A data center temperature prediction control method
By combining a Generative Moment Matching Neural Network (GMMN) with a temperature and humidity sensor, the problems of accuracy and energy consumption in data center temperature control are solved, enabling rapid and accurate temperature prediction and control, reducing air conditioning energy consumption, and achieving energy conservation and emission reduction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- DEZHOU OURUI ELECTRONICS COMM EQUIP MFG CO LTD
- Filing Date
- 2023-05-09
- Publication Date
- 2026-06-05
AI Technical Summary
Traditional data center temperature control methods suffer from problems such as low density of temperature acquisition points, incomplete coverage, data being greatly affected by location, difficulty in accurately reflecting temperature distribution, large workload of variable data acquisition, resulting in control lag and high energy consumption.
By employing a Generative Moment Matching Neural Network (GMMN) combined with temperature and humidity sensors, a data center temperature prediction model is generated through a generator and encoder. The model is then optimized using the MMD loss function and mean square error, and trained using a deep neural network to achieve accurate prediction and control of data center temperature.
It enables precise and rapid control of data center temperature, reduces air conditioning energy consumption, and achieves energy conservation and emission reduction.
Smart Images
Figure CN116744634B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data center temperature control technology, and more specifically to a data center temperature prediction and control method. Background Technology
[0002] With the development of internet technology and the rise of big data, the increasing amount of data to be processed means that data centers need more servers. To ensure the normal operation of these servers, a certain amount of air conditioning equipment is required for cooling. However, real-time temperature control in data centers suffers from lag, making real-time temperature prediction and subsequent temperature regulation crucial. Traditional temperature control involves placing temperature sensors at several dispersed locations within the data center. When a temperature at a certain point exceeds a preset level, the air conditioning fan speed is increased to control the temperature. The traditional method has several drawbacks: firstly, the density of temperature sensors is low, making it difficult to cover all effective areas of the data center, and the collected data is greatly affected by its location, failing to accurately reflect the actual temperature distribution and operating conditions of the equipment; secondly, changing the temperature sensor locations is labor-intensive and time-consuming, requiring redesign; and thirdly, it is difficult to determine the relationship between the temperature at different locations within the data center and factors such as the airflow, temperature, and location of the air conditioning vents. This not only fails to provide timely and effective temperature control but also increases energy consumption, hindering energy conservation and emission reduction. Summary of the Invention
[0003] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a data center temperature prediction and control method that can not only quickly and effectively control the temperature of the data center, but also achieve energy saving and emission reduction.
[0004] The technical solution adopted by this invention to solve its technical problem is:
[0005] A data center temperature prediction and control method includes the following steps:
[0006] (1) The temperature of the air outlet of the air duct, the gas flow rate, the temperature of the return air outlet and the indoor humidity are obtained by the temperature and humidity sensor. When the predicted temperature of the natural air outlet of the air duct is higher than or equal to the preset temperature of the equipment, the air outlet of the natural air duct is closed. If the predicted temperature of the natural air outlet of the air duct is lower than the preset temperature of the equipment, the parameters measured by the temperature and humidity sensor are input into the data center temperature prediction model to obtain the temperature information of the data center equipment.
[0007] (2) Determine whether the predicted data center temperature meets the preset temperature of the data center. If it does, use the above-mentioned air conditioner outlet parameters and air conditioner location information to control the temperature. If it does not meet, the user re-enters the air conditioner outlet parameters until the data center equipment temperature reaches the preset temperature.
[0008] Furthermore, this invention utilizes a Generative Moment Matching Neural Network (GMMN) to predict data center temperature, where the GMMN consists of a generator and an autoencoder. Specifically, random noise is sampled from a predetermined prior distribution and used as input to the trained generative network, and the output is inversely normalized to obtain a data center temperature prediction model that conforms to the corresponding law.
[0009] Furthermore, the generator concept is as follows: for the prior distribution Z ~ P Z(Z) Sampling is performed to obtain a random noise vector Z, and then the random noise vector Z and the actual temperature curve X are constructed. t The mapping relationship between them, i.e.
[0010] X g =f g (ZW g +B g ),
[0011] In the formula: X g Temperature profile generated by the generator; f g W is the activation function for the generator network. g B is the weight matrix; g This is the bias vector.
[0012] Furthermore, to avoid training difficulties, MMD, used to describe distribution distance, is proposed as a loss function to build the model, measuring the generated temperature curve X. g and actual temperature curve X t The distribution differences between them
[0013]
[0014] Where N and M are the number of actual temperature curves and the number of generated temperature curves, respectively; θ is a regenerating kernel transformation.
[0015] Expanding the above equation and applying a Gaussian transform to ensure that the generated temperature curve and the actual temperature curve have the same distribution pattern, we obtain:
[0016]
[0017] In the formula, k represents a kernel transformation, x′ represents x in the new space, and σ is the window width. The resulting loss function allows for simpler updates to the weights and facilitates the training process.
[0018] Furthermore, during the encoding process, the encoder receives the actual temperature curve X as input data. t After processing through multiple fully connected layers, the low-dimensional latent variable H is obtained, resulting in:
[0019] H = f e (X t W e +B e ),
[0020] In the formula, f e W is the activation function of the encoder. e B is the weight matrix of the encoder; e It is the bias vector;
[0021] During the decoding process, the input data to the decoder is H, which is processed to obtain the reconstructed result X. d The decoding formula is:
[0022] X d =f d (HW d +B d ),
[0023] In the formula: f d W is the activation function of the decoder. d B is the weight matrix of the decoder; d Let it be its bias vector.
[0024] Furthermore, to make the input and output data as similar as possible, the loss function L is defined as the mean squared error (MSE), and the formula is:
[0025]
[0026] In the formula: m is the number of elements in each temperature curve; x t，i x is the i-th element of the actual temperature curve; d，i This is the i-th element for reconstructing the temperature curve.
[0027] Furthermore, in step (2), the data center temperature prediction model uses the pre-collected air duct outlet temperature, gas flow rate, return air outlet temperature, indoor humidity, data center air conditioner location information, data center equipment temperature information, and data center equipment location information as independent variables, and the data center temperature at the next moment as the dependent variable, and uses the neural network model to train and obtain the data center temperature.
[0028] Furthermore, in step (2), the training process includes:
[0029] Obtain information on natural air duct outlet parameters, natural air duct location, natural air duct outlet location, data center equipment location, data center air conditioning outlet location, and data center air conditioning outlet parameters;
[0030] The data center temperature prediction model is obtained by using the temperature and location information of the data center equipment collected at each moment as output, and the natural air duct outlet parameter information, natural air duct outlet location information, data center air conditioning outlet parameter information, and data center air conditioning outlet location information collected at that moment as input, and training a deep neural network model.
[0031] Technical effects of the present invention:
[0032] Compared with existing technologies, the data center temperature prediction and control method of the present invention can accurately reflect the temperature distribution and actual operating conditions of the equipment in the data center. It is convenient to determine the relationship between the temperature at various locations in the data center and factors such as the air volume, temperature and location of the air conditioning outlet. It achieves accurate, fast and effective control of the temperature of data center equipment, while using natural air ducts to reduce air conditioning energy consumption, thus playing a role in energy conservation and emission reduction. Attached Figure Description
[0033] Figure 1 This is a flowchart of the temperature prediction and control method of the present invention;
[0034] Figure 2 A flowchart is provided for establishing the data center temperature prediction model of this invention. Detailed Implementation
[0035] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.
[0036] Example 1:
[0037] like Figure 1 As shown, this embodiment relates to a data center temperature prediction and control method, which includes the following steps:
[0038] (1) Taking a certain moment as a reference, the temperature of the air outlet of the air duct, the gas flow rate, the temperature of the return air outlet, and the indoor humidity are obtained by installing temperature and humidity sensors. A model is established based on the layout, location, and equipment parameters of the data center. When the predicted temperature of the natural air outlet of the air duct is higher than or equal to the preset temperature of the equipment, the natural air outlet of the air duct is closed; if the predicted temperature of the natural air outlet of the air duct is lower than the preset temperature of the equipment, the parameters measured by the temperature and humidity sensors are input into the data center temperature prediction model to obtain the temperature information of the data center equipment. Specifically, this invention uses a generative moment matching neural network (GMMN) to predict the temperature of the data center, wherein the generative moment matching neural network is divided into a generator and an autoencoder.
[0039] 1) Generator
[0040] The basic idea of a generator is to use a simple prior distribution Z~P Z(Z) Sampling is performed to obtain a random noise vector Z, and then the random noise vector Z and the actual temperature curve X are constructed. t The mapping relationship between them, i.e.
[0041] X g =f g (ZW g +B g ),
[0042] In the formula: X g Temperature profile generated by the generator; f g W is the activation function for the generator network. g B is the weight matrix; g This is the bias vector.
[0043] To avoid training difficulties, MMD, used to describe distribution distance, is proposed as a loss function to build the model, measuring the generated temperature curve X. g and actual temperature curve X t Distribution differences between
[0044]
[0045] Where N and M are the number of actual temperature curves and the number of generated temperature curves, respectively; θ is a regenerating kernel transformation.
[0046] Expanding the above equation and applying a Gaussian transform to ensure that the generated temperature curve and the actual temperature curve have the same distribution pattern, we obtain:
[0047]
[0048] In the formula, k represents a kernel transformation, x′ represents x in the new space, and σ is the window width. The resulting loss function allows for simpler updates to the weights and facilitates the training process.
[0049] 2) Encoder
[0050] In the encoding process of this invention, the input data is the actual temperature curve X. t After processing through multiple fully connected layers, the low-dimensional latent variable H is obtained, resulting in:
[0051] H = f e (X t W e +B e ),
[0052] In the formula, f e W is the activation function of the encoder. e B is the weight matrix of the encoder; e This is the bias vector.
[0053] During the decoding process, the input data to the decoder is H, which is processed to obtain the reconstructed result X. d The decoding formula is:
[0054] X d =f d (HW d +B d ),
[0055] In the formula: f d W is the activation function of the decoder. d B is the weight matrix of the decoder; d Let it be its bias vector.
[0056] To make the input and output data as similar as possible, the loss function L is defined as the mean squared error (MSE), and the formula is:
[0057]
[0058] In the formula: m is the number of elements in each temperature curve; x t，i x is the i-th element of the actual temperature curve; d，i This is the i-th element for reconstructing the temperature curve.
[0059] 3) Temperature curve generation
[0060] The above-described generative moment matching neural network can generate predicted data center temperature curves. The specific temperature prediction model generation process is as follows: Figure 2As shown, random noise is obtained by sampling the set prior distribution and inputting it into the trained generator network. By inverse normalizing the output, a data center predicted temperature curve that follows the corresponding law can be obtained.
[0061] (2) Determine whether the predicted data center equipment temperature meets the preset temperature of the data center. That is, if there is an error of ±2° between the predicted temperature and the preset temperature, the predicted data center temperature is considered to meet the expected ideal temperature of the data center. If it meets the requirements, the temperature is controlled using the air conditioning outlet parameters and air conditioning location information mentioned above. If it does not meet the requirements, the user can re-enter the air conditioning outlet parameters until the data center equipment temperature reaches the preset temperature. The above data center temperature prediction model obtains the data center temperature by training a deep neural network model based on the pre-collected natural air duct outlet parameters, data center air conditioning outlet parameter information, data center air conditioning location information, data center equipment temperature information, and data center equipment location information.
[0062] The prediction model training process described in this invention includes:
[0063] Obtain information on natural air duct outlet parameters, natural air duct location, natural air duct outlet location, data center equipment location, data center air conditioning outlet location, and data center air conditioning outlet parameters;
[0064] The data center temperature prediction model is obtained by using the temperature and location information of the data center equipment collected at each moment as output, and the natural air duct outlet parameter information, natural air duct outlet location information, data center air conditioning outlet parameter information, and data center air conditioning outlet location information collected at that moment as input, and training a deep neural network model.
[0065] The above-described specific embodiments are merely specific examples of the present invention. The patent protection scope of the present invention includes, but is not limited to, the above-described specific embodiments. Any appropriate changes or modifications made by a person skilled in the art that conform to the claims of the present invention should fall within the patent protection scope of the present invention.
Claims
1. A data center temperature prediction and control method, characterized in that: Includes the following steps: (1) The temperature of the natural air duct outlet, the temperature of the return air duct, and the indoor humidity are obtained by temperature and humidity sensors, and the gas flow rate is obtained by wind speed sensor. When the predicted temperature of the natural air duct outlet is higher than or equal to the preset temperature of the data center, the natural air duct outlet is closed. If the predicted temperature of the natural air duct outlet is lower than the preset temperature of the data center, the parameters measured by the temperature and humidity sensor are input into the data center temperature prediction model to obtain the data center temperature information. (2) Determine whether the predicted data center temperature meets the preset temperature of the data center. If it does, use the above natural air duct outlet parameters and air conditioner location information to control the temperature. If it does not meet, the user re-enters the natural air duct outlet parameters until the data center temperature reaches the preset temperature. The generating moment matching neural network is used to predict the temperature of a data center. The generating moment matching neural network consists of a generator and an autoencoder. The generator concept is as follows: Based on the prior distribution... Sampling is performed to obtain a random noise vector. Then construct a random noise vector. and actual temperature curve The mapping relationship between them, i.e. ， In the formula: The temperature profile generated by the generator; The activation function for generating the network; This is the weight matrix; It is the bias vector; The model is built using MMD as the loss function, and the generated temperature curve is measured. and actual temperature curve The distribution differences between them ， Where N and M are the number of actual temperature curves and the number of generated temperature curves, respectively; For a regenerating kernel transformation; Expanding the above equation and applying a Gaussian transform to ensure that the generated temperature curve and the actual temperature curve have the same distribution pattern, we obtain: ， In the formula, k represents a certain kernel transformation. Let x be the representation of x in the new space. ; During the encoding process, the encoder inputs actual temperature curves as data. Low-dimensional latent variables are obtained after processing through multiple fully connected layers. ,have to: ， In the formula, This is the activation function for the encoder; This is the weight matrix of the encoder; It is the bias vector; During the decoding process, the input data to the decoder is The reconstructed result is obtained after processing. The decoding formula is: ， In the formula: is the activation function for the decoder; This is the weight matrix of the decoder; Its bias vector; Define loss function The mean squared error is given by the formula: ， In the formula: The number of elements in each temperature curve; The first of the actual temperature curves One element; To reconstruct the temperature profile Each element.
2. The data center temperature prediction and control method according to claim 1, characterized in that: In step (2), the data center temperature prediction model uses the pre-collected natural air duct outlet temperature, gas flow rate, return air outlet temperature, indoor humidity, data center air conditioning location information, data center temperature information and data center location information as independent variables, and the data center temperature at the next moment as the dependent variable, and uses a neural network model to train and obtain the data center temperature.
3. The data center temperature prediction and control method according to claim 1, characterized in that: In step (2), the training process includes: Obtain information on the parameters of the natural air duct outlet, the location of the natural air duct, the location of the natural air duct outlet, the location of the data center, the location of the data center air conditioning outlet, and the parameters of the data center air conditioning outlet. The data center temperature and location information collected at each moment are used as outputs, and the natural air duct outlet parameters, natural air duct outlet locations, data center air conditioning outlet parameters, and data center air conditioning outlet locations collected at that moment are used as inputs. A data center temperature prediction model is obtained by training a deep neural network model.