Machine room temperature field prediction method and model training method
By determining the operating condition category and generating input parameters based on the current operating conditions in the temperature field prediction of data center computer rooms, calling a dedicated temperature prediction model, and combining it with a neural network model for prediction, the problem of complex and time-consuming calculations in existing technologies has been solved, and efficient temperature field prediction has been achieved.
Patent Information
- Application Number
- CN202511547350.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-27
- Publication Date
- 2025-12-16
AI Technical Summary
In the prediction of temperature fields in data center computer rooms, existing technologies rely on computational fluid dynamics (CFD) simulation methods, which are computationally complex and time-consuming, failing to meet real-time operational requirements. In data-driven models, existing prediction models based on machine learning methods cannot effectively extract and integrate key physical features from the physical spatial structure and complex and ever-changing operating conditions of the computer room, resulting in poor prediction performance.
By determining the operating condition category based on the current operating conditions, input parameters including physical boundary feature maps and initial temperature fields are generated. A dedicated temperature prediction model is called, and a neural network model is used for prediction. By deeply integrating physical prior knowledge with real-time operating status, a rapid three-dimensional temperature field output is achieved.
It enables intelligent identification and classification of complex operating states, improves the accuracy and adaptability of the prediction model, and solves the problem of poor prediction results caused by insufficient generalization ability of existing methods.
Smart Images
Figure CN121144962A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of model training technology, and in particular to a method for predicting the temperature field of a computer room and a method for training a model. Background Technology
[0002] In the field of predicting fluid temperature fields in data center computer rooms, computational fluid dynamics (CFD) simulation technology is commonly used. Although this method can obtain three-dimensional temperature field distribution, its calculation process is complex and consumes a lot of computing resources and time. Completing a single simulation usually takes several hours or even longer, which is difficult to meet the stringent requirements of real-time operation for data timeliness.
[0003] To address this, related technologies have attempted to employ traditional algorithms such as proper orthogonal decomposition (POD) and principal component analysis (PCA) for rapid prediction of data center temperature fields. However, these methods primarily rely on a purely data-driven approach to learn flow field information, failing to effectively extract and integrate key physical features from the data center's physical spatial structure and complex, variable operating conditions. This results in poor prediction performance under certain operating conditions. Summary of the Invention
[0004] This application provides a method for predicting the temperature field of a computer room and a method for training a model, which are used to predict the temperature field of a computer room.
[0005] In a first aspect, this application provides a method for predicting the temperature field of a computer room. The method includes: determining the operating condition category of the target computer room based on its current operating conditions; generating input parameters based on the current operating conditions and the physical layout of the target computer room; the input parameters include a physical boundary feature map and a first initial temperature field; determining the corresponding temperature prediction model according to the operating condition category; one operating condition category corresponds to one temperature prediction model; and inputting the input parameters into the temperature prediction model to obtain the predicted temperature field of the target computer room.
[0006] In some embodiments, determining the operating condition category of the target data center based on its current operating condition includes: acquiring feature data of the current operating condition; calculating the distance between the feature data and the central feature data of multiple clusters; and taking the operating condition category of the cluster with the closest distance among the multiple clusters as the operating condition category of the target data center.
[0007] Each cluster corresponds to a working condition category.
[0008] In some embodiments, the current operating conditions include the power of the first cabinet and the setting parameters of the first air conditioner; based on the current operating conditions and the physical layout of the target computer room, input parameters are generated, including: generating a first initial temperature field based on the power of the first cabinet and the setting parameters of the first air conditioner; generating a physical boundary feature map based on the physical layout of the target computer room; the physical boundary feature map is used to characterize the relative positional relationship between any point in the target computer room space and the nearest physical boundary.
[0009] In some embodiments, the method further includes: when the error between the predicted temperature field and the actual temperature field of the target computer room is greater than an error threshold, performing incremental learning on the temperature prediction model based on the current operating conditions.
[0010] The technical solution provided in this application brings at least the following beneficial effects: This technical solution determines the operating condition category by the current operating condition, realizing intelligent identification and classification of complex operating states, laying the foundation for subsequent accurate prediction; then, based on the physical layout and the current operating condition, it generates input parameters containing physical boundary feature maps and initial temperature fields, deeply integrating physical prior knowledge with real-time operating states, providing high-quality input for the model; on this basis, it calls the corresponding dedicated temperature prediction model according to the operating condition category, avoiding the dilemma of accuracy and generalization ability of a single model in dealing with all operating conditions; finally, it inputs the input parameters into the dedicated model for prediction, enabling the model to quickly output a three-dimensional temperature field under constraints, effectively solving the problem of poor prediction results caused by insufficient model generalization ability in existing methods when various operating conditions occur.
[0011] Secondly, this application provides a training method for a data center temperature field prediction model. The method includes: determining multiple operating condition categories based on the historical operating conditions of the target data center; for each operating condition category, selecting operating conditions belonging to the category from the historical operating conditions; determining the dataset of the neural network model corresponding to the operating condition category based on the operating conditions and the physical layout of the target data center; one neural network model corresponding to one operating condition category; the dataset includes the physical boundary feature map of the target data center, a second initial temperature field, and the actual temperature field; iteratively training the neural network model according to the dataset until the loss value of the neural network model is lower than a preset threshold, and using the neural network model as the data center temperature field prediction model for the operating condition category.
[0012] In some embodiments, the operating conditions of the operating condition category include the power of the second cabinet and the setting parameters of the second air conditioner; based on the operating conditions and the physical layout of the target computer room, the dataset of the neural network model corresponding to the operating condition category is determined, including: generating a second initial temperature field according to the power of the second cabinet and the setting parameters of the second air conditioner; generating a physical boundary feature map based on the physical layout of the target computer room; the physical boundary feature map is used to characterize the relative positional relationship of any point in the target computer room space to the nearest physical boundary; and pairing the physical boundary feature map, the second initial temperature field and the actual temperature field corresponding to the operating condition in time to generate a dataset.
[0013] In some embodiments, the neural network model is iteratively trained based on the dataset until the loss value of the neural network model is lower than a preset threshold, and the neural network model is used as the computer room temperature field prediction model for the working condition category. This includes: inputting the physical boundary feature map and the second initial temperature field into the neural network model to obtain the temperature field prediction result; calculating the difference between the temperature field prediction result and the actual temperature field in the dataset to obtain the loss value of the current iteration; and adjusting the parameters of the neural network model based on the loss value until the loss value is lower than the preset threshold, and using the neural network model as the computer room temperature field prediction model.
[0014] In some embodiments, multiple operating condition categories are determined based on the historical operating conditions of the target data center, including: acquiring feature data of historical operating conditions; clustering the feature data to determine the clustering results of the feature data; the clustering results include multiple clusters; one cluster corresponds to one operating condition category; and multiple operating condition categories are determined based on multiple clusters.
[0015] In some embodiments, clustering feature data and determining the clustering result of the feature data includes: calculating the distance between feature data points and the center points of multiple clusters; iteratively updating the cluster to which the feature data points belong and the center points of multiple clusters according to the nearest neighbor principle until the distance is less than a distance threshold, and determining the clustering result.
[0016] The technical solution provided in this application provides at least the following beneficial effects: Multiple operating condition categories are determined based on historical operating conditions, enabling a systematic summary of complex historical operating patterns and laying the foundation for building specialized models; then, specific operating condition data is selected for each operating condition category, ensuring the relevance of training data and providing a high-quality data source for training high-precision specialized models; based on this, a dataset containing physical boundary feature maps, initial temperature fields, and actual temperature fields is constructed for each operating condition category, providing strong constraints and clear objectives for model learning; finally, the dataset is used to iteratively train each neural network model, ensuring that the multiple models trained can accurately capture the temperature field characteristics under different operating conditions, thereby significantly improving the overall accuracy and adaptability of the prediction model.
[0017] Thirdly, this application provides a data center temperature field prediction device, the device comprising: a processing unit; the processing unit being configured to determine the operating condition category of the target data center based on its current operating conditions; the processing unit being further configured to generate input parameters based on the current operating conditions and the physical layout of the target data center; the input parameters including a physical boundary feature map and a first initial temperature field; the processing unit being further configured to determine a corresponding temperature prediction model according to the operating condition category; one operating condition category corresponds to one temperature prediction model; the processing unit being further configured to input the input parameters into the temperature prediction model to obtain the predicted temperature field of the target data center.
[0018] Fourthly, this application provides a training device for a data center temperature field prediction model. The device includes: a processing unit; the processing unit is used to determine multiple operating condition categories based on the historical operating conditions of the target data center; the processing unit is further used to filter out operating conditions belonging to each operating condition category from the historical operating conditions for each operating condition category; the processing unit is further used to determine the dataset of the neural network model corresponding to the operating condition category based on the operating conditions and the physical layout of the target data center; one operating condition category corresponds to one neural network model; the dataset includes the physical boundary feature map of the target data center, a second initial temperature field, and the actual temperature field; the processing unit is further used to iteratively train the neural network model according to the dataset until the loss value of the neural network model is lower than a preset threshold, and use the neural network model as the data center temperature field prediction model for the operating condition category.
[0019] Fifthly, this application provides an electronic device comprising: a processor and a memory; the memory storing processor-executable instructions; when the processor is configured to execute the instructions, causing the electronic device to implement the room temperature field prediction method of the first aspect and any possible implementation thereof, or the training method of the room temperature field prediction model of the second aspect and any possible implementation thereof.
[0020] Sixthly, this application provides a computer-readable storage medium that, when the instructions in the computer-readable storage medium are executed by a processor of a server, enables the server to execute the data center temperature field prediction method provided by the first aspect, the second aspect, and any possible implementation thereof; or, when the instructions in the computer-readable storage medium are executed by a processor of a client, enables the client to execute the data center temperature field prediction method of the first aspect and any possible implementation thereof, or the data center temperature field prediction model training method of the second aspect and any possible implementation thereof.
[0021] Seventhly, this application provides a computer program product. The computer program product includes computer program instructions, which, when executed on a computer (e.g., a receiving node), cause the computer to perform a data center temperature field prediction method as described in the first aspect and any possible implementation thereof, or a data center temperature field prediction model training method as described in the second aspect and any possible implementation thereof.
[0022] The beneficial effects of aspects two through seven mentioned above can be referred to aspect one, and will not be repeated here. Attached Figure Description
[0023] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0024] Figure 1 A schematic diagram of a computer room temperature field prediction system provided in this application embodiment; Figure 2 A flowchart illustrating a method for predicting the temperature field of a computer room, provided in an embodiment of this application; Figure 3 A schematic diagram of a three-dimensional distance distribution provided in an embodiment of this application; Figure 4 A schematic diagram illustrating a method for estimating the temperature field in a computer room, provided as an embodiment of this application; Figure 5 A flowchart illustrating another method for predicting the temperature field of a computer room provided in this application embodiment; Figure 6 A flowchart illustrating a training method for a computer room temperature field prediction model provided in this application embodiment; Figure 7 A schematic diagram of a measured three-dimensional temperature field provided for an embodiment of this application; Figure 8 A schematic diagram illustrating a method for predicting the temperature field in a computer room, as provided in an embodiment of this application. Figure 9 A schematic diagram of a computer room temperature field prediction device provided in an embodiment of this application; Figure 10 A schematic diagram of the structure of a training device for a computer room temperature field prediction model provided in an embodiment of this application; Figure 11 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation
[0025] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0026] It should be noted that in the embodiments of this application, the words "exemplarily" or "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design scheme described as "exemplarily" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of the words "exemplarily" or "for example" is intended to present the relevant concepts in a specific manner.
[0027] To facilitate a clear description of the technical solutions of the embodiments of this application, the terms "first" and "second" are used in the embodiments of this application to distinguish the same or similar items with essentially the same function and effect. Those skilled in the art can understand that the terms "first" and "second" are not intended to limit the quantity or execution order.
[0028] With the increasing demand for green and low-carbon transformation of critical information infrastructure such as data centers, the power efficiency of newly built and expanded large and super-large data centers needs to be reduced to below 1.25, demonstrating the significant technical challenges still faced in energy conservation and carbon reduction. The rationality of airflow organization within the data center directly determines the energy consumption level of the cooling system and the operational reliability of IT equipment, making it a core element in achieving efficient, safe, and stable operation.
[0029] Currently, computational fluid dynamics (CFD) methods are commonly used to perform three-dimensional numerical simulations of the temperature field within computer rooms to evaluate and optimize cooling efficiency. However, CFD methods require complex discrete solutions and iterative calculations, consuming significant computational resources, and a single simulation typically takes several hours or more, failing to meet the high timeliness requirements for rapid prediction and real-time control of the temperature field during actual data center operations.
[0030] Based on this, with the development of sensing and monitoring and Internet of Things technologies, data centers are now able to continuously collect and store massive amounts of historical operational data, laying the foundation for building data-driven intelligent temperature field prediction models.
[0031] To overcome the aforementioned problems, existing research has explored combining machine learning methods with CFD simulations. Under preset boundary and initial conditions, deep learning models such as convolutional neural networks, generative adversarial networks, and autoencoders are used to achieve rapid reconstruction of temperature fields and intelligent inference of physical fields. However, this type of method still suffers from the following significant drawbacks: First, data center environments are characterized by high-dimensional, heterogeneous spatial layouts and strong nonlinear thermodynamics. Traditional feature extraction methods based on linear assumptions (such as POD and PCA) have limited characterization capabilities, complex feature selection processes, and difficulty in achieving high-precision predictions. Secondly, most existing prediction models rely on a pure data-driven approach to learn flow field information, failing to effectively extract and integrate key physical features from the physical spatial structure of the computer room and the complex and ever-changing operating conditions, resulting in reduced reliability of prediction results under drastic changes in boundary conditions or extreme operating conditions. Moreover, the predictive models rely heavily on simulation data for training, while the CFD method itself has limitations in terms of computational efficiency and model accuracy. In addition, the high dimensionality and nonlinearity of actual working conditions result in insufficient coverage and diversity of training data, which significantly limits the generalization performance of the model.
[0032] Therefore, improving the accuracy of room temperature field prediction when various working conditions occur is a problem that needs to be studied.
[0033] Based on this, this application provides a method for predicting the temperature field of a computer room. It determines the operating condition category based on the current operating conditions, achieving intelligent identification and classification of complex operating states, laying the foundation for subsequent accurate prediction. Furthermore, it generates input parameters containing physical boundary feature maps and an initial temperature field based on the physical layout and current operating conditions, deeply integrating prior physical knowledge with real-time operating status to provide high-quality input for the model. On this basis, it calls the corresponding dedicated temperature prediction model according to the operating condition category, avoiding the dilemma of accuracy and generalization ability of a single model for all operating conditions. Finally, it inputs the input parameters into the dedicated model for prediction, enabling the model to quickly output a three-dimensional temperature field under constraints, effectively solving the problem of poor prediction results caused by insufficient model generalization ability in existing methods when various operating conditions occur.
[0034] Figure 1 This is a schematic diagram of the structure of a computer room temperature field prediction system provided in an embodiment of this application, as shown below. Figure 1 As shown, the computer room temperature field prediction system 100 includes at least one prediction module 101, a communication line 102, and at least one acquisition module 103. The prediction module 101 and the acquisition module 103 are connected via the communication line 102.
[0035] In some embodiments, the acquisition module 103 is used to acquire the current operating status of the target computer room and send the current operating status to the prediction module 101 through the communication line 102.
[0036] Accordingly, the prediction module 101 receives the current operating conditions from the acquisition module 103.
[0037] In some embodiments, the prediction module 101 is used to determine the operating condition category of the target computer room based on its current operating condition, and to generate input parameters based on the current operating condition and the physical layout of the target computer room. The input parameters include a physical boundary feature map and a first initial temperature field. Then, according to the operating condition category, the corresponding temperature prediction model is determined. One operating condition category corresponds to one temperature prediction model. Finally, the input parameters are input into the temperature prediction model to obtain the predicted temperature field of the target computer room.
[0038] Through the above description of the implementation methods, those skilled in the art will clearly understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the network node can be divided into different functional modules to complete all or part of the functions described above. The specific working process of the system, modules, and network nodes described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.
[0039] Figure 2 This is a flowchart illustrating a method for predicting the temperature field of a computer room according to an embodiment of this application. The method for predicting the temperature field of a computer room provided in this embodiment can be applied to, for example... Figure 1 In the computer room temperature field prediction system shown, the computer room temperature field prediction method provided in this application embodiment can be implemented through the following steps.
[0040] S101. Based on the current operating conditions of the target computer room, determine the operating condition category to which the target computer room belongs.
[0041] As one possible implementation of S101 above, the feature data of the current operating condition is obtained; the distance between the feature data and the central feature data of multiple clusters is calculated; and the operating condition category corresponding to the cluster with the closest distance among the multiple clusters is taken as the operating condition category to which the target computer room belongs.
[0042] Each cluster corresponds to a working condition category.
[0043] For example, the feature data of the current operating condition can be a multi-dimensional vector, such as (total power: 150kW, air conditioning set temperature: 18°C, total air volume: 30,000 CFM). Assume that three cluster centers have been generated through historical data clustering, representing the three operating condition categories of "high power, high heat density," "medium power, balanced operation," and "low power, energy saving," respectively. The system calculates the Euclidean distance between the current feature vector and these three cluster centers, finding that the distance to the "high power, high heat density" cluster center is the smallest; therefore, the current operating condition is classified as belonging to the "high power, high heat density" category.
[0044] For example, the distance between feature data and the central feature data of multiple clusters can be determined using the following formula 1. .
[0045] Formula 1 in, For feature data, It is a cluster The mean vector.
[0046] S102. Generate input parameters based on the current operating conditions and the physical layout of the target computer room.
[0047] The input parameters include the physical boundary feature map and the first initial temperature field.
[0048] In some embodiments, the current operating conditions include the power of the first cabinet and the first air conditioning settings. The first air conditioning settings include the air conditioning temperature and the air conditioning airflow.
[0049] As one possible implementation of the above S102, a first initial temperature field is generated based on the power of the first cabinet and the first air conditioning setting parameters; a physical boundary feature map is generated based on the physical layout of the target computer room.
[0050] Among them, the physical boundary feature map is used to characterize the relative positional relationship between any point in the target computer room space and the nearest physical boundary.
[0051] For example, when generating physical boundary feature maps, the Signed Distance Function (SDF) can be used. For any point (x, y, z) in the three-dimensional space of the computer room, calculate its distance to the nearest physical boundary such as a wall, cabinet, or air conditioner. The distance is positive inside the boundary and negative outside, generating a map like this. Figure 3 The diagram shows a three-dimensional distance distribution.
[0052] In some embodiments, when generating the first initial temperature field, estimation can be based on simple heat balance rules. For example, the temperature of the air conditioner vent area can be initialized to the air conditioner set temperature (e.g., 18°C), and the heat can be diffused to the surrounding area according to the rack power attenuation function over distance, forming an initial guess of a high-temperature zone near the rack heat source. For example, the server room space is divided into server racks, enclosed corridors, and other areas. The locations of the air conditioners in different areas are assigned values according to Formula 2, forming a structure like... Figure 4 The diagram shows a possible estimation of the temperature field in a computer room.
[0053] Formula 2 in, It is a function determined by the air conditioning temperature. It is a function determined by the rack power. It is a function determined by the air conditioning air volume.
[0054] S103. Determine the corresponding temperature prediction model based on the operating condition category.
[0055] One operating condition category corresponds to one temperature prediction model.
[0056] For example, if it is determined in S101 that the current working condition belongs to the "high power high heat density" category, then a "high power high heat density model" trained specifically for this category is loaded from the pre-stored model library. This working condition-specific, model-specific strategy can effectively improve prediction accuracy and speed. S104. Input the input parameters into the temperature prediction model to obtain the predicted temperature field of the target computer room.
[0057] For example, the physical boundary feature map generated in S102 and the first initial temperature field are used as input tensors and input into a high-power, high-heat-density model. This model is a convolutional neural network based on the U-Net architecture, which performs forward inference through an encoder-decoder structure and outputs a three-dimensional matrix with the same resolution as the input within seconds. Each element of this matrix represents the predicted temperature value at the corresponding spatial location, i.e., the complete three-dimensional predicted temperature field. The technical solution provided in this application provides at least the following beneficial effects: It determines the operating condition category based on the current operating condition, achieving intelligent identification and classification of complex operating states, laying the foundation for subsequent accurate prediction; furthermore, it generates input parameters containing physical boundary feature maps and initial temperature fields based on the physical layout and current operating conditions, deeply integrating prior physical knowledge with real-time operating states to provide high-quality input for the model; on this basis, it calls the corresponding dedicated temperature prediction model according to the operating condition category, avoiding the dilemma of accuracy and generalization ability of a single model for all operating conditions; finally, it inputs the input parameters into the dedicated model for prediction, enabling the model to quickly output a three-dimensional temperature field under constraints, effectively solving the problem of poor prediction results caused by insufficient model generalization ability in existing methods when various operating conditions occur.
[0058] like Figure 5 As shown, the method for predicting the temperature field of a computer room also includes the following steps.
[0059] S201. When the error between the predicted temperature field and the actual temperature field in the target computer room is greater than the error threshold, the temperature prediction model is incrementally learned based on the current operating conditions.
[0060] In one possible implementation, For example, the system continuously compares the predicted temperature with the actual sensor readings at key points in the computer room (such as the air inlets of each rack). When the error in a certain area is found to exceed 3°C (error threshold), the system automatically records the data packet at that moment, such as the current operating conditions, the generated input parameters, the predicted temperature field, and the actual sensor data.
[0061] Furthermore, the Elastic Weight Consolidation (EWC) algorithm is used to incrementally learn the corresponding "expert model". By introducing a regularization term, the EWC algorithm optimizes the model to fit new data while constraining the magnitude of changes in parameters that are important for previously learned tasks, thereby effectively mitigating the forgetting problem and enabling the model to continuously adapt to new operating states without retraining from scratch. For example, by integrating the EWC algorithm with neural networks, specific regularization constraints can be introduced into the loss function during the training process to limit the model's modification of key parameters of old tasks when learning new tasks.
[0062] Specifically, initial training: based on the dataset from initial scenario A. The neural network model is trained to obtain a set of optimized model parameters θA. The importance Fi of each parameter in θA to task A is calculated (usually based on the Fisher information matrix). When faced with a new dataset in scenario B... When the objective function is generated, it can be obtained using the following formula 3.
[0063] Formula 3 Among them, A and B are two independent scenarios. and These are the training datasets for each scenario.
[0064] Figure 6 This is a flowchart illustrating a training method for a computer room temperature field prediction model provided in an embodiment of this application. The model training method provided in this embodiment can be implemented through the following steps.
[0065] S401. Based on the historical operating conditions of the target computer room, determine multiple operating condition categories.
[0066] As one possible implementation of the above S401, the characteristic data of historical operating conditions are obtained, the characteristic data are clustered and divided, the clustering results of the characteristic data are determined, and multiple operating condition categories are determined based on multiple clusters.
[0067] The clustering results include multiple clusters; each cluster corresponds to a working condition category.
[0068] For example, operational records of the data center under different seasons and loads over the past three months were collected. Each record included the total power of the server racks, air conditioning settings, etc. Principal component analysis (PCA) was first used to reduce the dimensionality of the high-dimensional data, preserving the main features.
[0069] Furthermore, the K-Means clustering algorithm is adopted, and the silhouette score is used as an evaluation index to determine the optimal number of clusters K (e.g., K=4).
[0070] For example, the optimal number of clusters K can be determined using the following formula 4.
[0071] Formula 4 Where a is the average distance between a certain feature data and other feature data in its cluster; b is the average distance between a certain feature data and feature data in other clusters.
[0072] Ultimately, the historical operating conditions were divided into four clusters, named "Summer High Load", "Winter Normal Load", "Nighttime Low Load" and "Special Operation and Maintenance Mode". In some embodiments, clustering feature data and determining the clustering result of the feature data includes: calculating the distance between feature data points and the center points of multiple clusters; iteratively updating the cluster to which the feature data points belong and the center points of multiple clusters according to the nearest neighbor principle until the distance is less than a distance threshold, and determining the clustering result.
[0073] For example, the iterative process is as follows: First, randomly initialize K cluster centers; second, for each historical working condition data point, calculate its Euclidean distance to all cluster centers and assign it to the nearest cluster; then recalculate the mean of each cluster as the new cluster center; repeat the distance step and the assignment step until the position change of the cluster center is less than a preset distance threshold (such as 0.001), then the clustering is confirmed to be complete. S402. For each of the multiple operating condition categories, select the operating conditions belonging to the operating condition category from the historical operating conditions.
[0074] For example, taking "high summer load" as an example, all data belonging to this cluster are filtered out from the historical total data. This data usually has similar characteristics, such as high ambient temperature, continuous rack power operation above 80% load, and high air conditioning cooling capacity settings. S403. Based on the operating conditions and the physical layout of the target computer room, determine the dataset of the neural network model corresponding to the operating condition category.
[0075] One working condition category corresponds to one neural network model; the dataset includes the physical boundary feature map of the target computer room, the second initial temperature field, and the actual temperature field.
[0076] In some embodiments, the operating conditions of the operating condition category include the power of the second cabinet and the setting parameters of the second air conditioner.
[0077] As a possible implementation of the above S403, a second initial temperature field is generated based on the power of the second cabinet and the setting parameters of the second air conditioner; a physical boundary feature map is generated based on the physical layout of the target computer room; the physical boundary feature map is used to characterize the relative positional relationship between any point in the target computer room space and the nearest physical boundary; the physical boundary feature map, the second initial temperature field and the actual temperature field corresponding to the operating conditions in time are paired to generate a dataset.
[0078] For example, for the operating condition record of the "summer high load" category, the second cabinet power and the second air conditioner setting parameters in the record are used to generate a second initial temperature field. It should be noted that the method of generating the second initial temperature field is the same as that of generating the first initial temperature field, as detailed in S103 above, and will not be repeated here.
[0079] Furthermore, the actual temperature field can be derived from high-fidelity CFD simulation results corresponding to the recorded moment, or from a measured three-dimensional temperature field obtained by deploying a large number of sensors and using the Delaunay tetrahedral interpolation method, specifically as follows: Figure 7 As shown.
[0080] Finally, the physical boundary feature map and the second initial temperature field are used as input pairs, and the actual temperature field is used as the label to form training samples. S404. Iteratively train the neural network model based on the dataset until the loss value of the neural network model is lower than the preset threshold, and use the neural network model as the prediction model for the temperature field of the computer room under the working condition category.
[0081] As a possible implementation of the above S404, the physical boundary feature map and the second initial temperature field are input into the neural network model to obtain the temperature field prediction result; the difference between the temperature field prediction result and the actual temperature field in the dataset is calculated to obtain the loss value of the current iteration; based on the loss value, the parameters of the neural network model are adjusted until the loss value is lower than the preset threshold, and the neural network model is used as the computer room temperature field prediction model.
[0082] For example, the neural network model can employ a Transformer architecture that includes a self-attention mechanism. During training, the Fisher matrix is used to compute the regularization loss function.
[0083] Formula 5 in, Let the loss function be the current task. Used to set the correlation between old and new tasks. These are the model parameters generated during the current network training. These are the model parameters generated from previous network training.
[0084] Furthermore, the Adam optimizer is used for backpropagation to iteratively update the model weights. When the average MAE loss on the validation set remains below 0.5°C (a preset threshold), training is stopped and the model parameters are saved. The resulting model is a data center temperature field prediction model specifically designed for predicting "high-load summer" operating conditions. The technical solution provided in this application provides at least the following beneficial effects: Multiple operating condition categories are determined based on historical operating conditions, enabling a systematic summary of complex historical operating patterns and laying the foundation for building specialized models; then, specific operating condition data is selected for each operating condition category, ensuring the relevance of training data and providing a high-quality data source for training high-precision specialized models; based on this, a dataset containing physical boundary feature maps, initial temperature fields, and actual temperature fields is constructed for each operating condition category, providing strong constraints and clear objectives for model learning; finally, the dataset is used to iteratively train each neural network model, ensuring that the multiple models trained can accurately capture the temperature field characteristics under different operating conditions, thereby significantly improving the overall accuracy and adaptability of the prediction model.
[0085] like Figure 8 As shown, the overall process of the computer room temperature field prediction method provided in the embodiments of this application is introduced.
[0086] First, the current operating conditions are obtained, and based on the current operating conditions and the physical layout of the target data center, input parameters (such as the first initial temperature field) are generated. Second, based on the current operating conditions of the target data center, the operating condition category of the target data center is determined. According to the operating condition category, a temperature prediction model is determined, and then the input parameters are input into the temperature prediction model to obtain the predicted temperature field of the target data center.
[0087] Based on this, the predicted temperature field is compared with the actual temperature field. If the error between the predicted temperature field and the actual temperature field of the target computer room is greater than the error threshold, the temperature prediction model is incrementally learned based on the current operating conditions.
[0088] The above describes the overall process of the computer room temperature field prediction method provided in the embodiments of this application.
[0089] The following provides a detailed overview of the training process for the computer room temperature field prediction model provided in this application embodiment, combined with... Figure 8 As shown.
[0090] First, historical operating conditions are acquired, and multiple operating condition categories are determined based on these conditions for the target data center. Second, for each operating condition category, operating conditions belonging to that category are selected from the historical data. Based on these operating conditions and the physical layout of the target data center, the dataset for the neural network model corresponding to each operating condition category (such as physical boundary feature maps, a second initial temperature field, and the actual temperature field) is determined. Then, the neural network model is iteratively trained using this dataset until its loss value is below a preset threshold. This neural network model is then used as the data center temperature field prediction model for each operating condition category.
[0091] In an exemplary embodiment, this application also provides a data center temperature field prediction device. Figure 9 This is a schematic diagram of the structure of the computer room temperature field prediction device provided in an embodiment of this application. Figure 9 As shown, the computer room temperature field prediction device 1000 includes a processing unit 1001 and an acquisition unit 1002.
[0092] In some embodiments, the processing unit 1001 is configured to determine the operating condition category of the target computer room based on its current operating condition; the processing unit 1001 is also configured to generate input parameters based on the current operating condition and the physical layout of the target computer room; the input parameters include a physical boundary feature map and a first initial temperature field; the processing unit 1001 is also configured to determine the corresponding temperature prediction model according to the operating condition category; one operating condition category corresponds to one temperature prediction model; the processing unit 1001 is also configured to input the input parameters into the temperature prediction model to obtain the predicted temperature field of the target computer room.
[0093] In some embodiments, the processing unit 1001 is specifically used to acquire feature data of the current operating condition; calculate the distance between the feature data and the central feature data of multiple clusters; and take the operating condition category corresponding to the cluster with the closest distance among the multiple clusters as the operating condition category to which the target computer room belongs; wherein, one cluster corresponds to one operating condition category.
[0094] In some embodiments, the current operating conditions include the power of the first cabinet and the setting parameters of the first air conditioner; the processing unit 1001 is specifically used to generate a first initial temperature field based on the power of the first cabinet and the setting parameters of the first air conditioner; and to generate a physical boundary feature map based on the physical layout of the target computer room; the physical boundary feature map is used to characterize the relative positional relationship between any point in the target computer room space and the nearest physical boundary.
[0095] In some embodiments, the processing unit 1001 is further configured to perform incremental learning on the temperature prediction model based on the current operating conditions when the error between the predicted temperature field and the actual temperature field of the target computer room is greater than the error threshold.
[0096] In an exemplary embodiment, this application also provides a training device for a computer room temperature field prediction model. Figure 10 A schematic diagram of the structure of a training device for a computer room temperature field prediction model provided in this application embodiment is shown below. Figure 10 As shown, the training device 2000 for the computer room temperature field prediction model includes a processing unit 2001 and an acquisition unit 2002.
[0097] In some embodiments, the processing unit 2001 is configured to determine multiple operating condition categories based on the historical operating conditions of the target data center; the processing unit 2001 is further configured to filter out operating conditions belonging to each of the multiple operating condition categories from the historical operating conditions; the processing unit 2001 is further configured to determine the dataset of the neural network model corresponding to the operating condition category based on the operating conditions and the physical layout of the target data center; one operating condition category corresponds to one neural network model; the dataset includes the physical boundary feature map of the target data center, the second initial temperature field, and the actual temperature field; the processing unit 2001 is further configured to iteratively train the neural network model according to the dataset until the loss value of the neural network model is lower than a preset threshold, and use the neural network model as the data center temperature field prediction model for the operating condition category.
[0098] In some embodiments, the operating conditions of the operating condition category include the power of the second cabinet and the setting parameters of the second air conditioner; the processing unit 2001 is specifically used to generate a second initial temperature field based on the power of the second cabinet and the setting parameters of the second air conditioner; generate a physical boundary feature map based on the physical layout of the target computer room; the physical boundary feature map is used to characterize the relative positional relationship between any point in the target computer room space and the nearest physical boundary; and pair the physical boundary feature map, the second initial temperature field and the actual temperature field corresponding to the operating condition in time to generate a dataset.
[0099] In some embodiments, the processing unit 2001 is specifically used to input the physical boundary feature map and the second initial temperature field into the neural network model to obtain the temperature field prediction result; calculate the difference between the temperature field prediction result and the actual temperature field in the dataset to obtain the loss value of the current iteration; and adjust the parameters of the neural network model based on the loss value until the loss value is lower than a preset threshold, and use the neural network model as the computer room temperature field prediction model.
[0100] In some embodiments, the processing unit 2001 is specifically used to acquire feature data of historical operating conditions; to perform clustering on the feature data and determine the clustering result of the feature data; the clustering result includes multiple clusters; one cluster corresponds to one operating condition category; and multiple operating condition categories are determined based on multiple clusters.
[0101] In some embodiments, the processing unit 2001 is specifically used to calculate the distance between the feature data point and the center point of multiple clusters; iteratively update the cluster to which the feature data point belongs and the center point of multiple clusters according to the nearest neighbor principle until the distance is less than the distance threshold, and determine the clustering result.
[0102] Of course, the computer room temperature field prediction device 1000 and the computer room temperature field prediction training device 2000 include, but are not limited to, the unit modules listed above. Furthermore, the specific functions that the above-mentioned functional units can implement include, but are not limited to, the functions corresponding to the method steps in the above embodiments. For detailed descriptions of other modules of the computer room temperature field prediction device 1000 and the computer room temperature field prediction device 2000, please refer to the detailed descriptions of their corresponding method steps; these will not be repeated here in the embodiments of this application.
[0103] In an exemplary embodiment, this application also provides a computer program product that, when run on a computer, causes the computer to execute the aforementioned related method steps to implement the computer room temperature field prediction method in the above embodiments.
[0104] In an exemplary embodiment, this application also provides an electronic device. Figure 11 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 11 As shown, the electronic device may include a processor 1101 and a memory 1102; the memory 1102 stores instructions executable by the processor 1101; when the processor 1101 is configured to execute the instructions, the electronic device implements the method as described in the foregoing method embodiments.
[0105] In an exemplary embodiment, this application also provides a computer-readable storage medium storing computer program instructions thereon; when the computer program instructions are executed by an electronic device, the electronic device performs the method as described in the foregoing embodiments. The computer-readable storage medium may be a non-transitory computer-readable storage medium, such as a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device.
[0106] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented using software programs, implementation can be, in whole or in part, in the form of a computer program product. This computer program product includes one or more computer-executable instructions. When these computer-executable instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer-executable instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer-executable instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device containing one or more servers, data centers, etc., that can be integrated with the medium. The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state disks, SSDs).
[0107] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, disclosure, and appended claims, will understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple instances. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.
[0108] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of this application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from the spirit and scope of this application. Thus, if such modifications and modifications of this application fall within the scope of the claims of this application and their equivalents, this application is also intended to include such modifications and modifications.
[0109] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A method for predicting the temperature field of a computer room, characterized in that, The method includes: Based on the current operating conditions of the target data center, determine the operating condition category to which the target data center belongs; Based on the current operating conditions and the physical layout of the target computer room, input parameters are generated; the input parameters include a physical boundary feature map and a first initial temperature field. Based on the aforementioned operating condition categories, a corresponding temperature prediction model is determined; one operating condition category corresponds to one temperature prediction model. The input parameters are input into the temperature prediction model to obtain the predicted temperature field of the target computer room.
2. The method according to claim 1, characterized in that, The determination of the operating condition category of the target data center based on its current operating condition includes: Obtain the characteristic data of the current operating condition; Calculate the distance between the feature data and the central feature data of multiple clusters; The operating condition category of the cluster that is closest to the target data center is taken as the operating condition category of the target data center. Each cluster corresponds to a working condition category.
3. The method according to claim 1, characterized in that, The current operating conditions include the power of the first cabinet and the setting parameters of the first air conditioner; The input parameters generated based on the current operating conditions and the physical layout of the target data center include: The first initial temperature field is generated based on the power of the first cabinet and the setting parameters of the first air conditioner; Based on the physical layout of the target computer room, a physical boundary feature map is generated; the physical boundary feature map is used to characterize the relative positional relationship of any point in the target computer room space to the nearest physical boundary.
4. The method according to claim 1 or 2, characterized in that, The method further includes: If the error between the predicted temperature field and the actual temperature field in the target computer room is greater than the error threshold, the temperature prediction model is incrementally learned based on the current operating conditions.
5. A training method for a computer room temperature field prediction model, characterized in that, The method includes: Based on the historical operating conditions of the target data center, multiple operating condition categories are determined; For each of the multiple operating condition categories, operating conditions belonging to that category are selected from the historical operating conditions. Based on the operating conditions and the physical layout of the target computer room, a dataset of neural network models corresponding to the operating condition categories is determined; one operating condition category corresponds to one neural network model; the dataset includes the physical boundary feature map of the target computer room, the second initial temperature field, and the actual temperature field; The neural network model is iteratively trained based on the dataset until the loss value of the neural network model is lower than a preset threshold, and then the neural network model is used as the computer room temperature field prediction model for the operating condition category.
6. The method according to claim 5, characterized in that, The operating conditions of the aforementioned operating condition category include the power of the second cabinet and the setting parameters of the second air conditioner; The dataset for determining the neural network model corresponding to the operating condition category based on the operating conditions and the physical layout of the target data center includes: The second initial temperature field is generated based on the power of the second cabinet and the setting parameters of the second air conditioner; Based on the physical layout of the target computer room, a physical boundary feature map is generated; the physical boundary feature map is used to characterize the relative positional relationship of any point in the target computer room space to the nearest physical boundary; The dataset is generated by pairing the physical boundary feature map, the second initial temperature field, and the actual temperature field corresponding to the operating condition in time.
7. The method according to claim 5 or 6, characterized in that, The step of iteratively training the neural network model based on the dataset until the loss value of the neural network model is lower than a preset threshold, and then using the neural network model as the data center temperature field prediction model for the operating condition category, includes: The physical boundary feature map and the second initial temperature field are input into the neural network model to obtain the temperature field prediction result; The difference between the predicted temperature field and the actual temperature field in the dataset is calculated to obtain the loss value for the current iteration. Based on the loss value, the parameters of the neural network model are adjusted until the loss value is lower than the preset threshold, and the neural network model is used as the prediction model for the computer room temperature field.
8. The method according to claim 5, characterized in that, Based on the historical operating conditions of the target data center, multiple operating condition categories are determined, including: Obtain the characteristic data of the historical operating conditions; The feature data is clustered to determine the clustering results; the clustering results include multiple clusters; each cluster corresponds to a working condition category; The multiple operating condition categories are determined based on the multiple clusters.
9. The method according to claim 8, characterized in that, The step of clustering the feature data and determining the clustering results of the feature data includes: Calculate the distance between the feature data points and the center points of multiple clusters; The clustering result is determined by iteratively updating the cluster to which the feature data point belongs and the center point of the multiple clusters according to the nearest neighbor principle until the distance is less than the distance threshold.
10. An electronic device, characterized in that, A processor and a memory; the memory is used to store computer programs or instructions, and the processor is used to run the computer programs or instructions to implement the computer room temperature field prediction method as described in any one of claims 1-4, or to implement the computer room temperature field prediction model training method as described in any one of claims 5-9.
11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed by a computer, enable the computer to perform the computer room temperature field prediction method according to any one of claims 1-4, or to implement the training method for the computer room temperature field prediction model according to any one of claims 5-9.
12. A computer program product, characterized in that, The computer program product includes instructions that, when executed on a computer, enable the computer to perform the room temperature field prediction method as described in any one of claims 1-4, or to implement the training method for the room temperature field prediction model as described in any one of claims 5-9.