Data center rack power management method and system

Abstract

The invention relates to the technical field of power management, in particular to a data center rack power management method and system, and the method comprises the following steps: collecting the CPU utilization rate, the memory utilization rate and the rack temperature of multiple racks of a data center, measuring the power output of each device through a rack power unit, and carrying out the integration to obtain energy consumption basic data. According to the method, the CPU utilization rate, the memory utilization rate, the rack temperature and other data of the multiple racks of the data center are integrated, the power output of each device is accurately measured, the comprehensive energy consumption basic data view is provided, and through data integration and weight analysis, energy distribution is more accurate, dynamic adjustment can be conducted according to the actual running state of the device, and the energy consumption is reduced. The energy utilization efficiency is improved, the neural network model is applied to predict energy demands and optimize power output and cooling system setting, accurate matching and efficiency maximization of power supply are ensured, prediction deviation is effectively corrected through real-time monitoring and application of the sliding window method, and excessive waste of resources is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of power management, and in particular to a data center rack power management method and system. Background Art

[0002] The field of power management technology is the study and development of various methods and devices for efficiently managing the use of electrical energy. This includes, but is not limited to, improving the stability, efficiency, and safety of power supplies, especially in environments that require intensive energy consumption, such as data centers. Within this technology area, the focus includes optimizing power supply design, reducing energy waste, improving energy conversion efficiency, and monitoring and adjusting energy consumption. Power management technology is critical to ensuring reliable operation of equipment and reducing operating costs, especially in high-energy consumption environments such as data centers.

[0003] Among them, data center rack power management methods refer to a series of technologies and strategies for managing and optimizing rack power configuration and power distribution within the data center. The purpose of this management method is to ensure that all servers and network equipment can receive stable and efficient power supply, while reducing power waste and improving energy efficiency throughout the system. By precisely controlling the power supply of each device, the operating efficiency and reliability of the data center can be significantly improved, while reducing the equipment failure rate caused by power problems. This approach usually includes the use of intelligent power distribution units (PDUs), environmental monitoring tools, and other advanced power technologies to achieve fine-grained management of data center power.

[0004] Existing power management technologies in data center applications usually configure power based on the maximum load demand of the equipment without considering the actual operating status, resulting in over-configuration and energy waste. In addition, the lack of real-time analysis and optimization adjustment capabilities makes the data center less flexible when demand changes, and unable to adjust resource configuration in real time to cope with actual power and cooling needs. Existing methods also lack accurate data analysis and prediction tools, making it difficult to achieve optimal energy configuration, which may lead to insufficient resources during peak hours and energy waste during low loads, affecting economic benefits and environmental effects. Summary of the invention

[0005] The purpose of the present invention is to solve the shortcomings in the prior art and to propose a data center rack power management method and system.

[0006] In order to achieve the above object, the present invention adopts the following technical solution: a data center rack power management method, comprising the following steps:

[0007] S1: Collect the CPU usage, memory usage, and rack temperature of multiple racks in the data center, measure the power output of each device through the rack power unit, integrate to obtain basic energy consumption data, calculate the information entropy and critical weight of each data item, and obtain the weight calculation result;

[0008] S2: using the weight calculation result, adjusting the weight for the data input, setting the structure of the neural network including the configuration of the activation function and the power distribution unit, starting the network for iterative training until the parameters converge, and obtaining a training optimization model;

[0009] S3: input the real-time monitoring data into the training optimization model, use the power monitoring system to predict the short-term energy demand peak, use the sliding window method to smooth the prediction results, correct the model prediction deviation, and output the energy demand prediction results;

[0010] S4: according to the energy demand prediction result, apply linear programming to optimize the power output and cooling pipeline settings of the rack, calculate the optimal rack power configuration, adjust the power output and cooling pipeline settings, and generate a dynamic power configuration plan;

[0011] S5: Implement the dynamic power configuration solution, and continuously track energy utilization efficiency through the power monitoring system, record changes in power usage and cooling efficiency, perform deviation analysis compared with expected targets, update the power configuration, and obtain efficiency monitoring results.

[0012] The weight calculation results specifically include information entropy and critical weights; the training optimization model specifically includes network structure, activation function configuration, and power distribution unit settings; the energy demand forecasting results include short-term energy demand peak and forecast correction deviation; the dynamic power supply configuration scheme specifically includes power output optimization and cooling pipeline configuration; the efficiency monitoring results specifically refer to energy utilization efficiency, power usage changes, and cooling efficiency.

[0013] As a further solution of the present invention, the step of obtaining the weight calculation result is specifically:

[0014] S111: Collect CPU usage, memory usage, and rack temperature of multiple racks in the data center, measure the power output of each device through the rack power unit, integrate the data, calculate the energy consumption of each data point, and generate basic energy consumption data;

[0015] S112: Based on the basic energy consumption data, information entropy is calculated for each data item, and information entropy data results are obtained by applying statistical analysis;

[0016] S113: Using the information entropy data results, perform a criticality assessment on each data item using the formula:

[0017]

[0018] Process the variability and relative position of data items to obtain the critical weight of each data item, and then generate critical weight data results;

[0019] Among them, w i represents the critical weight of the i-th data item, H i represents the information entropy of the i-th data item, R i Indicates the relative position of the data item in the entire data set and is used to adjust the weight distribution. i is the load factor of the data item, which indicates the usage load or criticality of the data item. α is the coefficient for adjusting the influence of information entropy, which is used to amplify or reduce the influence of information entropy. β is the position inverse adjustment coefficient, which increases the influence of position criticality by adjusting the coefficient. γ is the weight adjustment coefficient of the load factor, which is used to adjust the contribution of the load factor to the weight.

[0020] S114: Integrate the key weight data results to form a weight calculation result.

[0021] As a further solution of the present invention, the step of obtaining the training optimization model is specifically as follows:

[0022] S211: using the weight calculation result, adjusting the input weight of the neural network, setting the node weight of each layer in the network one by one, and then configuring the type and parameters of the activation function to generate an adjusted network configuration;

[0023] S212: Based on the adjusted network configuration, setting a power distribution unit to optimize energy consumption, starting iterative training of a neural network, using a target data set for feedback learning, and generating training process data;

[0024] S213: Using the training process data, real-time monitoring of parameter convergence is performed using the formula:

[0025]

[0026] Calculate parameter adjustments and generate real-time parameter adjustment data C n+1 ;

[0027] Among them, C n represents the parameter value of the current iteration, η is the learning rate, and the square root and absolute value ensure that the adjustment step size will not become too large when the parameter value is extremely small or the gradient change is not significant, thus avoiding over-adjustment. is the loss function with respect to C n The gradient of ,∈ is a small constant to avoid the denominator being zero;

[0028] S214: Utilize the real-time parameter adjustment data and iterate until the change in all parameters is less than a preset threshold, thereby generating a training optimization model.

[0029] As a further solution of the present invention, the steps of obtaining the energy demand forecast result are specifically as follows:

[0030] S311: input the real-time monitoring data into the training optimization model, combine it with the power supply monitoring system, predict the peak energy demand in the short term, and obtain preliminary energy demand prediction data;

[0031] S312: Applying a sliding window method to smooth the preliminary energy demand forecast data to reduce volatility and error of the forecast, thereby obtaining smoothed energy demand data;

[0032] S313: Perform deviation correction on the smoothed energy demand data using the formula:

[0033]

[0034] Calculate and generate energy demand forecast results E corr ;

[0035] Among them, E smooth represents the smoothed energy demand data, E min and E max They represent the minimum and maximum values ​​in the previous energy demand data respectively, and α is an adjustment coefficient used to adjust the forecast deviation according to previous data.

[0036] As a further solution of the present invention, the step of obtaining the dynamic power configuration solution is specifically:

[0037] S411: using the energy demand prediction result, setting the input parameters of the linear programming model, including the objective function and the constraint conditions, and performing the optimal configuration of the rack power and the cooling system to obtain the initial parameters of the rack configuration;

[0038] S412: performing linear programming optimization on the initial parameters of the rack configuration, adjusting the power output and cooling pipeline settings to match the predicted energy demand peak, and calculating the optimized rack configuration parameters;

[0039] S413: According to the optimized rack configuration parameters, the formula is used:

[0040]

[0041] Calculate the adjusted power output and generate a dynamic power configuration scheme P new ;

[0042] Among them, P opt represents the optimized power output, ΔP is the difference between the predicted demand and the current configuration, P max , Pmin are the maximum and minimum estimated demands, β and γ are adjustment coefficients, and ∈ is a small amount added to avoid division by zero.

[0043] As a further solution of the present invention, the steps of obtaining the efficiency monitoring results are specifically as follows:

[0044] S511: Implement the dynamic power configuration solution, continuously track real-time data of power usage and cooling efficiency through a power monitoring system, and generate preliminary performance monitoring data;

[0045] S512: Analyze the preliminary performance monitoring data, compare it with the preset efficiency target, determine the existing deviation, and obtain the deviation analysis result;

[0046] S513: According to the deviation analysis result, the power supply configuration is updated using the formula:

[0047]

[0048] Calculates adjusted power settings and generates efficiency monitoring results;

[0049] Among them, R new Represents the current efficiency result, R current Represents the efficiency data currently recorded, Target is the preset efficiency target, α and β are adjustment coefficients used to adjust the influence of target deviation and the amplitude of scaling deviation.

[0050] A data center rack power management system, the data center rack power management system is used to execute the above data center rack power management method, the system comprises:

[0051] The energy consumption data collection module collects CPU usage, memory usage, rack temperature and power output based on a multi-rack environment to obtain basic energy consumption data;

[0052] The weight calculation module calculates the information entropy of multiple data items based on the energy consumption basic data, extracts the key weights, and generates a weight calculation result;

[0053] The network training module sets the neural network structure based on the weight calculation result, including the activation function and the power distribution unit configuration, performs iterative training until the parameters converge, and generates a training optimization model;

[0054] The demand forecasting module inputs the real-time monitoring data based on the training optimization model, predicts the short-term energy demand peak, applies the sliding window method to smooth the prediction results, corrects the prediction deviation, and obtains the energy demand prediction results;

[0055] The optimization configuration module uses linear programming to optimize the rack power output and cooling pipeline settings based on the energy demand prediction results, calculates the optimal power configuration, and generates a dynamic power configuration plan;

[0056] Based on the dynamic power configuration scheme, the efficiency monitoring module continuously tracks energy utilization efficiency through the power monitoring system, records changes in power usage and cooling efficiency, performs deviation analysis compared with expected targets, updates the power configuration, and obtains efficiency monitoring results.

[0057] Compared with the prior art, the advantages and positive effects of the present invention are:

[0058] In the present invention, by integrating the CPU usage, memory usage, rack temperature and other data of multiple racks in the data center, the power output of each device is accurately measured, providing a comprehensive basic data view of energy consumption. Through this data integration and weight analysis, energy allocation is more accurate, and can be dynamically adjusted according to the actual operating status of the equipment to improve energy efficiency. The use of neural network models predicts energy demand and optimizes power output and cooling system settings, ensuring accurate matching of power supply and maximization of efficiency. The application of real-time monitoring and sliding window method effectively corrects prediction deviations, reduces excessive waste of resources, and improves the overall reliability and stability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is a schematic diagram of the workflow of the present invention;

[0060] Figure 2 A flow chart of the steps for obtaining the weight calculation results of the present invention;

[0061] Figure 3 A flowchart of the steps for obtaining the training optimization model of the present invention;

[0062] Figure 4 A flow chart of the steps for obtaining the energy demand prediction results of the present invention;

[0063] Figure 5 A flowchart of the steps for obtaining the dynamic power configuration solution of the present invention;

[0064] Figure 6 The figure is a flow chart of the steps for obtaining the efficiency monitoring results of the present invention. DETAILED DESCRIPTION

[0065] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0066] In the description of the present invention, it should be understood that the terms "length", "width", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, in the description of the present invention, "multiple" means two or more, unless otherwise clearly and specifically defined.

[0067] Embodiment 1

[0068] See also Figure 1 The present invention provides a technical solution: a data center rack power management method, comprising the following steps:

[0069] S1: Collect the CPU usage, memory usage, and rack temperature of multiple racks in the data center, measure the power output of each device through the rack power unit, integrate to obtain basic energy consumption data, calculate the information entropy and critical weight of each data item, and obtain the weight calculation result;

[0070] S2: Using the weight calculation results, adjust the weights for the data input, set the structure of the neural network including the configuration of the activation function and the power distribution unit, start the network for iterative training until the parameters converge, and obtain the training optimization model;

[0071] S3: Input the real-time monitoring data into the training optimization model, use the power monitoring system to predict the short-term energy demand peak, use the sliding window method to smooth the prediction results, correct the model prediction deviation, and output the energy demand prediction results;

[0072] S4: Based on the energy demand forecast results, linear programming is applied to optimize the power output and cooling pipeline settings of the racks, calculate the optimal rack power configuration, adjust the power output and cooling pipeline settings, and generate a dynamic power configuration plan;

[0073] S5: Implement a dynamic power configuration solution and continuously track energy utilization efficiency through the power monitoring system, record changes in power usage and cooling efficiency, perform deviation analysis compared with expected targets, update the power configuration, and obtain efficiency monitoring results.

[0074] The weight calculation results specifically include information entropy and critical weights. The training optimization model specifically includes network structure, activation function configuration, and power distribution unit settings. The energy demand forecast results include short-term energy demand peak and forecast correction deviation. The dynamic power configuration plan specifically includes power output optimization and cooling pipeline configuration. The efficiency monitoring results specifically refer to energy utilization efficiency, power usage changes, and cooling efficiency.

[0075] See also Figure 2 , the specific steps for obtaining the weight calculation results are:

[0076] S111: Collect CPU usage, memory usage, and rack temperature of multiple racks in the data center, measure the power output of each device through the rack power unit, integrate the data, calculate the energy consumption of each data point, and generate basic energy consumption data;

[0077] Based on the collection of basic energy consumption data of multiple racks in the data center, the execution process includes starting from the CPU and memory usage monitoring of each device, while measuring the rack temperature and the power output of each device, and calculating the total energy consumption of the rack by integrating the data, including real-time data collection and preliminary data processing of each data point to accurately calculate the energy consumption, and preliminarily sorting and screening the collected data to exclude those data points with outliers and large deviations to ensure the accuracy and reliability of subsequent information entropy calculations. The generated basic energy consumption data is intended to provide input for the next step of information entropy analysis, and the results directly affect the calculation accuracy of information entropy and the quality of data evaluation.

[0078] S112: Based on the basic energy consumption data, information entropy is calculated for each data item, and information entropy data results are obtained by applying statistical analysis;

[0079] Starting from the generated basic energy consumption data, the information entropy calculation of the data items involves extracting the utilization rate and power output of each data point from the basic data, and applying statistical analysis methods to perform variation analysis and entropy value calculation on the extracted data. In this stage, each item of data is assigned an entropy value, which is obtained by probability distribution analysis and entropy formula calculation of the data items, ensuring the integrity of data processing and the scientific nature of the calculation results. The calculation of entropy values ​​provides a scientific basis for the next step of key weight analysis, and the generated information entropy data results are the preliminary data input for weight calculation.

[0080] S113: Use the information entropy data results to perform a criticality assessment on each data item using the formula:

[0081]

[0082] Process the variability and relative position of data items to obtain the critical weight of each data item, and then generate critical weight data results;

[0083] Among them, w i represents the critical weight of the i-th data item, H i represents the information entropy of the ith data item, R i Indicates the relative position of the data item in the entire data set and is used to adjust the weight distribution. iis the load factor of the data item, which indicates the usage load or criticality of the data item. α is the coefficient for adjusting the influence of information entropy, which is used to amplify or reduce the influence of information entropy. β is the position inverse adjustment coefficient, which increases the influence of position criticality by adjusting the coefficient. γ is the weight adjustment coefficient of the load factor, which is used to adjust the contribution of the load factor to the weight.

[0084] formula:

[0085]

[0086] The benefit of the formula is that by introducing the load factor L i and position index R i , so that the weight not only refers to the information entropy, but also reflects the actual load of the data item and its criticality in the total data set, thereby improving the accuracy and practicality of weight allocation.

[0087] Detailed explanation of the formula and the process of formula calculation and derivation:

[0088] Set the data to H i =0.5, R i =2,L i =0.75, and the adjustment coefficients are α=0.7, β=0.3, γ=0.5, the calculation process is as follows:

[0089]

[0090]

[0091]

[0092]

[0093] The results show that the calculated criticality weight is about 0.514. This value represents the comprehensive criticality of the data item in the weight analysis. It is the result of comprehensive information entropy, position criticality and load factor. A high weight value means that the data item contributes more to the overall data analysis and needs to be given priority.

[0094] S114: Integrate the key weight data results to form a weight calculation result.

[0095] The key weighted data results obtained by integration calculation are focused on during the integration process. The data items with high weights are used as key monitoring points for more rigorous data analysis and data processing, which includes not only the merging and summarizing of data, but also secondary verification and optimization of data to ensure that the weight calculation results obtained have high accuracy and reliability. The generated weight calculation results are directly used to evaluate and optimize the energy efficiency management strategy of the data center. The accurate acquisition of these results is of key significance for energy consumption management and equipment optimization.

[0096] See also Figure 3 , the specific steps for obtaining the training optimization model are:

[0097] S211: using the weight calculation result, adjusting the input weight of the neural network, setting the node weight of each layer in the network one by one, and then configuring the type and parameters of the activation function to generate an adjusted network configuration;

[0098] Using the weight calculation results, adjust the input weight of the neural network, set the node weights of each layer in the network one by one, and then configure the type and parameters of the activation function. By configuring the weight of each node layer by layer, determine the optimal activation function type for each layer of nodes, adjust the parameters to match the requirements of the target function, thereby optimizing the overall learning and response speed of the network, ensuring that the network can process input data and improve the accuracy of the output, and generate an adjusted network configuration.

[0099] S212: Based on the adjusted network configuration, a power distribution unit is set to optimize energy consumption, iterative training of the neural network is started, feedback learning is performed using a target data set, and training process data is generated;

[0100] Based on the adjusted network configuration, the power distribution unit is set to optimize energy consumption, the iterative training of the neural network is started, and the target data set is used for feedback learning. By adjusting the power distribution unit, the optimal balance between power supply and consumption in each training stage is ensured, reducing energy consumption while ensuring the continuity and efficiency of training. The training parameters are adjusted using real-time data feedback, and the energy consumption changes and efficiency improvements during training are monitored to generate training process data.

[0101] S213: Use the training process data to monitor the parameter convergence in real time, using the formula:

[0102]

[0103] Calculate parameter adjustments and generate real-time parameter adjustment data C n+1 ;

[0104] Among them, C nrepresents the parameter value of the current iteration, η is the learning rate, and the square root and absolute value ensure that the adjustment step size does not become too large when the parameter value is extremely small or the gradient change is not significant. is the loss function with respect to C n The gradient of ,∈ is a small constant to avoid the denominator being zero;

[0105] formula:

[0106]

[0107] The benefit of the formula is that by introducing square root and absolute value operations, the influence of the learning rate is adjusted, so that the parameter update step avoids excessive adjustments when facing small gradient values, increasing the stability and accuracy of the parameter update, which is especially critical in high-dimensional data processing and avoids the problem of gradient explosion or disappearance during training.

[0108] Detailed explanation of the formula and the process of formula calculation and derivation:

[0109] Set the parameter value C of the current iteration n is 0.5, the learning rate η is 0.01, and the gradient is 0.05, and the adjustment term ∈ is 1e-6, then

[0110] Q′C n =0.0025

[0111]

[0112]

[0113] C n+1 =0.5-0.19996·0.05=0.5-0.009998≈0.490002

[0114] The results show that through gradient adjustment and learning rate control, the parameter C n After one iteration, it was adjusted from 0.5 to 0.490002, showing a small adjustment, which is in line with the expected training optimization direction. This adjustment helps the neural network gradually converge to the optimal solution during the iteration process.

[0115] S214: Use the real-time parameter adjustment data and iterate until the change in all parameters is less than a preset threshold to generate a training optimization model.

[0116] Using real-time parameter adjustment data, continue to iterate until the change in all parameters is less than the preset threshold. By continuously monitoring the change in parameters after each iteration and comparing it with the preset threshold, ensure that the results of each iteration tend to be stable. This method helps to avoid overfitting while improving the generalization ability of the model. When the monitored parameter change is lower than the threshold for several consecutive iterations, it can be considered that the model has stabilized and a training optimization model is generated. This model will be able to perform the predetermined tasks with high precision, improving the accuracy and reliability of the prediction.

[0117] See also Figure 4 , the specific steps for obtaining energy demand forecast results are:

[0118] S311: input the real-time monitoring data into the training optimization model, combine it with the power supply monitoring system, predict the peak energy demand in the short term, and obtain preliminary energy demand forecast data;

[0119] Real-time monitoring data is input into the training optimization model. The model learns and analyzes based on previous and current energy usage data. Through deep learning technology, the model adaptively adjusts internal parameters to match the changing trend of energy demand. Through multivariate regression analysis of previous data, the model identifies the key factors and peak periods of energy use. Combined with the power supply monitoring system, the model predicts the peak energy demand in the short term and obtains preliminary energy demand forecast data. The data will be used for subsequent processing and analysis.

[0120] S312: Applying a sliding window method to smooth the preliminary energy demand forecast data to reduce the volatility and error of the forecast, thereby obtaining smoothed energy demand data;

[0121] The sliding window method is used to smooth the preliminary energy demand forecast data. This process involves selecting a matching window size to minimize the forecast error. The window is determined by past actual data to ensure that the volatility and outliers of the data are effectively controlled. Through sliding average or weighted smoothing technology, the processed data shows the true trend of energy demand. The smoothed energy demand data is more stable, reducing the impact of occasional fluctuations on forecast accuracy. The smoothed energy demand data provides a basis for the next step of deviation correction.

[0122] S313: Perform deviation correction on the smoothed energy demand data using the formula:

[0123]

[0124] Calculate and generate energy demand forecast results E corr ;

[0125] Among them, E smooth represents the smoothed energy demand data, Emin and E max They represent the minimum and maximum values ​​in the previous energy demand data respectively, and α is an adjustment coefficient used to adjust the forecast deviation according to previous data.

[0126] formula:

[0127]

[0128] The benefit of the formula is that it refers to the volatility of energy demand data and moderates the volatility by introducing an adjustment coefficient α, thereby obtaining a more stable and reliable forecast result.

[0129] Detailed explanation of the formula and the process of formula calculation and derivation:

[0130] First, obtain the smoothed energy demand data E smooth The minimum demand is 150MW. min The maximum demand is 100MW. max The power generation capacity is 200MW, and the adjustment factor α is set to 0.5, which is calculated as follows:

[0131]

[0132] Then calculate the adjustment:

[0133] 0.5×0.707=0.3535

[0134] E corr =150-0.3535=149.6465MW

[0135] The results show that the adjusted forecast value is closer to the actual demand, avoiding over-estimation or underestimation, and can guide energy management strategies more accurately.

[0136] See also Figure 5 , the specific steps for obtaining the dynamic power configuration scheme are:

[0137] S411: using the energy demand forecast result, setting the input parameters of the linear programming model, including the objective function and the constraint conditions, and optimizing the configuration of the rack power and the cooling system to obtain the initial parameters of the rack configuration;

[0138] The energy demand forecast results are used as input parameters. The parameters are obtained from the real-time monitoring system, including real-time power consumption and expected load peaks. Through data analysis, the initial parameters of the rack configuration are determined, including the current level of power output and the assessment of cooling demand. The linear programming model is used to process the data to optimize the energy efficiency of the entire data center. The model operation refers to the maximum load capacity and cooling efficiency of the equipment to ensure that the system operates effectively within the predicted demand range. The generated initial parameters of the rack configuration will serve as the basis for the next step of optimization.

[0139] S412: performing linear programming optimization on initial parameters of the rack configuration, adjusting power output and cooling pipeline settings to match the predicted energy demand peak, and calculating optimized rack configuration parameters;

[0140] Based on the initial parameters of the rack configuration, a linear programming algorithm is used for calculation. This process involves the adjustment of multiple variables to meet the goal of maximizing energy efficiency, including the gradual adjustment of power output and cooling systems. The optimal point is solved through algorithm iteration. In the process, the interaction and constraints of multiple components, such as power supply limitations and cooling capacity, are referred to. The optimized rack configuration parameters are adjusted through this method to ensure that each part operates in the optimal state, thereby obtaining the adjusted rack configuration parameters to provide input for the implementation of the dynamic power configuration solution.

[0141] S413: According to the optimized rack configuration parameters, the formula is used:

[0142]

[0143] Calculate the adjusted power output and generate a dynamic power configuration scheme P new ;

[0144] Among them, P opt represents the optimized power output, ΔP is the difference between the predicted demand and the current configuration, P max , P min are the maximum and minimum estimated demands, β and γ are adjustment coefficients, and ∈ is a small amount added to avoid division by zero.

[0145] formula:

[0146]

[0147] The benefit of the formula is that by referring to the nonlinear response of power adjustment and the dynamic changes of actual energy demand, the matching of the system to sudden load changes is improved, while maximizing the energy utilization efficiency.

[0148] Detailed explanation of the formula and the process of formula calculation and derivation:

[0149] The following data is obtained:opt =150kW, representing the optimized power output, ΔP = 20kW, which is the difference between the predicted demand and the current configuration, P max =200kW, P min =100kW, the maximum and minimum demands respectively, β=0.05, γ=0.1, the adjustment coefficient, obtained from the previous operation data analysis, ∈=1kW, added to avoid the zero division error in the calculation, substituted into the formula to calculate:

[0150]

[0151] The results show that the current power output is adjusted to 156.06kW, which means that the system has adjusted the output based on real-time data and prediction models to meet the expected energy demand while ensuring the optimization of operating efficiency. This dynamic adjustment strategy helps data centers manage energy consumption more effectively, ensuring sufficient power output during peak demand and avoiding problems such as insufficient or excessive power consumption.

[0152] See also Figure 6 , the specific steps for obtaining the efficiency monitoring results are:

[0153] S511: Implement a dynamic power configuration solution to continuously track real-time data on power usage and cooling efficiency through the power monitoring system to generate preliminary performance monitoring data;

[0154] In the process of implementing the dynamic power configuration solution, the real-time data collection function of the modern power monitoring system is utilized. The system first continuously collects data related to power usage and cooling efficiency through sensors installed on multiple key nodes, including power consumption peak, time period energy efficiency ratio and temperature changes. The monitoring data undergoes preliminary data cleaning and formatting to ensure the accuracy and real-time nature of the data. Through the data, the monitoring system can display the energy efficiency status of the equipment operation in real time, and conduct preliminary analysis of the data to identify energy waste points or performance bottlenecks, providing a scientific basis for subsequent data analysis and system optimization.

[0155] S512: Analyze the preliminary performance monitoring data, compare it with the preset efficiency target, determine the existing deviation, and obtain the deviation analysis result;

[0156] For the preliminary performance monitoring data obtained from the power monitoring system, engineers analyze the data, including comparing real-time data with previous performance data, analyzing trends and anomalies in the data, and using regression analysis and cluster analysis to identify patterns and potential problems in the data. Through these analyses, performance deviations and potential optimization points in the system can be accurately identified, providing data support for the formulation of more effective energy management strategies. Through this step, the efficiency of power use and the performance of the cooling system can be understood, so as to formulate targeted optimization measures.

[0157] S513: Update the power configuration according to the deviation analysis result, using the formula:

[0158]

[0159] Calculates adjusted power settings and generates efficiency monitoring results;

[0160] Among them, R new Represents the current efficiency result, R current Represents the efficiency data currently recorded, Target is the preset efficiency target, α and β are adjustment coefficients used to adjust the influence of target deviation and the amplitude of scaling deviation.

[0161] formula:

[0162]

[0163] The benefit of the formula is that the difference between the target and current efficiency can be flexibly adjusted by adjusting the coefficients β and α, thereby adjusting the power configuration, which helps to maintain the system operating at the optimal energy efficiency level.

[0164] Detailed explanation of the formula and the process of formula calculation and derivation:

[0165] Set the current efficiency R current is 80%, the target efficiency Target is set to 85%, the adjustment coefficient β is 0.05, and the stability coefficient α is 0.1. First, calculate the difference between the target and current efficiency and get:

[0166] Target-R current =85%-80%=5%

[0167] Apply the adjustment factor in the formula to calculate:

[0168] β×5%=0.05×5%=0.0025=0.25%

[0169] Calculate the denominator:

[0170]

[0171] Substituting into the formula, we get:

[0172]

[0173] The results show that the new efficiency after adjustment is slightly lower than the current efficiency. This is due to the overly conservative adjustment factor, indicating that it is necessary to re-refer to the values ​​of β and α when adjusting the power supply configuration to achieve an efficiency closer to the target.

[0174] A data center rack power management system, the data center rack power management system is used to execute the above data center rack power management method, the system includes:

[0175] The energy consumption data collection module collects CPU usage, memory usage, rack temperature and power output based on a multi-rack environment to obtain basic energy consumption data;

[0176] The weight calculation module calculates the information entropy of multiple data items based on the basic energy consumption data, extracts the key weights, and generates the weight calculation results;

[0177] The network training module sets the neural network structure based on the weight calculation results, including the activation function and power distribution unit configuration, performs iterative training until the parameters converge, and generates a training optimization model;

[0178] The demand forecasting module is based on the training optimization model, inputs real-time monitoring data, predicts the short-term energy demand peak, applies the sliding window method to smooth the prediction results, corrects the prediction deviation, and obtains the energy demand prediction results;

[0179] The optimization configuration module uses linear programming to optimize the rack power output and cooling pipeline settings based on the energy demand forecast results, calculates the optimal power configuration, and generates a dynamic power configuration plan;

[0180] The efficiency monitoring module is based on a dynamic power configuration solution. It continuously tracks energy utilization efficiency through the power monitoring system, records changes in power usage and cooling efficiency, performs deviation analysis compared with expected targets, updates the power configuration, and obtains efficiency monitoring results.

[0181] The above are only preferred embodiments of the present invention and are not intended to limit the present invention in other forms. Any technician familiar with the profession may use the technical contents disclosed above to change or modify them into equivalent embodiments with equivalent changes and apply them to other fields. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. A data center rack power management method, characterized in that: The following steps are involved: Collect CPU usage, memory usage, and rack temperature of multiple racks in the data center, measure the power output of each device through the rack power unit, integrate to obtain basic energy consumption data, calculate the information entropy and critical weight of each data item, and obtain the weight calculation results; Using the weight calculation results, adjusting the weights for data input, setting the structure of the neural network including the configuration of the activation function and the power distribution unit, starting the network for iterative training until the parameters converge, and obtaining a training optimization model; Input the real-time monitoring data into the training optimization model, use the power monitoring system to predict the peak energy demand in the short term, use the sliding window method to smooth the prediction results, correct the model prediction deviation, and output the energy demand prediction results; According to the energy demand prediction results, linear programming is applied to optimize the power output and cooling pipeline settings of the rack, calculate the optimal rack power configuration, adjust the power output and cooling pipeline settings, and generate a dynamic power configuration plan; Implement the dynamic power configuration solution, and continuously track energy utilization efficiency through the power monitoring system, record changes in power usage and cooling efficiency, perform deviation analysis compared with expected goals, update the power configuration, and obtain efficiency monitoring results.

2. The data center rack power management method according to claim 1, characterized in that: The weight calculation results specifically include information entropy and critical weights; the training optimization model specifically includes network structure, activation function configuration, and power distribution unit settings; the energy demand forecasting results include short-term energy demand peak and forecast correction deviation; the dynamic power supply configuration scheme specifically includes power output optimization and cooling pipeline configuration; the efficiency monitoring results specifically refer to energy utilization efficiency, power usage changes, and cooling efficiency.

3. The data center rack power management method according to claim 2, characterized in that: The steps for obtaining the weight calculation result are specifically as follows: Collect CPU usage, memory usage, and rack temperature of multiple racks in the data center, measure the power output of each device through the rack power unit, integrate the data, calculate the energy consumption of each data point, and generate basic energy consumption data; Based on the basic energy consumption data, information entropy is calculated for each data item, and information entropy data results are obtained by applying statistical analysis; Using the information entropy data results, each data item is evaluated for criticality using the formula: Process the variability and relative position of data items to obtain the critical weight of each data item, and then generate critical weight data results; Among them, w i represents the critical weight of the i-th data item, H i represents the information entropy of the ith data item, R i Indicates the relative position of the data item in the entire data set and is used to adjust the weight distribution. i is the load factor of the data item, which indicates the usage load or criticality of the data item. α is the coefficient for adjusting the influence of information entropy, which is used to amplify or reduce the influence of information entropy. β is the position inverse adjustment coefficient, which increases the influence of position criticality by adjusting the coefficient. γ is the weight adjustment coefficient of the load factor, which is used to adjust the contribution of the load factor to the weight. The key weight data results are integrated to form a weight calculation result.

4. The data center rack power management method according to claim 3, characterized in that: The steps for obtaining the training optimization model are specifically as follows: Using the weight calculation results, the input weights of the neural network are adjusted, the node weights of each layer in the network are set one by one, and then the type and parameters of the activation function are configured to generate an adjusted network configuration; Based on the adjusted network configuration, setting a power distribution unit to optimize energy consumption, starting iterative training of the neural network, using a target data set for feedback learning, and generating training process data; Using the training process data, the parameter convergence is monitored in real time, using the formula: Calculate parameter adjustments and generate real-time parameter adjustment data C n+1 ; Among them, C n represents the parameter value of the current iteration, η is the learning rate, and the square root and absolute value ensure that the adjustment step size will not become too large when the parameter value is extremely small or the gradient change is not significant, thus avoiding over-adjustment. is the loss function with respect to C n The gradient of ,∈ is a small constant to avoid the denominator being zero; The real-time parameter adjustment data is used to iterate until the change in all parameters is less than a preset threshold, thereby generating a training optimization model.

5. The data center rack power management method according to claim 4, characterized in that: The steps for obtaining the energy demand forecast result are specifically as follows: Inputting the real-time monitoring data into the training optimization model, combining it with the power supply monitoring system, predicting the peak energy demand in the short term, and obtaining preliminary energy demand forecast data; Applying a sliding window method to smooth the preliminary energy demand forecast data to reduce volatility and error of the forecast, thereby obtaining smoothed energy demand data; The energy demand data after the smoothing process is corrected for deviation using the formula: Calculate and generate energy demand forecast results E corr ; Among them, E smooth represents the smoothed energy demand data, E min and E max They represent the minimum and maximum values ​​in the previous energy demand data respectively, and α is an adjustment coefficient used to adjust the forecast deviation according to previous data.

6. The data center rack power management method according to claim 5, characterized in that: The steps for obtaining the dynamic power configuration scheme are specifically as follows: Using the energy demand forecast results, setting the input parameters of the linear programming model, including the objective function and constraints, and optimizing the configuration of the rack power and cooling system to obtain the initial parameters of the rack configuration; Performing linear programming optimization on the initial parameters of the rack configuration, adjusting the power output and cooling pipeline settings to match the predicted energy demand peak, and calculating the optimized rack configuration parameters; According to the optimized rack configuration parameters, the formula is adopted: Calculate the adjusted power output and generate a dynamic power configuration scheme P new ; Among them, P opt represents the optimized power output, ΔP is the difference between the predicted demand and the current configuration, P max , P min are the maximum and minimum estimated demands, β and γ are adjustment coefficients, and ∈ is a small amount added to avoid division by zero.

7. The data center rack power management method according to claim 6, characterized in that: The steps for obtaining the efficiency monitoring results are specifically as follows: Implement the dynamic power configuration scheme, continuously track real-time data of power usage and cooling efficiency through a power monitoring system, and generate preliminary performance monitoring data; Analyze the preliminary performance monitoring data, determine the existing deviations by comparing them with the preset efficiency targets, and obtain deviation analysis results; According to the deviation analysis results, the power supply configuration is updated using the formula: Calculates adjusted power settings and generates efficiency monitoring results; Among them, R new Represents the current efficiency result, R current Represents the efficiency data currently recorded, Target is the preset efficiency target, α and β are adjustment coefficients used to adjust the influence of target deviation and the amplitude of scaling deviation.

8. A data center rack power management system, characterized in that: According to any one of claim 17, the data center rack power management method comprises: The energy consumption data collection module collects CPU usage, memory usage, rack temperature and power output based on a multi-rack environment to obtain basic energy consumption data; The weight calculation module calculates the information entropy of multiple data items based on the energy consumption basic data, extracts the key weights, and generates a weight calculation result; The network training module sets the neural network structure based on the weight calculation result, including the activation function and the power distribution unit configuration, performs iterative training until the parameters converge, and generates a training optimization model; The demand forecasting module inputs the real-time monitoring data based on the training optimization model, predicts the short-term energy demand peak, applies the sliding window method to smooth the prediction results, corrects the prediction deviation, and obtains the energy demand prediction results; The optimization configuration module uses linear programming to optimize the rack power output and cooling pipeline settings based on the energy demand prediction results, calculates the optimal power configuration, and generates a dynamic power configuration plan; Based on the dynamic power configuration scheme, the efficiency monitoring module continuously tracks energy utilization efficiency through the power monitoring system, records changes in power usage and cooling efficiency, performs deviation analysis compared with expected targets, updates the power configuration, and obtains efficiency monitoring results.

Images