A transformer-based data center energy-saving control method

Abstract

This invention provides a Transformer-based energy-saving control method for data centers, comprising the following steps: Step S10, acquiring operational data of the data center, including cold storage indicators, terminal equipment indicators, and environmental indicators; Step S20, preprocessing the acquired operational data; Step S30, extracting features from the preprocessed operational data to construct a dataset; Step S40, creating a PUE prediction model based on Transformer, and training the PUE prediction model using the dataset; Step S50, publishing the trained PUE prediction model, performing online optimization of the index parameters of the published PUE prediction model, and controlling the operation of the data center based on the optimized index parameters. The advantage of this invention is that it significantly improves the energy-saving effect of data centers.

Description

Technical Field

[0001] This invention relates to the field of data center control technology, and in particular to a data center energy-saving control method based on Transformer. Background Technology

[0002] The widespread adoption of cloud computing and its rapid growth in scale have spurred the development of large-scale data centers. The operation and maintenance of these massive data centers are inseparable from excessive electricity consumption. For data center operators, data centers consume enormous amounts of energy, with electricity costs accounting for a significant portion of operating costs. Reducing energy consumption to save on electricity costs is equivalent to increasing corporate profits. Therefore, how to promote effective energy-saving practices through artificial intelligence has become an important goal that companies continue to focus on.

[0003] Traditionally, there are two main approaches to energy-saving control in data centers: 1. Product-level energy-saving technologies, such as PID control for air conditioning cooling systems. This method uses proportional, integral, and derivative calculations to optimize control quantities. While simple to design and inexpensive, it is difficult to achieve ideal operating results through simple proportional control. In practical applications, the conditions of each data center's cooling station, terminal equipment, service load, and environmental changes vary greatly. This method cannot consider the interaction and comprehensive impact between the cooling station and the environment. 2. Artificial intelligence optimization technology, namely the neural network models used in industry, such as those by Google. This involves building AI prediction models based on historical data on the environment, air conditioning, and energy consumption, and then using these AI prediction models to adjust the operation of the data center. However, most current AI prediction models, facing complex environmental changes, specialized cooling systems, and numerous control parameters, simply input all the available indicators without analyzing the coupling characteristics between indicators or considering the interference analysis of irrelevant indicators. This makes it difficult to achieve the optimal long-term PUE (Power Usage Effectiveness) close to 1.

[0004] Therefore, how to provide a Transformer-based energy-saving control method for data centers to improve energy efficiency has become an urgent technical problem to be solved. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide a data center energy-saving control method based on Transformer, so as to improve the energy-saving effect of data centers.

[0006] This invention is implemented as follows: A data center energy-saving control method based on Transformer, comprising the following steps:

[0007] Step S10: Obtain operational data for the data center, including cold storage indicators, terminal equipment indicators, and environmental indicators;

[0008] Step S20: Preprocess the acquired running data;

[0009] Step S30: Extract features from the preprocessed running data to construct a dataset;

[0010] Step S40: Create a PUE prediction model based on Transformer, and train the PUE prediction model using the dataset;

[0011] Step S50: Publish the trained PUE prediction model, perform online tuning of the index parameters of the published PUE prediction model, and control the operation of the data center based on the tuned index parameters.

[0012] Further, in step S10, the chiller plant indicators include at least the number of operating cooling towers, cooling tower cooling water outlet temperature, number of operating chilled water pumps, number of operating cooling water pumps, chilled water main supply pressure, cooling water main supply pressure, cooling water main return temperature, wet bulb temperature, number of operating chiller units, chiller unit cooling water return temperature, chiller unit evaporator temperature difference, chiller unit chilled water flow rate, total active power input to chilled pumps, total active power input to cooling pumps, cooling water main return pressure, cooling water main return temperature, chilled water main flow rate, chiller unit current percentage, chiller unit chilled water outlet temperature, chiller unit condenser temperature difference, chiller unit output frequency, and cooling tower fan frequency conversion feedback.

[0013] The specifications for the terminal equipment include at least the number of water pumps, the number of CTs, the number of CHs, the frequency, and the outlet water temperature.

[0014] The environmental indicators include at least IT load rate, outdoor temperature, outdoor humidity, and outdoor wind speed.

[0015] Further, step S20 specifically includes:

[0016] First, the obtained running data is standardized using the z-score method to unify the units. Then, duplicate running data is removed using the deletion method. Next, abnormal running data is removed using the quartile method. Then, missing data in the running data is filled in using the mean of adjacent n positions. Finally, the parameters of the running data are merged using the index normalization method to complete the preprocessing of the running data.

[0017] Further, step S30 specifically includes:

[0018] First, the characteristic indicators related to PUE in the preprocessed operating data are extracted using the variance method;

[0019] Next, the collinearity relationship between the feature indicators is extracted using the Pearson correlation coefficient method, and feature indicators with strong collinearity are selected based on the collinearity relationship.

[0020] Next, the dimensionality of each of the strongly collinear feature indices is reduced using the PCA method to remove the strongly collinear feature indices.

[0021] Finally, a dataset is constructed based on the remaining feature indicators.

[0022] Furthermore, in step S40, the PUE prediction model includes a DSSM multi-tower network, a Transformer network, and a DNN output network.

[0023] The DSSM multi-tower network includes a cold storage tower sub-model created based on Embedding and DNN, and an end-device tower model created based on Embedding and DNN.

[0024] The input of the Transformer network is connected to the output of the cooling plant tower model and the output of the terminal equipment tower model, and the output is connected to the input of the DNN output network.

[0025] Further, in step S40, the loss function of the PUE prediction model is the mean squared error (RMSE) to evaluate the error between the predicted value and the actual value.

[0026] During the training process of the PUE prediction model, the index parameters of the PUE prediction model are optimized by incorporating the Warmup strategy of the AdamW optimizer, so as to minimize the loss value of the loss function.

[0027] Further, step S50 specifically includes:

[0028] The trained PUE prediction model is published to the server. The ant colony algorithm is used to fine-tune the index parameters of the published PUE prediction model online to bring the PUE to the optimal value. The fine-tuned index parameters are filtered in multiple layers based on the SLA protocol. The operation of the data center is controlled based on the filtered index parameters.

[0029] The advantages of this invention are:

[0030] 1. By acquiring operational data including cold storage indicators, terminal equipment indicators, and environmental indicators, the operational data is preprocessed and features are extracted to construct a dataset. The dataset is then used to train a PUE prediction model based on Transformer. The trained PUE prediction model is then released and optimized online. Finally, the optimized indicator parameters are used to control the operation of the data center. This approach fully considers cold storage indicators, terminal equipment indicators, and environmental indicators. The PUE prediction model has independent cold storage tower models and terminal equipment tower models to reduce the interference characteristics of unrelated indicators. At the same time, the coupling characteristics of related indicators are captured through the Transformer network, so that the trained and optimized PUE prediction model can output the best indicator parameters to control the operation of the data center, thereby greatly improving the energy-saving effect of the data center.

[0031] 2. By preprocessing the running data to unify the units, remove duplicate data, delete abnormal data, fill in missing data, and merge parameters, and then extracting features and reducing dimensionality of the preprocessed running data to construct a dataset, the dataset is greatly standardized and the amount of useless data in the dataset is reduced. This not only reduces the training pressure of the PUE prediction model, but also effectively improves the prediction accuracy of the PUE prediction model. Attached Figure Description

[0032] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0033] Figure 1 This is a flowchart of a data center energy-saving control method based on Transformer according to the present invention.

[0034] Figure 2 This is a schematic diagram of the PUE prediction model of the present invention. Detailed Implementation

[0035] The overall idea of ​​the technical solution in this application embodiment is as follows: By acquiring operational data including cold station indicators, terminal equipment indicators, and environmental indicators, and fully considering the correlation and comprehensive impact among cold station indicators, terminal equipment indicators, and environmental indicators, independent cold station tower models and terminal equipment tower models are set up to reduce the interference characteristics of unrelated indicators. At the same time, the coupling characteristics of related indicators are captured through the Transformer network, so that the trained and optimized PUE prediction model can output the best indicator parameters to control the operation of the data center, thereby improving the energy-saving effect of the data center.

[0036] Please refer to Figures 1 to 2 As shown, a preferred embodiment of the data center energy-saving control method based on Transformer of the present invention includes the following steps:

[0037] Step S10: Obtain operational data for the data center, including cold storage indicators, terminal equipment indicators, and environmental indicators;

[0038] Step S20: Preprocess the acquired running data using automated processing tools; that is, perform preprocessing such as cleaning, noise reduction, and normalization on the running data.

[0039] Step S30: Use feature processing tools to extract features from the preprocessed running data to construct a dataset;

[0040] Step S40: Create a PUE prediction model based on Transformer, and train the PUE prediction model using the dataset;

[0041] Step S50: Publish the trained PUE prediction model, perform online tuning of the index parameters of the published PUE prediction model, and control the operation of the data center based on the tuned index parameters.

[0042] In step S10, the chiller plant indicators include at least the number of operating cooling towers, cooling tower cooling water outlet temperature, number of operating chilled water pumps, number of operating cooling water pumps, chilled water main supply pressure, cooling water main supply pressure, cooling water main return temperature, wet bulb temperature, number of operating chiller units, chiller unit cooling water return temperature, chiller unit evaporator temperature difference, chiller unit chilled water flow rate, total active power input to chilled pumps, total active power input to cooling pumps, cooling water main return pressure, cooling water main return temperature, chilled water main flow rate, chiller unit current percentage, chilled water outlet temperature of chiller units, chiller unit condenser temperature difference, chiller unit output frequency, and cooling tower fan frequency conversion feedback.

[0043] The specifications for the terminal equipment include at least the number of water pumps, the number of CTs, the number of CHs, the frequency, and the outlet water temperature.

[0044] The environmental indicators include at least IT load rate, outdoor temperature, outdoor humidity, and outdoor wind speed.

[0045] Step S20 specifically involves:

[0046] First, the obtained running data is standardized using the z-score method to unify the units. Then, duplicate running data is removed using the deletion method. Next, abnormal running data is removed using the quartile method. Then, missing data in the running data is filled in using the mean of adjacent n positions. Finally, the parameters of the running data are merged using the index normalization method to complete the preprocessing of the running data.

[0047] Step S30 specifically involves:

[0048] First, the characteristic indicators related to PUE in the preprocessed operating data are extracted using the variance method;

[0049] Next, the collinearity relationship between the feature indicators is extracted using the Pearson correlation coefficient method, and feature indicators with strong collinearity are selected based on the collinearity relationship.

[0050] Next, the dimensionality of each of the strongly collinear feature indices is reduced using the PCA method to remove the strongly collinear feature indices.

[0051] Finally, a dataset is constructed based on the remaining feature indicators; by reducing the dimensionality of the feature indicators, the workload of modeling the PUE prediction model can be reduced.

[0052] In step S40, the PUE prediction model includes a DSSM multi-tower network, a Transformer network, and a DNN output network; the DNN output network is used to convert the output of the Transformer network into a predicted value of PUE energy efficiency.

[0053] The DSSM multi-tower network includes a cold storage tower sub-model created based on Embedding and DNN, and an end-device tower model created based on Embedding and DNN.

[0054] The input of the Transformer network is connected to the output of the cooling plant tower model and the output of the terminal equipment tower model, and the output is connected to the input of the DNN output network.

[0055] Data centers collect numerous metrics from their cooling stations and terminal equipment, and these metrics exhibit complex coupling relationships and interference characteristics. Traditional control methods, such as PID control, struggle to accurately describe these complex coupling relationships, while traditional DNN models are ineffective at eliminating these complex interference characteristics. This invention's PUE prediction model utilizes a DSSM multi-tower network to create separate sub-models for cooling stations and terminal equipment, reducing interference from irrelevant metrics. Simultaneously, it integrates a Transformer network to capture the complex coupling relationships between metrics, effectively eliminating interference and describing coupling characteristics. Through supervised learning, it fits the energy consumption of cooling stations and terminal equipment to accurately establish a nonlinear mapping relationship between input metrics and output PUE.

[0056] The Transformer network is used to simultaneously perform self-attention and cross-attention calculations on the cold storage tower model and the terminal equipment tower model, outputting X. att It captures the complex coupling relationships between indicators and learns their correlation with global features.

[0057] Output X' of the cold station tower model 冷站 The formula for calculating self-attention is as follows:

[0058]

[0059]

[0060] Terminal equipment tower model output X' 末端 The formula for calculating self-attention is as follows:

[0061]

[0062]

[0063] Output X' of the cold station tower model 冷站 and the output X' of the terminal equipment tower model 末端 The formula for calculating cross-attention is as follows:

[0064]

[0065]

[0066] Finally, it is transformed into X through weighted sum calculation. att :c i =∑α ij h ij .

[0067] In step S40, the loss function of the PUE prediction model is the mean squared error (RMSE) to evaluate the error between the predicted value and the actual value.

[0068] During the training process of the PUE prediction model, the index parameters of the PUE prediction model are optimized by incorporating the Warmup strategy of the AdamW optimizer, so as to minimize the loss value of the loss function.

[0069] Step S50 specifically involves:

[0070] The trained PUE prediction model is published to the server. The ant colony algorithm is used to fine-tune the index parameters of the published PUE prediction model online to bring the PUE to the optimal value. The fine-tuned index parameters are filtered in multiple layers based on the SLA protocol. The operation of the data center is controlled based on the filtered index parameters.

[0071] This involves joint modeling of the cooling plant and terminal equipment, considering both the interference characteristics of unrelated indicators and the coupling characteristics between related indicators, while also optimizing all indicator parameters and comprehensively considering the impact of all indicator parameters on PUE.

[0072] In summary, the advantages of this invention are:

[0073] 1. By acquiring operational data including cold storage indicators, terminal equipment indicators, and environmental indicators, the operational data is preprocessed and features are extracted to construct a dataset. The dataset is then used to train a PUE prediction model based on Transformer. The trained PUE prediction model is then released and optimized online. Finally, the optimized indicator parameters are used to control the operation of the data center. This approach fully considers cold storage indicators, terminal equipment indicators, and environmental indicators. The PUE prediction model has independent cold storage tower models and terminal equipment tower models to reduce the interference characteristics of unrelated indicators. At the same time, the coupling characteristics of related indicators are captured through the Transformer network, so that the trained and optimized PUE prediction model can output the best indicator parameters to control the operation of the data center, thereby greatly improving the energy-saving effect of the data center.

[0074] 2. By preprocessing the running data to unify the units, remove duplicate data, delete abnormal data, fill in missing data, and merge parameters, and then extracting features and reducing dimensionality of the preprocessed running data to construct a dataset, the dataset is greatly standardized and the amount of useless data in the dataset is reduced. This not only reduces the training pressure of the PUE prediction model, but also effectively improves the prediction accuracy of the PUE prediction model.

[0075] While specific embodiments of the present invention have been described above, those skilled in the art should understand that the specific embodiments described are merely illustrative and not intended to limit the scope of the present invention. Equivalent modifications and variations made by those skilled in the art in accordance with the spirit of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A data center energy-saving control method based on Transformer, characterized in that: Includes the following steps: Step S10: Obtain operational data for the data center, including cold storage indicators, terminal equipment indicators, and environmental indicators; Step S20: Preprocess the acquired running data; Step S30: Extract features from the preprocessed running data to construct a dataset; Step S40: Create a PUE prediction model based on Transformer, and train the PUE prediction model using the dataset; Step S50: Publish the trained PUE prediction model, perform online tuning of the index parameters of the published PUE prediction model, and control the operation of the data center based on the tuned index parameters.

2. The data center energy-saving control method based on Transformer as described in claim 1, characterized in that: In step S10, the chiller plant indicators include at least the number of operating cooling towers, cooling tower cooling water outlet temperature, number of operating chilled water pumps, number of operating cooling water pumps, chilled water main supply pressure, cooling water main supply pressure, cooling water main return temperature, wet bulb temperature, number of operating chiller units, chiller unit cooling water return temperature, chiller unit evaporator temperature difference, chiller unit chilled water flow rate, total active power input to chilled pumps, total active power input to cooling pumps, cooling water main return pressure, cooling water main return temperature, chilled water main flow rate, chiller unit current percentage, chilled water outlet temperature of chiller units, chiller unit condenser temperature difference, chiller unit output frequency, and cooling tower fan frequency conversion feedback. The specifications for the terminal equipment include at least the number of water pumps, the number of CTs, the number of CHs, the frequency, and the outlet water temperature. The environmental indicators include at least IT load rate, outdoor temperature, outdoor humidity, and outdoor wind speed.

3. The data center energy-saving control method based on Transformer as described in claim 1, characterized in that: Step S20 specifically involves: First, the obtained running data is standardized using the z-score method to unify the units. Then, duplicate running data is removed using the deletion method. Next, abnormal running data is removed using the quartile method. Then, missing data in the running data is filled in using the mean of adjacent n positions. Finally, the parameters of the running data are merged using the index normalization method to complete the preprocessing of the running data.

4. The data center energy-saving control method based on Transformer as described in claim 1, characterized in that: Step S30 specifically involves: First, the characteristic indicators related to PUE in the preprocessed operating data are extracted using the variance method; Next, the collinearity relationship between the feature indicators is extracted using the Pearson correlation coefficient method, and feature indicators with strong collinearity are selected based on the collinearity relationship. Next, the dimensionality of each of the strongly collinear feature indices is reduced using the PCA method to remove the strongly collinear feature indices. Finally, a dataset is constructed based on the remaining feature indicators.

5. The data center energy-saving control method based on Transformer as described in claim 1, characterized in that: In step S40, the PUE prediction model includes a DSSM multi-tower network, a Transformer network, and a DNN output network. The DSSM multi-tower network includes a cold storage tower sub-model created based on Embedding and DNN, and an end-device tower model created based on Embedding and DNN. The input of the Transformer network is connected to the output of the cooling plant tower model and the output of the terminal equipment tower model, and the output is connected to the input of the DNN output network.

6. The data center energy-saving control method based on Transformer as described in claim 1, characterized in that: In step S40, the loss function of the PUE prediction model is the mean squared error (RMSE) to evaluate the error between the predicted value and the actual value. During the training process of the PUE prediction model, the index parameters of the PUE prediction model are optimized by incorporating the Warmup strategy of the AdamW optimizer, so as to minimize the loss value of the loss function.

7. The data center energy-saving control method based on Transformer as described in claim 1, characterized in that: Step S50 specifically involves: The trained PUE prediction model is published to the server. The ant colony algorithm is used to fine-tune the index parameters of the published PUE prediction model online to bring the PUE to the optimal value. The fine-tuned index parameters are filtered in multiple layers based on the SLA protocol. The operation of the data center is controlled based on the filtered index parameters.

Images