saving method described in this application, the calculation steps for calculating the spatial directed distance from each IT device or temperature sensor in the cold aisle to the precision air conditioner using a spatial directed distance calculation model that includes a direction correction factor include:

[0019] Based on the three-dimensional coordinates of the precision air conditioner outlet and the three-dimensional coordinates of the IT equipment or the temperature sensor in the cold aisle, calculate the direction vector from the precision air conditioner outlet to the IT equipment or the temperature sensor in the cold aisle, and calculate the three-dimensional Euclidean straight-line distance between the precision air conditioner outlet and the IT equipment or the temperature sensor in the cold aisle.

[0020] Calculate the angle between the air supply direction of the precision air conditioner outlet and the direction vector of the corresponding precision air conditioner outlet pointing to the IT equipment or the temperature sensor in the cold aisle, and calculate the direction correction factor based on the angle.

[0021] The product of the three-dimensional Euclidean straight-line distance and the corresponding direction correction factor is calculated to obtain the spatial directional distance from the corresponding IT equipment or temperature sensor in the cold aisle to the corresponding precision air conditioner.

[0022] As a preferred embodiment of the data center cooling and energy-saving method described in this application, the direction correction factor is: calculating the difference between 1 and the cosine of the included angle, multiplying the difference by 2, and adding the resulting product to 1 to obtain the direction correction factor;

[0023] The direction vector of the precision air conditioner outlet pointing towards the IT equipment or the temperature sensor in the cold aisle is calculated as follows: the three-dimensional coordinates of the precision air conditioner outlet are defined as follows: The three-dimensional coordinates of temperature sensors in IT equipment or cold aisles are The airflow direction of the precision air conditioner's air outlet is The direction vector is calculated as follows: ;

[0024] in, , , These represent the X-axis, Y-axis, and Z-axis coordinates of the precision air conditioner's air outlet. , , These are the X-axis, Y-axis, and Z-axis coordinates of temperature sensors in IT equipment or cold aisles, respectively. , , The components of the air supply direction of the precision air conditioner outlet on the X-axis, Y-axis, and Z-axis are given.

[0025] As a preferred embodiment of the data center cooling and energy-saving method described in this application, the baseline operating parameter library is used to provide temperature setting baselines and safety thresholds under different operating conditions, and its construction method is as follows:

[0026] The index is built using season, day and night, and time of day as dimensions; the values ​​for the season dimension include {spring, summer, autumn, winter}, where spring is March to May, summer is June to August, autumn is September to November, and winter is December to February;

[0027] The day-night dimension includes {day, night}, where day is from 8:00 to 20:00 and night is from 20:00 to 8:00.

[0028] The time period dimension includes {valley, plateau, peak}, where valley is from 0 to 8, plateau is from 8 to 10 and 17 to 24, and peak is from 10 to 17;

[0029] Each index record corresponds to a set of baseline operating parameters; the baseline operating parameters include the precision air conditioning operating temperature setpoint, the upper limit of the cold aisle temperature, the lower limit of the cold aisle temperature, the upper limit of the hot aisle temperature, and the lower limit of the hot aisle temperature.

[0030] As a preferred embodiment of the data center cooling and energy-saving method described in this application, the lightweight weighted fusion decision model is constructed as follows:

[0031] Using the three-dimensional coordinates of each device in the three-dimensional coordinate system of the computer room, the benchmark operating parameter library and multi-source operating condition data as input data, the key features of the input data are extracted through four dedicated branches: spatial correlation feature branch, benchmark strategy feature branch, operating condition feature branch and energy consumption feature branch. The fusion weight of each dedicated branch under the current operating condition is dynamically calculated through the gating network to obtain the global fusion feature vector.

[0032] Based on the global fusion feature vector, the control formula for the temperature of the precision air conditioner within the control cycle, the operating temperature setpoint for each precision air conditioner, and the PUE prediction value of the data center are generated through the multi-task output layer.

[0033] As a preferred embodiment of the data center cooling and energy-saving method described in this application, the spatial correlation feature branch, the baseline strategy feature branch, the operating condition feature branch, and the energy consumption feature branch each contain a fully connected layer, which respectively maps their respective input data into a 64-dimensional continuous feature vector.

[0034] The spatial association feature branch is based on the three-dimensional coordinate data of each device in the three-dimensional coordinate system of the computer room. It calculates the spatial directed distance between each IT device and cold aisle temperature sensor in the computer room and each precision air conditioner in the computer room, and encodes the spatial semantics of the spatial directed distance into a continuous feature vector. At the same time, it establishes a spatial distance lookup table from near to far.

[0035] As a preferred embodiment of the data center cooling and energy-saving method described in this application, wherein: the baseline strategy feature branch encodes the baseline operating parameter library into a continuous feature vector;

[0036] The operating condition feature branch encodes the precision air conditioner return air temperature, cold aisle temperature, hot aisle temperature, outdoor ambient temperature, weather forecast temperature, IT equipment inlet air temperature, IT equipment outlet air temperature, and IT equipment CPU temperature from the multi-source operating condition data into continuous feature vectors.

[0037] The energy consumption feature branch encodes IT equipment energy consumption data, refrigeration equipment energy consumption data, and other energy consumption data into continuous feature vectors.

[0038] As a preferred embodiment of the data center cooling and energy-saving method described in this application, the lightweight weighted fusion decision model outputs the global fusion feature vector through the gating network in the following manner:

[0039] The continuous feature vectors output by the four dedicated branches are concatenated and fused to obtain the global context vector. The global context vector is then input into the gating network to solve for the fusion weights, resulting in the gating vector g=[g1,g2,g3,g4].

[0040] The sum of the weight coefficients g1, g2, g3 and g4 is 1, and g1, g2, g3 and g4 correspond to the real-time fusion weights of the spatial correlation feature branch, the baseline strategy feature branch, the operating condition feature branch and the energy consumption feature branch, respectively.

[0041] The continuous feature vectors output by the four dedicated branches are multiplied by the corresponding real-time fusion weights in the gating vector, and then summed element by element to obtain the global fusion feature vector.

[0042] As a preferred embodiment of the data center cooling and energy-saving method described in this application, the multi-task output layer includes a control formula output head, a precision air conditioner operating temperature setpoint output head, and a PUE prediction value output head.

[0043] The inputs to the output head of the control formula include the global fusion feature vector, the three-dimensional coordinates of the precision air conditioner, the spatial correlation feature vector corresponding to the precision air conditioner, and the temperature anomaly identification vector. The output is the control coefficient of the corresponding precision air conditioner.

[0044] The precision air conditioner's operating temperature setpoint output head uses the control coefficient generated by the control formula output head as weight to perform a weighted summation of the corresponding real-time temperature and the precision air conditioner's operating temperature setpoint in the multi-source operating condition data and the benchmark operating parameters, thereby obtaining the corresponding precision air conditioner's operating temperature setpoint.

[0045] Secondly, this application provides a data center cooling and energy-saving system, including a three-dimensional modeling module, a benchmark library construction module, a model building module, and a control execution module;

[0046] The 3D modeling module is used to construct a 3D coordinate system for the computer room and to label the 3D coordinates of each cabinet, IT equipment, precision air conditioner and temperature sensor, as well as the air supply direction of each precision air conditioner.

[0047] The 3D modeling module also uses a spatial directional distance calculation model that includes a direction correction factor to calculate the spatial directional distance from each IT device or temperature sensor in the cold aisle to the precision air conditioner based on the 3D coordinates and air supply direction.

[0048] The benchmark library construction module builds a benchmark operating parameter library based on season, day and night, and time period, and acquires multi-source operating condition data from the data center in real time;

[0049] The model building module constructs a lightweight weighted fusion decision model based on the spatial directional distance, the benchmark operating parameter library, and multi-source operating condition data.

[0050] The control execution module generates control coefficients and operating temperature setpoints for each precision air conditioner through the lightweight weighted fusion decision model, and controls each precision air conditioner based on the control coefficients and operating temperature setpoints.

[0051] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0052] By constructing a three-dimensional coordinate system for the computer room and marking the three-dimensional coordinates of each cabinet, IT equipment, precision air conditioner, and temperature sensor, as well as the airflow direction of each precision air conditioner, a standardized foundation is provided for subsequent spatial directed distance calculations. Based on the three-dimensional coordinates and airflow direction, a direction correction factor is introduced to calculate the spatial directed distance from each IT device or temperature sensor in the cold aisle to the precision air conditioner. This solves the problem that traditional spatial distance calculations do not consider the azimuth angle between the airflow direction of the precision air conditioner and the target point, failing to accurately reflect the actual delivery efficiency of cold air and easily leading to misjudgments such as "short straight-line distance but opposite airflow direction and extremely low cooling efficiency," resulting in misaligned hotspot response and ineffective cooling capacity delivery. Furthermore, by constructing a benchmark operating parameter library, data... Based on the mature experience of precision air conditioning control in the center, a base library of benchmark parameters and safety threshold rules categorized by season, day / night, and time period has been established. This provides a stable policy prior and safety baseline for the subsequent lightweight weighted fusion decision model. By collecting multi-source operating condition data from the data center, comprehensive perception from chip thermal status to global energy efficiency is achieved, providing comprehensive and multi-granular input support for the lightweight weighted fusion decision model. By extracting four types of features—spatial correlation, benchmark strategy, operating condition, and energy consumption—and dynamically weighting and fusing them through a gating network, the model adapts to the decision-making focus under different operating conditions. Under the premise of ensuring safety, it achieves precise, efficient, and adaptive cooling control, thereby reducing energy consumption and minimizing manual intervention. Attached Figure Description

[0053] Figure 1 A flowchart of a data center cooling and energy-saving method provided in this application;

[0054] Figure 2 Architecture diagram of the lightweight weighted fusion decision model provided in this application;

[0055] Figure 3 A framework diagram of a data center cooling and energy-saving system provided in this application. Detailed Implementation

[0056] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof.

[0057] Example 1

[0058] like Figure 1 As shown in the figure, this embodiment introduces a data center cooling and energy-saving method, including the following steps:

[0059] A three-dimensional coordinate system for the data center is constructed, and the three-dimensional coordinates of each cabinet, IT equipment, precision air conditioner, and temperature sensor, as well as the airflow direction of each precision air conditioner, are marked. The construction method of the three-dimensional coordinate system is as follows: taking one corner of the data center as the origin, the horizontal arrangement direction of the equipment is defined as the X-axis, the vertical arrangement direction is defined as the Y-axis, and the vertical height direction is defined as the Z-axis. Based on the CAD drawings of each data center and the size data of its internal equipment, the three-dimensional coordinates of the center of each cabinet, the air inlet of the IT equipment, the air outlet of the precision air conditioner, the cold aisle temperature sensor, and the hot aisle temperature sensor in the coordinate system are determined. The airflow direction of each precision air conditioner outlet is represented as a unit direction vector in the three-dimensional coordinate system of the data center; for example, an airflow direction of (1,0,0) indicates airflow along the X-axis.

[0060] Based on the aforementioned three-dimensional coordinates and airflow direction, a spatial directional distance calculation model incorporating a direction correction factor is used to calculate the spatial directional distance from each IT device or temperature sensor within the cold aisle to the precision air conditioner. Specifically, the spatial directional distance is calculated as follows:

[0061] Define the three-dimensional coordinates of the precision air conditioner's air outlet as follows: The three-dimensional coordinates of temperature sensors in IT equipment or cold aisles are The airflow direction of the precision air conditioner's air outlet is ;in, The X-axis coordinate of the precision air conditioner's air outlet. The Y-axis coordinate of the precision air conditioner's air outlet. The Z-axis coordinate of the precision air conditioner's air outlet; The X-axis coordinate of the temperature sensor in the IT equipment or cold aisle. For the Y-axis coordinate of temperature sensors in IT equipment or cold aisles, Z-axis coordinates for temperature sensors in IT equipment or cold aisles; This represents the component of the airflow direction from the precision air conditioner's outlet on the X-axis. This represents the Y-axis component of the airflow direction from the precision air conditioner's outlet. This represents the Z-axis component of the airflow direction from the precision air conditioner's outlet.

[0062] Calculate the direction vector of the precision air conditioner's air outlet pointing towards IT equipment or temperature sensors in the cold aisle. It also calculates the three-dimensional Euclidean linear distance between the precision air conditioner outlet and the temperature sensor in the IT equipment or cold aisle. In two scenarios—where the precision air conditioner's air outlet is directly facing IT equipment or a temperature sensor in the cold aisle, or where it is facing away from IT equipment or a temperature sensor in the cold aisle—even with the same Euclidean distance, the cooling effect will be drastically different. Therefore, a direction correction factor is introduced to correct for the three-dimensional Euclidean distance. This direction correction factor is a coefficient calculated based on the airflow direction of the precision air conditioner's air outlet and the direction vector pointing from the air outlet to the IT equipment or temperature sensor in the cold aisle. The calculation method is as follows: calculate the angle between the airflow direction of the precision air conditioner's air outlet and the corresponding direction vector pointing from the air outlet to the IT equipment or temperature sensor in the cold aisle. And based on the included angle The direction correction factor is calculated as follows: ;in Angle The cosine value.

[0063] For any IT device or cold aisle temperature sensor, calculate the three-dimensional Euclidean straight-line distance between it and each precision air conditioner outlet, and calculate the spatial directional distance between each IT device or cold aisle temperature sensor and each precision air conditioner according to the corresponding direction correction factor; the spatial directional distance is the product of each three-dimensional Euclidean straight-line distance and the corresponding direction correction factor.

[0064] A baseline operating parameter library is constructed using season, day / night cycle, and time period as dimensions, and multi-source operating condition data from the data center is acquired in real time. This baseline operating parameter library provides temperature setting baselines and safety thresholds under different operating conditions. Its construction method is as follows: an index is created using season, day / night cycle, and time period as dimensions. The values ​​for the season dimension include {spring, summer, autumn, winter}, where spring is March-May, summer is June-August, autumn is September-November, and winter is December-February. The values ​​for the day / night cycle dimension include {daytime, nighttime}, where daytime is 8-20 hours and nighttime is 20-8 hours. The values ​​for the time period dimension include {valley, plateau, peak}, where valley is 0-8 hours, plateau is 8-10 hours and 17-24 hours, and peak is 10-17 hours. Each index record corresponds to a set of baseline operating parameters. The baseline operating parameters include the precision air conditioning operating temperature setpoint, cold aisle temperature upper limit, cold aisle temperature lower limit, hot aisle temperature upper limit, and hot aisle temperature lower limit. The precision air conditioning operating temperature setpoint is the default set temperature of the precision air conditioning under the corresponding operating conditions, in °C, and is manually entered by maintenance personnel based on historical maintenance experience of the data center or design manuals. The cold aisle temperature upper limit is the maximum allowable temperature of the cold aisle, in °C, and the cold aisle temperature lower limit is the minimum allowable temperature of the cold aisle, in °C. The hot aisle temperature upper limit is the maximum allowable temperature of the hot aisle, in °C, and the hot aisle temperature lower limit is the minimum allowable temperature of the hot aisle, in °C. All of these are manually configured and entered by data center maintenance personnel in the system according to relevant national standards, industry design specifications (such as the "Data Center Design Specification" GB 50174), and the operating temperature and humidity range recommended by IT equipment manufacturers.

[0065] The multi-source operating condition data includes precision air conditioning return air temperature, cold aisle temperature, hot aisle temperature, outdoor ambient temperature, weather forecast temperature, IT equipment inlet air temperature, IT equipment outlet air temperature, IT equipment CPU temperature, IT equipment energy consumption data, cooling equipment energy consumption data, and other energy consumption data. The precision air conditioning return air temperature is obtained from the precision air conditioning communication interface via the Modbus protocol. The cold aisle and hot aisle temperatures are obtained from the temperature sensor communication interface in the corresponding aisle via the Modbus protocol. The outdoor ambient temperature is obtained from the outdoor temperature sensor communication interface via the Modbus protocol. The weather forecast temperature is obtained from the open interface provided by a third-party weather forecast service provider via the HTTP protocol. The IT equipment inlet air temperature, outlet air temperature, and CPU temperature are obtained from the IT equipment's BMC via the IPMI or SNMP protocol. The IT equipment energy consumption data, cooling equipment energy consumption data, and other energy consumption data are obtained from the smart meter communication interface in the data center power distribution cabinet via the Modbus protocol.

[0066] A lightweight weighted fusion decision model is constructed based on the spatial directional distance, the benchmark operating parameter library, and multi-source operating condition data. The control coefficient and operating temperature setpoint of each precision air conditioner are generated through the lightweight weighted fusion decision model, and the control coefficient and operating temperature setpoint are used to control each precision air conditioner.

[0067] like Figure 2 As shown, the lightweight weighted fusion decision model adopts a three-layer progressive architecture of "multi-branch extraction → gating fusion → multi-task output". It takes the three-dimensional coordinate data of each device in the three-dimensional coordinate system of the computer room, the benchmark operating parameter library and multi-source operating condition data as input data. Through four dedicated branches, namely spatial correlation feature branch, benchmark strategy feature branch, operating condition feature branch and energy consumption feature branch, the key features of the input data are extracted respectively. The fusion weight of each dedicated branch under the current operating condition is dynamically calculated through the gating network to achieve adaptive weighted fusion. Through the multi-task output layer, the control formula of the precision air conditioner temperature within the control cycle, the operating temperature setpoint of each precision air conditioner and the PUE prediction value of the data center are generated respectively to realize the dynamic control of data center cooling and energy saving.

[0068] Specifically, the spatial association feature branch calculates the spatial directed distance between each IT device and cold aisle temperature sensor in the computer room and each precision air conditioner in the computer room based on the three-dimensional coordinate data of each device in the three-dimensional coordinate system of the computer room. The spatial semantics of the spatial directed distance is encoded into a continuous feature vector, and a spatial distance lookup table is established from near to far. When the temperature of any IT device or cold aisle temperature sensor is abnormal, the lightweight weighted fusion decision model quickly locates the precision air conditioner with the smallest spatial directed distance to the corresponding IT device or cold aisle temperature sensor according to the spatial distance lookup table and performs dynamic control.

[0069] The abnormal temperature of the IT equipment or cold aisle temperature sensor refers to the corresponding temperature exceeding the corresponding safety threshold range in the baseline operating parameter library, specifically including any of the following situations: the temperature of the IT equipment or cold aisle temperature sensor is higher than the corresponding upper limit of the cold aisle temperature, the temperature of the IT equipment or cold aisle temperature sensor is lower than the corresponding lower limit of the cold aisle temperature, the inlet air temperature of the IT equipment, the outlet air temperature of the IT equipment, or the CPU temperature of the IT equipment exceeds the preset safe operating temperature range of the equipment.

[0070] In this embodiment, the safe operating temperature range of the equipment is set as follows: based on the recommended operating temperature range indicated in the equipment specification or technical parameter document provided by the IT equipment manufacturer, the upper limit values ​​of the IT equipment's inlet air temperature, outlet air temperature, and CPU temperature are set respectively; the lower limit value of the safe operating temperature range of the equipment is the minimum ambient temperature requirement for normal operation of the equipment, which is configured by the maintenance personnel according to the actual environmental conditions of the computer room.

[0071] The core of the benchmark strategy feature branch is to provide a benchmark reference for dynamic control of the lightweight weighted fusion decision model, that is, the default set temperature and safety threshold that the precision air conditioner should follow under different seasons, day and night, and time periods; the benchmark strategy feature branch encodes the benchmark operating parameter library into a continuous feature vector.

[0072] The core of the operating condition feature branch is to integrate data such as the return air temperature of the precision air conditioner, cold aisle temperature, hot aisle temperature in the computer room, outdoor ambient temperature, weather forecast temperature, and the air intake temperature, air outlet temperature, and CPU temperature of IT equipment to form a real-time picture of the current operating condition of the computer room. The operating condition feature branch encodes the precision air conditioner return air temperature, cold aisle temperature, hot aisle temperature, outdoor ambient temperature, weather forecast temperature, air intake temperature of IT equipment, air outlet temperature of IT equipment, and CPU temperature of IT equipment from the multi-source operating condition data into continuous feature vectors.

[0073] The core of the energy consumption feature branch is to provide energy consumption dimension information input for the energy-saving optimization decision of the lightweight weighted fusion decision model based on the energy consumption data of IT equipment, cooling equipment and other energy consumption data in the multi-source operating condition data; the energy consumption feature branch encodes the energy consumption data of IT equipment, cooling equipment and other energy consumption data into continuous feature vectors.

[0074] In the lightweight weighted fusion decision model, the spatial correlation feature branch, the baseline strategy feature branch, the operating condition feature branch, and the energy consumption feature branch each contain a fully connected layer, which maps their respective input data into a 64-dimensional continuous feature vector, thus completing the standardized extraction of single-branch features.

[0075] The lightweight weighted fusion decision model employs a dynamic gating fusion mechanism. The gating network dynamically determines the fusion weights of each specialized branch based on the current input data. This is achieved by concatenating and fusing the 64-dimensional continuous feature vectors output from the four specialized branches to obtain a 256-dimensional global context vector. This global context vector is then input into the gating network to solve for the fusion weights. The gating network consists of two fully connected layers. The first fully connected layer reduces the 256-dimensional global context vector to 32 dimensions and performs a nonlinear transformation using the ReLU activation function. The second fully connected layer... The network maps a 32-dimensional vector to 4 dimensions and normalizes it using the Softmax function to generate a gated vector g=[g1,g2,g3,g4]. The sum of the weight coefficients g1, g2, g3, and g4 is 1, and g1, g2, g3, and g4 correspond to the real-time fusion weights of the spatial correlation feature branch, the baseline strategy feature branch, the operating condition feature branch, and the energy consumption feature branch, respectively. The 64-dimensional continuous feature vectors output by the four dedicated branches are multiplied by the corresponding real-time fusion weights in the gated vector, and then summed element by element to obtain a 64-dimensional global fusion feature vector.

[0076] The lightweight weighted fusion decision model is equipped with a multi-task output layer at the end, which is connected to the output of the gating network. It includes a control formula output head, a precision air conditioner operating temperature setpoint output head, and a PUE prediction value output head.

[0077] The input to the output head of the control formula consists of four parts: the 64-dimensional global fusion feature vector, the three-dimensional coordinates of the precision air conditioner, the spatial correlation feature vector corresponding to the precision air conditioner, and a 4-dimensional temperature anomaly identification vector; the spatial correlation feature vector is the 64-dimensional continuous feature vector output by the spatial correlation feature branch; after concatenating the above inputs into a 135-dimensional comprehensive vector, it is sequentially processed through three layers of network to complete feature transformation and parameter output: first, the 135-dimensional comprehensive vector is mapped to 64 dimensions through a fully connected layer with ReLU activation function, then a Dropout layer is used to complete feature regularization, and finally a fully connected layer is used to map to 9-dimensional features, outputting the 9 control coefficients corresponding to the precision air conditioner.

[0078] Each dimension of the 4-dimensional temperature anomaly identification vector corresponds to a type of temperature anomaly, including: cold aisle temperature exceeding the upper limit, i.e., the cold aisle temperature sensor temperature is higher than the upper limit of the cold aisle temperature; cold aisle temperature below the lower limit, i.e., the cold aisle temperature sensor temperature is lower than the lower limit of the cold aisle temperature; IT equipment temperature anomaly, i.e., the IT equipment inlet air temperature, IT equipment outlet air temperature, or IT equipment CPU temperature exceeds the corresponding safe operating temperature range of the equipment; and sensor data anomaly, i.e., the temperature sensor data acquisition fails.

[0079] The precision air conditioner's operating temperature setpoint output head takes as input nine control coefficients generated by the control formula output head, the corresponding real-time temperature from the multi-source operating condition data, and the precision air conditioner's operating temperature setpoint from the benchmark operating parameters. It calculates and outputs the corresponding precision air conditioner's operating temperature setpoint through weighted summation. The real-time temperatures are the precision air conditioner's return air temperature, cold aisle temperature, hot aisle temperature, outdoor ambient temperature, weather forecast temperature, IT equipment inlet air temperature, IT equipment outlet air temperature, and IT equipment CPU temperature. The weight coefficients for each item are the nine control coefficients generated by the control formula output head.

[0080] The PUE prediction output head takes a 64-dimensional global fusion feature vector as input and completes feature transformation and prediction solution step by step: the 64-dimensional global fusion feature vector is reduced to 32 dimensions by a fully connected layer with ReLU activation function, and then overfitting is suppressed by a Dropout layer; a fully connected layer with Sigmoid activation function is used to map it to a 1-dimensional initial prediction value, and the 1-dimensional initial prediction value is linearly scaled to map the value to the interval [1.0, 3.0] covering the typical PUE range of the data center, thus obtaining the data center PUE prediction value.

[0081] In this embodiment, the training method of the lightweight weighted fusion decision model is as follows: Historical operational data of at least one complete operating cycle (covering all four seasons) of the data center is acquired; the historical operational data includes three-dimensional coordinate data of the data center equipment at each moment, baseline operating parameter indexes (season, day / night, time period), multi-source operating condition data (precision air conditioning return air temperature, cold aisle temperature, hot aisle temperature, outdoor ambient temperature, weather forecast temperature, IT equipment inlet air temperature, IT equipment outlet air temperature, IT equipment CPU temperature, IT equipment energy consumption, cooling equipment energy consumption, and other energy consumption), as well as the actual operating temperature setpoint of the precision air conditioning at that moment and the actual P value of the data center. UE value; preprocessing the historical operating data, including: removing outliers caused by temperature sensor failure, filling missing values ​​with linear interpolation, unifying all temperature data to degrees Celsius, and unifying energy consumption data to kilowatt-hours; dividing the preprocessed historical data into training and validation sets according to time series, with a ratio of 8:2; each training sample includes the three-dimensional coordinates of each device required for the spatial correlation feature branch and the spatial distance lookup table calculated therefrom, the baseline operating parameters required for the baseline strategy feature branch based on the current season, day / night, and time period index, the various real-time temperature data required for the operating condition feature branch, and the various energy consumption data required for the energy consumption feature branch. The lightweight weighted fusion decision model is trained using a multi-task joint training method. The total loss function of the training is composed of a weighted sum of temperature setpoint prediction loss, PUE prediction loss, and a safety constraint penalty term, with the weights of temperature setpoint prediction loss (1.0), PUE prediction loss (0.5), and safety constraint penalty term (0.1). The temperature setpoint prediction loss uses a mean squared error loss function to calculate the mean squared error between the predicted operating temperature setpoints of all precision air conditioners and the actual operating temperature setpoints of the precision air conditioners. The PUE prediction loss also uses a mean squared error loss function to calculate the mean squared error between the predicted PUE value and the actual PUE value of the data center. The safety constraint penalty term applies an additional penalty loss when the operating temperature setpoint of the precision air conditioner output by the lightweight weighted fusion decision model causes the corresponding cold aisle predicted temperature to exceed the upper limit or fall below the lower limit of the cold aisle temperature. The penalty loss is the square of the difference between the predicted temperature and the corresponding threshold.The Adam optimizer is used to iteratively update all learnable parameters of the lightweight weighted fusion decision model. The initial learning rate of the Adam optimizer is set to 0.001, the batch size is 64, and the weight decay coefficient is 0.0001. Each training iteration includes the following steps: randomly sampling a batch of training samples from the training set, feeding the input data of the training samples into the four dedicated branches for forward propagation, and obtaining the predicted temperature setpoint, predicted PUE, and control coefficient through a gating network and a multi-task output layer; calculating the total loss function value and calculating the gradient of the total loss function with respect to all model parameters through backpropagation; updating the model parameters using the Adam optimizer based on the gradient; calculating the validation loss on the validation set after each training cycle, and triggering an early stopping mechanism when the validation loss no longer decreases for several consecutive cycles, saving the model parameters with the minimum validation loss as the final trained lightweight weighted fusion decision model.

[0082] Example 2

[0083] This embodiment is the second embodiment of this application; it is based on the same inventive concept as Embodiment 1, and refers to... Figure 3 This embodiment introduces a data center cooling and energy-saving system, including a 3D modeling module, a benchmark library construction module, a model building module, and a control execution module;

[0084] The 3D modeling module is used to construct a 3D coordinate system for the computer room and to label the 3D coordinates of each cabinet, IT equipment, precision air conditioner and temperature sensor, as well as the air supply direction of each precision air conditioner.

[0085] The 3D modeling module also uses a spatial directional distance calculation model that includes a direction correction factor to calculate the spatial directional distance from each IT device or temperature sensor in the cold aisle to the precision air conditioner based on the 3D coordinates and air supply direction.

[0086] The benchmark library construction module builds a benchmark operating parameter library based on season, day and night, and time period, and acquires multi-source operating condition data of the data center in real time. The multi-source operating condition data includes precision air conditioning return air temperature, cold aisle temperature, hot aisle temperature, outdoor ambient temperature, weather forecast temperature, IT equipment inlet air temperature, IT equipment outlet air temperature, IT equipment CPU temperature, IT equipment energy consumption data, cooling equipment energy consumption data, and other energy consumption data.

[0087] The model building module constructs a lightweight weighted fusion decision model based on the spatial directional distance, the benchmark operating parameter library, and multi-source operating condition data.

[0088] The control execution module generates control coefficients and operating temperature setpoints for each precision air conditioner through the lightweight weighted fusion decision model, and controls each precision air conditioner based on the control coefficients and operating temperature setpoints.

[0089] The specific functions of each of the above steps are described in the relevant content of the data center cooling and energy-saving method in Embodiment 1, and will not be repeated here.

[0090] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0091] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of protection of this application, and these forms are all within the protection scope of this application.

Claims

1. A data center cooling and energy-saving method, characterized in that, Includes the following steps: Construct a three-dimensional coordinate system for the computer room and mark the three-dimensional coordinates of each cabinet, IT equipment, precision air conditioner and temperature sensor, as well as the air supply direction of each precision air conditioner; Based on the three-dimensional coordinates and air supply direction, a spatial directional distance calculation model including a direction correction factor is used to calculate the spatial directional distance from each IT device or temperature sensor in the cold aisle to the precision air conditioner. A baseline operating parameter library is constructed based on season, day / night, and time of day, and multi-source operating condition data of the data center is acquired in real time; A lightweight weighted fusion decision model is constructed based on the spatial directed distance, the benchmark operating parameter library, and multi-source operating condition data. The lightweight weighted fusion decision model generates the control coefficient and operating temperature setpoint for each precision air conditioner, and controls each precision air conditioner based on the control coefficient and operating temperature setpoint.

2. The data center cooling and energy-saving method as described in claim 1, characterized in that, The specific method for constructing the three-dimensional coordinate system of the computer room is as follows: Using a corner of the computer room as the origin of the coordinate system, the horizontal arrangement of the equipment is defined as the X-axis, the vertical arrangement as the Y-axis, and the vertical height as the Z-axis. Based on the CAD drawings of each server room in the data center and the size data of its internal equipment, determine the three-dimensional coordinates of each rack center, IT equipment air inlet, precision air conditioner air outlet, cold aisle temperature sensor and hot aisle temperature sensor in the coordinate system. The airflow direction of each precision air conditioner outlet is represented as a unit direction vector in the three-dimensional coordinate system of the computer room.

3. The data center cooling and energy-saving method as described in claim 2, characterized in that, The calculation steps for determining the spatial directional distance from each IT device or temperature sensor within the cold aisle to the precision air conditioner, using a spatial directional distance calculation model that includes a direction correction factor, include: Based on the three-dimensional coordinates of the precision air conditioner outlet and the three-dimensional coordinates of the IT equipment or the temperature sensor in the cold aisle, calculate the direction vector from the precision air conditioner outlet to the IT equipment or the temperature sensor in the cold aisle, and calculate the three-dimensional Euclidean straight-line distance between the precision air conditioner outlet and the IT equipment or the temperature sensor in the cold aisle. Calculate the angle between the air supply direction of the precision air conditioner outlet and the direction vector of the corresponding precision air conditioner outlet pointing to the IT equipment or the temperature sensor in the cold aisle, and calculate the direction correction factor based on the angle. The product of the three-dimensional Euclidean straight-line distance and the corresponding direction correction factor is calculated to obtain the spatial directional distance from the corresponding IT equipment or temperature sensor in the cold aisle to the corresponding precision air conditioner.

4. A data center cooling and energy-saving method as described in claim 3, characterized in that, The direction correction factor is calculated as follows: the difference between 1 and the cosine of the included angle is calculated, the difference is multiplied by 2, and the product is added to 1 to obtain the direction correction factor. The direction vector of the precision air conditioner outlet pointing towards the IT equipment or the temperature sensor in the cold aisle is calculated as follows: the three-dimensional coordinates of the precision air conditioner outlet are defined as follows: The three-dimensional coordinates of temperature sensors in IT equipment or cold aisles are The airflow direction of the precision air conditioner's air outlet is The direction vector is calculated as follows: ; in, , , These represent the X-axis, Y-axis, and Z-axis coordinates of the precision air conditioner's air outlet. , , These are the X-axis, Y-axis, and Z-axis coordinates of temperature sensors in IT equipment or cold aisles, respectively. , , The components of the air supply direction of the precision air conditioner outlet on the X-axis, Y-axis, and Z-axis are given.

5. A data center cooling and energy-saving method as described in claim 4, characterized in that, The baseline operating parameter library is used to provide temperature setting baselines and safety thresholds under different operating conditions, and its construction method is as follows: The index is built using season, day and night, and time of day as dimensions; the values ​​for the season dimension include {spring, summer, autumn, winter}, where spring is March to May, summer is June to August, autumn is September to November, and winter is December to February; The day-night dimension includes {day, night}, where day is from 8:00 to 20:00 and night is from 20:00 to 8:

00. The time period dimension includes {valley, plateau, peak}, where valley is from 0 to 8, plateau is from 8 to 10 and 17 to 24, and peak is from 10 to 17; Each index record corresponds to a set of baseline operating parameters; the baseline operating parameters include the precision air conditioning operating temperature setpoint, the upper limit of the cold aisle temperature, the lower limit of the cold aisle temperature, the upper limit of the hot aisle temperature, and the lower limit of the hot aisle temperature.

6. A data center cooling and energy-saving method as described in claim 1, characterized in that, The lightweight weighted fusion decision model is constructed as follows: Using the three-dimensional coordinates of each device in the three-dimensional coordinate system of the computer room, the benchmark operating parameter library and multi-source operating condition data as input data, the key features of the input data are extracted through four dedicated branches: spatial correlation feature branch, benchmark strategy feature branch, operating condition feature branch and energy consumption feature branch. The fusion weight of each dedicated branch under the current operating condition is dynamically calculated through the gating network to obtain the global fusion feature vector. Based on the global fusion feature vector, the control formula for the temperature of the precision air conditioner within the control cycle, the operating temperature setpoint for each precision air conditioner, and the PUE prediction value of the data center are generated through the multi-task output layer.

7. A data center cooling and energy-saving method as described in claim 6, characterized in that, The spatial correlation feature branch, the baseline strategy feature branch, the operating condition feature branch, and the energy consumption feature branch each contain a fully connected layer, which maps their respective input data into a 64-dimensional continuous feature vector. The spatial association feature branch is based on the three-dimensional coordinate data of each device in the three-dimensional coordinate system of the computer room. It calculates the spatial directed distance between each IT device and cold aisle temperature sensor in the computer room and each precision air conditioner in the computer room, and encodes the spatial semantics of the spatial directed distance into a continuous feature vector. At the same time, it establishes a spatial distance lookup table from near to far.

8. A data center cooling and energy-saving method as described in claim 7, characterized in that, The benchmark strategy feature branch encodes the benchmark running parameter library into a continuous feature vector; The operating condition feature branch encodes the precision air conditioner return air temperature, cold aisle temperature, hot aisle temperature, outdoor ambient temperature, weather forecast temperature, IT equipment inlet air temperature, IT equipment outlet air temperature, and IT equipment CPU temperature from the multi-source operating condition data into continuous feature vectors. The energy consumption feature branch encodes IT equipment energy consumption data, refrigeration equipment energy consumption data, and other energy consumption data into continuous feature vectors.

9. A data center cooling and energy-saving method as described in claim 8, characterized in that, The lightweight weighted fusion decision model outputs the global fusion feature vector through the gating network in the following way: The continuous feature vectors output by the four dedicated branches are concatenated and fused to obtain the global context vector. The global context vector is then input into the gating network to solve for the fusion weights, resulting in the gating vector g=[g1,g2,g3,g4]. The sum of the weight coefficients g1, g2, g3 and g4 is 1, and g1, g2, g3 and g4 correspond to the real-time fusion weights of the spatial correlation feature branch, the baseline strategy feature branch, the operating condition feature branch and the energy consumption feature branch, respectively. The continuous feature vectors output by the four dedicated branches are multiplied by the corresponding real-time fusion weights in the gating vector, and then summed element by element to obtain the global fusion feature vector.

10. A data center cooling and energy-saving method as described in claim 9, characterized in that, The multi-task output layer includes a control formula output head, a precision air conditioner operating temperature setpoint output head, and a PUE prediction value output head; The inputs to the output head of the control formula include the global fusion feature vector, the three-dimensional coordinates of the precision air conditioner, the spatial correlation feature vector corresponding to the precision air conditioner, and the temperature anomaly identification vector. The output is the control coefficient of the corresponding precision air conditioner. The precision air conditioner's operating temperature setpoint output head uses the control coefficient generated by the control formula output head as weight to perform a weighted summation of the corresponding real-time temperature and the precision air conditioner's operating temperature setpoint in the multi-source operating condition data and the benchmark operating parameters, thereby obtaining the corresponding precision air conditioner's operating temperature setpoint.

11. A data center cooling and energy-saving system, used to implement a data center cooling and energy-saving method as described in any one of claims 1-10, characterized in that, It includes a 3D modeling module, a benchmark library construction module, a model building module, and a control and execution module; The 3D modeling module is used to construct a 3D coordinate system for the computer room and to label the 3D coordinates of each cabinet, IT equipment, precision air conditioner and temperature sensor, as well as the air supply direction of each precision air conditioner. The 3D modeling module also uses a spatial directional distance calculation model that includes a direction correction factor to calculate the spatial directional distance from each IT device or temperature sensor in the cold aisle to the precision air conditioner based on the 3D coordinates and air supply direction. The benchmark library construction module builds a benchmark operating parameter library based on season, day and night, and time period, and acquires multi-source operating condition data from the data center in real time; The model building module constructs a lightweight weighted fusion decision model based on the spatial directional distance, the benchmark operating parameter library, and multi-source operating condition data. The control execution module generates control coefficients and operating temperature setpoints for each precision air conditioner through the lightweight weighted fusion decision model, and controls each precision air conditioner based on the control coefficients and operating temperature setpoints.

Images