A method and system for data center power distribution optimization

Abstract

The application discloses a data center power distribution optimization method and system, which comprises the following steps: monitoring the power consumption signal and environmental sensing data of the data center computing node, constructing a three-dimensional environment-power consumption correlation field model, deeply mining the influence law of environmental factors such as temperature and humidity on power consumption demand, extracting the probability power consumption parameters from the correlation field by using the probability inference analysis technology, establishing the power consumption event triggering threshold, forming the dynamic power consumption management boundary, identifying the quality demand matrix based on the service quality demand difference, constructing the hierarchical power consumption parameter system, obtaining the adaptive parameters through load-power consumption correlation analysis, identifying the power consumption redundant components and converting them into the power consumption enhancement buffer area, realizing the adaptive power consumption adjustment, finally constructing the power consumption control matrix, detecting the power consumption imbalance nodes and balancing pairing, generating the accurate power consumption management instruction, realizing the intelligent optimization control of the data center power consumption, effectively improving the energy utilization efficiency and guaranteeing the service quality.

Description

Technical Field

[0001] This invention relates to the field of energy-saving management technology, and in particular to a method and system for optimizing power distribution in data centers. Background Technology

[0002] With the rapid development of cloud computing, big data, and artificial intelligence technologies, data centers have become the core infrastructure of modern information society. However, the enormous energy consumption of data centers is becoming increasingly prominent, with their power consumption accounting for a continuously rising proportion of global electricity consumption, posing a serious challenge to energy supply and environmental protection.

[0003] Existing data center power management methods are primarily based on static threshold control or simple load balancing strategies, lacking a deep understanding of the complex relationship between environmental factors and power consumption requirements. These methods often ignore the impact of environmental parameters such as temperature and humidity on the power consumption characteristics of computing nodes, making it difficult to achieve fine-grained adjustment of power allocation, resulting in low power utilization efficiency and significant resource waste. Furthermore, traditional power management schemes are slow to respond to dynamic load changes, unable to effectively identify and utilize redundant power resources, and lack differentiated consideration of service quality requirements, making it difficult to achieve optimal power configuration while ensuring service performance. Therefore, a data center power optimization and control technology is urgently needed to solve at least one of the above problems. Summary of the Invention

[0004] This invention provides a method and system for optimizing power distribution in data centers. By exploring the inherent correlation between environmental parameters and power consumption requirements, an environment-power consumption correlation field model is established. Probabilistic inference analysis technology is used to obtain adaptive power consumption parameters, identify power redundancy components and construct a power enhancement caching mechanism, generate hierarchical power consumption configurations based on the differentiated characteristics of service quality requirements, and finally achieve precise correction and optimization control of imbalance nodes through a power consumption control matrix, providing a high-efficiency and intelligent power management solution for data centers.

[0005] The first aspect of this invention proposes a data center power distribution optimization method, comprising the following steps:

[0006] Monitor the power consumption signals of computing nodes and environmental sensor data in the data center, extract power demand information from the computing node power consumption signals, and correlate and fuse the power demand information with the environmental sensor data to form an environment-power consumption correlation field;

[0007] The probabilistic power consumption parameters are obtained by performing probabilistic inference analysis using the environment-power consumption correlation field. Event evaluation is performed using the probabilistic power consumption parameters to determine the power event triggering threshold. Based on the power event triggering threshold and the probabilistic power consumption parameters, power configuration is performed to determine the dynamic power management boundary.

[0008] Based on the dynamic power management boundary, service quality requirements are identified to generate a quality requirement matrix. Based on the quality requirement matrix, hierarchical power consumption parameters are generated, and power level allocation configuration is constructed using the hierarchical power consumption parameters.

[0009] For the power level allocation configuration, load-power correlation is performed to obtain adaptive parameters. Power redundancy components are identified from the adaptive parameters to generate redundancy adjustment parameters. The power redundancy components are converted into power enhancement buffers using the redundancy adjustment parameters. The adaptive parameters are then adaptively adjusted using the power enhancement buffers to form adjusted power consumption data.

[0010] A power control matrix is ​​constructed based on the regulated power consumption data and the power consumption event trigger threshold. Power imbalance nodes within the regulated power consumption data are detected in the power control matrix. The power imbalance nodes are paired with the regulated power consumption data to obtain a power optimization control group. Power management instructions are generated based on the power optimization control group.

[0011] A second aspect of this invention provides a data center power distribution optimization system, comprising:

[0012] The correlation and fusion module is used to monitor the power consumption signals of computing nodes and environmental sensor data in the data center, extract power demand information from the power consumption signals of computing nodes, and correlate and fuse the power demand information with the environmental sensor data to form an environment-power consumption correlation field.

[0013] The probability analysis module is used to perform probability inference analysis using the environment-power consumption correlation field to obtain probability power consumption parameters, perform event evaluation using the probability power consumption parameters to determine power consumption event trigger thresholds, and perform power consumption configuration based on the power consumption event trigger thresholds and the probability power consumption parameters to determine dynamic power consumption management boundaries.

[0014] The quality identification module is used to identify service quality requirements based on the dynamic power management boundary, generate a quality requirement matrix, generate hierarchical power consumption parameters based on the quality requirement matrix, and construct a power level allocation configuration using the hierarchical power consumption parameters.

[0015] The redundancy conversion module is used to perform load-power correlation to obtain adaptive parameters for the power level allocation configuration, identify power redundancy components from the adaptive parameters to generate redundancy adjustment parameters, use the redundancy adjustment parameters to convert the power redundancy components into power enhancement buffers, and use the power enhancement buffers to adaptively adjust the adaptive parameters to form adjusted power consumption data.

[0016] The control optimization module is used to construct a power control matrix based on the adjusted power consumption data and the power consumption event triggering threshold, detect power imbalance nodes in the adjusted power consumption data in the power control matrix, perform power balance pairing between the power imbalance nodes and the adjusted power consumption data to obtain a power optimization control group, and generate power management instructions based on the power optimization control group.

[0017] The beneficial effects of this invention are reflected in the following points: 1. By constructing an environment-power consumption correlation field and applying probabilistic inference analysis technology, the inherent correlation between environmental parameters such as temperature and humidity and the power consumption requirements of computing nodes is deeply explored. Statistical analysis methods are used to quantify the strength of linear and nonlinear correlations, achieving comprehensive perception and accurate characterization of data center power consumption characteristics. Compared with traditional static threshold methods based on experience, this invention can dynamically adjust the power consumption event trigger threshold according to environmental changes, improving the accuracy of power consumption anomaly detection and the intelligence level of system response. 2. By generating a quality demand matrix through service quality demand identification technology, a differentiated hierarchical power consumption parameter system is constructed, dividing data center services into different levels such as critical business, important business, and general business. Corresponding power consumption guarantee coefficients and adjustment priorities are configured for each level of service, ensuring that critical business services receive priority in power consumption resource allocation. Even under system resource constraints, the high-quality operation of core services can still be maintained, achieving precise matching between power consumption allocation and business importance. 3. An innovative power redundancy identification and cache conversion mechanism is introduced. By analyzing the power distribution characteristics of the activation cycle and the sleep cycle, idle power resources are transformed into power-enhancing cache areas for dynamic allocation. Combined with power control matrix technology, imbalance nodes are accurately located. Eigenvalue decomposition technology is used to generate the dominant imbalance vector and implement balance correction, which effectively solves the problems of uneven power distribution and lag in adjustment response. The generated structured power management instructions support multiple modes such as immediate execution, timed execution, and conditional execution. Overall, intelligent optimization and efficient utilization of data center power resources are realized. Attached Figure Description

[0018] The accompanying drawings illustrate specific examples of the technical solutions described in this invention and, together with the detailed embodiments, form part of the specification, serving to explain the technical solutions, principles, and effects of this invention.

[0019] Unless otherwise specified or defined, the same reference numerals in different figures represent the same or similar technical features, and different reference numerals may be used to represent the same or similar technical features.

[0020] Figure 1 This is a flowchart illustrating a data center power distribution optimization method according to the present invention.

[0021] Figure 2This is a structural block diagram of a data center power distribution optimization system according to the present invention. Detailed Implementation

[0022] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0023] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0024] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0025] The technical solutions of the embodiments of this application will be described below.

[0026] like Figure 1 As shown, this embodiment of the invention provides a data center power distribution optimization method, including the following steps S110-S150:

[0027] Step S110: Monitor the power consumption signals of computing nodes and environmental sensor data in the data center, extract power demand information from the power consumption signals of computing nodes, and correlate and fuse the power demand information with the environmental sensor data to form an environment-power consumption correlation field.

[0028] Specifically, the system monitors the power consumption signals and environmental sensor data of the computing nodes in the data center. A power monitoring device is deployed on each computing node in the data center to collect real-time data on the node's voltage, current, and power consumption. The power monitoring device uses high-precision current transformers and voltage dividers, achieving a measurement accuracy of 0.1%, with a sampling frequency set to 1kHz to ensure the capture of instantaneous changes in power consumption signals. Monitoring parameters are configured for different types of computing nodes: CPU-intensive nodes focus on monitoring processor power consumption fluctuations, memory-intensive nodes focus on monitoring memory power consumption patterns, and network-intensive nodes focus on monitoring communication interface power consumption characteristics. Simultaneously, a multi-type environmental sensor network is deployed within the data center, including temperature sensors, humidity sensors, barometric pressure sensors, and airflow sensors. The temperature sensors employ a thermocouple design, covering a temperature range of 15-45 degrees Celsius with an accuracy of ±0.2 degrees Celsius, and are layered on the ceiling, floor, and within equipment racks. The humidity sensors use capacitive humidity-sensitive elements, measuring a range of 20%-80%RH with a response time of less than 30 seconds, and are paired with the temperature sensors to form temperature and humidity monitoring nodes. Barometric pressure and flow rate sensors are primarily deployed at the air inlets and outlets of the air conditioning system to monitor the operating status of the air circulation system. All sensor data is collected and sent to a data acquisition server via both wired Ethernet and wireless WiFi, forming a comprehensive monitoring network covering the entire data center.

[0029] Power demand information is extracted from the power consumption signals of computing nodes. The collected power consumption time-domain signals are converted into a time-frequency spectrum using a short-time Fourier transform, revealing the distribution patterns of power consumption in the time and frequency dimensions. Three typical modes in the power consumption signals are identified: basic load, peak load, and instantaneous load. Basic load corresponds to the minimum power consumption of the node in idle state, peak load corresponds to the maximum power consumption of the node under full load, and instantaneous load corresponds to the temporary power consumption increase of the node when handling sudden tasks. A sliding window statistical method is used to analyze the changing characteristics of each power consumption mode, with a window length set to 5 minutes, and the mean, variance, and trend of power consumption within the window are statistically analyzed. Power gradient analysis is used to identify the power consumption adjustment capability of nodes; nodes with larger gradient values ​​indicate rapid power consumption adjustment response, while nodes with smaller gradient values ​​indicate relatively smooth power consumption changes. Power demand is classified and labeled according to the node's workload type: CPU-intensive loads correspond to processor power consumption dominating, memory-intensive loads correspond to memory power consumption dominating, and network-intensive loads correspond to communication interface power consumption dominating. The extracted power demand information is categorized and organized according to power consumption mode and load type to form a power demand dataset.

[0030] Power consumption demand information and environmental sensor data are correlated and fused to form an environment-power consumption correlation field. The linear correlation strength between environmental parameters and power consumption demand is quantified using the Pearson correlation coefficient, while nonlinear correlations are analyzed using the Spearman rank correlation coefficient. The impact of temperature changes on computing node power consumption is analyzed in detail; CPU and memory power consumption typically increases with rising ambient temperature. Humidity and air pressure changes have relatively smaller impacts on power consumption but still need to be included in the correlation analysis. Excessive humidity affects equipment heat dissipation efficiency, indirectly affecting power consumption, while air pressure changes affect the efficiency of the cooling system. The environment-power consumption correlation field is constructed using Kriging spatial interpolation technology, expanding discrete monitoring point data into a continuous field distribution covering the entire data center space. In an application at a large cloud computing center, the correlation field model identified the northeast corner of the server room as a temperature-sensitive area, where the power consumption of computing nodes in this area responds to temperature changes by approximately 30% more than in other areas. The correlation field is represented in the form of a three-dimensional scalar field, where the X and Y coordinates correspond to the spatial location of the data center, the Z coordinate represents the intensity of power consumption demand, and the scalar value represents the influence weight of environmental factors.

[0031] Step S120: Use the environment-power consumption correlation field to perform probabilistic inference analysis to obtain probabilistic power consumption parameters, use the probabilistic power consumption parameters to perform event evaluation to determine the power consumption event trigger threshold, and perform power consumption configuration based on the power consumption event trigger threshold and probabilistic power consumption parameters to determine the dynamic power consumption management boundary.

[0032] In some embodiments, obtaining probabilistic power consumption parameters by performing probabilistic inference analysis using the environment-power consumption correlation field includes: identifying sparse regions of probability distribution using the environment-power consumption correlation field; generating sparse enhancement parameters based on the sparse regions of probability distribution; reconfiguring the sparse regions using the sparse enhancement parameters to form a probabilistic inference optimization region; and obtaining probabilistic power consumption parameters based on the probabilistic inference optimization region.

[0033] An environment-power consumption correlation field is used to identify sparse regions in the probability distribution. The data point density of each 10m × 10m grid cell in the environment-power consumption correlation field is scanned, and the number of valid data points per unit area is counted. The data density value of each grid is measured, and regions with a density below 60% of the average density are marked as candidates for sparse regions. Neighboring sparse grids are merged through neighborhood connectivity analysis to form continuous sparse regions. The geometric dimensions of each sparse region are measured, including length, width, and coverage area. The causes of sparse regions are identified, including sensor failure, equipment relocation, and data transmission interruption. Regions are classified according to their sparsity level: regions with a data density below 30% of the average are marked as severely sparse, and those between 30% and 60% are marked as moderately sparse. The spatial coordinates and boundary contours of sparse regions are recorded to provide location references for subsequent data compensation. The proportion of sparse regions to the total area is counted; when the proportion exceeds 15%, data augmentation processing needs to be initiated. The identified sparse regions in the probability distribution are ranked according to importance, with priority given to addressing sparsity issues in core equipment areas.

[0034] Sparse enhancement parameters are generated based on probability distribution sparse regions. For each identified sparse region, the power consumption distribution characteristics of the surrounding normal regions are extracted as a reference benchmark for the enhancement parameters. The distance from the boundary of the sparse region to the nearest normal region is measured; the closer the normal region, the greater its contribution weight to the enhancement parameters. An inverse distance weighting method is used to determine the spatial distribution of the enhancement parameters, with the enhancement intensity decreasing exponentially with increasing distance. The sparse enhancement factor λ = ρ_target / ρ_current is calculated, where ρ_target is the target data density set to 6 data points per square meter, and ρ_current is the current actual density. Enhancement strategies are selected based on the shape characteristics of the sparse regions: linear interpolation enhancement is used for strip-shaped sparse regions, and radial diffusion enhancement is used for block-shaped sparse regions. Considering the airflow direction in the data center, the enhancement weight along the airflow direction is increased by 20%, and the weight in the opposite direction is decreased by 10%. An enhancement parameter matrix is ​​generated, with the matrix size consistent with the number of grids in the sparse region, and each matrix element corresponds to an enhancement coefficient. Boundary continuity processing is applied to the enhancement parameters to ensure a smooth transition between the enhancement coefficients at the boundary of the sparse region and the adjacent normal regions.

[0035] A probabilistic inference optimization region is formed by reconfiguring sparse regions using sparse enhancement parameters. The generated sparse enhancement parameters are applied to the corresponding sparse grid cells, and data points are redistributed according to parameter weights. Within each sparse grid, a corresponding number of data points are inserted according to the enhancement coefficient values, with the insertion positions uniformly and randomly distributed. A bicubic interpolation algorithm is used to assign power consumption values ​​to the new data points, using the power consumption data of the eight neighboring grids as a reference. The interpolated power consumption data is checked for reasonableness to ensure that the values ​​are within the historical power consumption range of the region. A noise model is introduced to simulate the uncertainty of actual measurements, adding Gaussian noise with a standard deviation of 5% to the interpolated data. The statistical characteristics of the reconfigured region are checked to ensure that the mean and variance of the data distribution are consistent with the surrounding areas. The KS test is used to check the consistency between the probability distribution of the reconfigured data and the target distribution; a p-value greater than 0.05 indicates a good distribution match. The reconfigured region is named the probabilistic inference optimization region, and its spatial boundaries and data density are recorded. During equipment relocation or maintenance in the data center, this data reconfiguration method can effectively fill temporary monitoring gaps and maintain the continuity of power consumption analysis. The fusion boundary between the measurement optimization region and the original associated field is used to ensure the spatial continuity of the data.

[0036] Probabilistic power consumption parameters are obtained by probabilistically inferring the optimization region. Statistical characteristic parameters, including the mean, median, standard deviation, and interquartile range of power consumption, are extracted from the complete data distribution of the optimization region. The probability density function of power consumption within the optimization region is fitted using kernel density estimation to identify the peak location and shape characteristics of the distribution. The Kolmogorov-Smirnov test is used to determine the most suitable probability distribution type; common types include normal, gamma, and Weibull distributions. Characteristic parameters of the distribution are extracted: for the normal distribution, the mean μ and variance σ² are extracted; for the gamma distribution, the shape parameter α and scale parameter β are extracted. The temporal autocorrelation of power consumption data within the optimization region is measured, and the periodic patterns of power consumption changes are identified through the autocorrelation function. The correlation coefficients at different time lags are calculated; the correlation coefficient with a lag of one time step is typically between 0.6 and 0.8. The variation of power consumption parameters under different environmental conditions is analyzed; for every 1 degree Celsius increase in temperature, the expected power consumption increases by 2-3 watts. The probabilistic power consumption parameters are categorized according to the type of computing node. The power consumption distribution skewness for CPU-intensive nodes is 0.4-0.7, while that for memory-intensive nodes is 0.2-0.5. The obtained probabilistic power consumption parameters are organized into parameter vectors, with each optimization region corresponding to a vector containing 6-8 parameters.

[0037] Event assessment using probabilistic power consumption parameters determines the trigger threshold for power consumption events. The expected value and standard deviation information from these parameters are used to analyze the risk level of abnormal power consumption events in each region. The 3σ principle of statistical process control is used to set the anomaly detection boundary; an event is considered abnormal when power consumption exceeds the expected value by ±3 times the standard deviation. Differentiated thresholds are set for grid regions with different power consumption modes: μ+2.5σ for high-power mode regions, μ+3σ for medium-power mode regions, and μ+3.5σ for low-power mode regions. Considering the impact of ambient temperature on power consumption, the trigger threshold is increased by 10% when the ambient temperature exceeds 28 degrees Celsius. Historical power consumption event data is analyzed to statistically determine the duration and impact range of various abnormal events. Minor anomalies last an average of 3-5 minutes, while severe anomalies can last 15-30 minutes. Threshold settings are optimized using ROC curves to maximize detection accuracy within a 5% false alarm rate constraint. In actual operation, it was found that the server cluster's power consumption reaches the expected peak during peak business periods. Adjusting the trigger threshold settings during these periods prevents normal high loads from being misjudged as abnormal and triggering unnecessary alarms. The determined power consumption event trigger thresholds are categorized and stored according to grid coordinates and node type. Each grid corresponds to a set of parameters including upper limit threshold, lower limit threshold, and adjustment coefficient.

[0038] In some embodiments, the step of determining the dynamic power management boundary by performing power configuration based on the power event triggering threshold and the probabilistic power consumption parameter includes: performing configuration matching analysis on the power event triggering threshold and the probabilistic power consumption parameter to generate a matching degree distribution; identifying configuration deviation regions from the matching degree distribution to generate deviation identification parameters; converting the configuration deviation regions into boundary calibration optimization regions through the deviation identification parameters; and determining the dynamic power management boundary based on the boundary calibration optimization regions.

[0039] A matching degree distribution is generated by performing configuration matching analysis on power event trigger thresholds and probabilistic power consumption parameters. The power event trigger threshold for each grid is compared with its corresponding probabilistic power consumption parameter to analyze the degree of matching. The deviation between the trigger threshold and the expected power consumption value is measured, and the matching degree index M = 1 - |T - μ| / μ is calculated, where T is the trigger threshold and μ is the expected power consumption value. Matching degree analysis is performed on each grid cell in the data center; a matching degree close to 1 indicates a reasonable configuration, while a matching degree close to 0 indicates an inappropriate configuration. The spatial distribution pattern of the matching degree is identified, and a matching degree contour map is drawn to show the matching level in different areas. The number of grids in each matching degree interval is counted: grids with a matching degree greater than 0.8 account for 65% of the total, those between 0.6 and 0.8 account for 25%, and those below 0.6 account for 10%. The correlation between the matching degree and environmental factors is analyzed; the matching degree is generally lower in areas with higher temperatures, with an average matching degree of 0.72. The clustering characteristics of the matching degree distribution are identified; grids with similar matching degrees tend to cluster spatially. The spatial autocorrelation coefficient of the matching degree distribution is measured; a correlation coefficient of 0.43 indicates moderate spatial clustering. The matching degree distribution was divided into 5 levels at 0.2 intervals to provide a grading reference for subsequent identification of deviation areas. The statistical characteristics of the matching degree distribution were recorded, including the global average matching degree of 0.76, the standard deviation of 0.18, and the coefficient of variation of 0.24.

[0040] The process involves identifying configuration deviation regions from the matching degree distribution and generating deviation identification parameters. Grids with a matching degree below 0.5 are selected as candidate units for configuration deviation regions. Adjacent low-matching degree grids are merged using eight-neighbor connectivity analysis to form continuous deviation regions. The geometric characteristics of each configuration deviation region are measured, including region area, perimeter, and shape complexity index. The distribution characteristics of matching degree within the deviation regions are analyzed, recording the minimum matching degree, average matching degree, and standard deviation of the matching degree. Configuration deviation regions are classified according to the severity of the deviation: an average matching degree of 0.3-0.5 indicates mild deviation, 0.1-0.3 indicates moderate deviation, and below 0.1 indicates severe deviation. The spatial distribution density of deviation regions is identified, and the nearest distance and distribution uniformity between regions are measured. The correspondence between deviation regions and device layout is analyzed, revealing that deviation regions are concentrated in areas with high-power device density and poor heat dissipation. The generated deviation identification parameters include five dimensions: region number, center coordinates, area size, deviation level, and shape parameters. Weights are assigned to the deviation identification parameters, and the processing priority is determined based on the deviation level and the area. Areas with severe deviations and large areas have the highest weight.

[0041] For example, the step of converting the configuration deviation region into a boundary calibration optimization region through the deviation identification parameter includes: identifying deviation intensity characteristics based on the deviation identification parameter and the configuration deviation region to determine a deviation analysis window, wherein the deviation intensity characteristics include deviation frequency, duration intensity, and influence range;

[0042] The deviation change process is traced along the deviation analysis window to form a deviation change trajectory diagram; the position coordinates of the calibration adjustment points are extracted from the deviation change trajectory diagram; the position coordinates are arranged according to the calibration effect to determine the boundary calibration optimization area.

[0043] The deviation analysis window is determined based on deviation identification parameters and the configuration of the deviation region to identify deviation intensity characteristics. Deviation frequency information is extracted from the deviation identification parameters, and the number of times the matching degree of the region falls below a threshold per unit time is counted. The duration and intensity of power consumption fluctuations within the deviation region are measured, and the severity of the fluctuations is quantified using the standard deviation and peak-to-peak value of power consumption changes. The influence range of the deviation region is determined, spreading outward from the center of the deviation region, and the distance at which the influence decays to 50% is measured as the influence radius. Deviation frequency, duration and intensity, and influence range are used as the three analytical dimensions of deviation intensity characteristics. The size of the analysis window is determined based on the quantification results of the deviation intensity characteristics; higher intensity deviations require a larger window range for analysis. The geometric parameters of the analysis window are set: a rectangular window with an aspect ratio of 3:1 is used for strip-shaped deviations, and a square window is used for concentrated deviations. The center of the analysis window is positioned at the centroid of the deviation region, ensuring that the window completely covers the deviation region. The orientation angle of the window is adjusted so that the principal axis of the window is consistent with the main extension direction of the deviation region.

[0044] A deviation trajectory graph is generated by tracing the deviation change process along the deviation analysis window. Power matching data is collected at 5-minute intervals within the deviation analysis window, recording the average deviation level at each time point. The rate of change of the deviation level is measured to identify three trends: deviation deterioration, deviation improvement, and deviation stabilization. The spatial gradient of the deviation level is analyzed to determine the propagation direction and diffusion speed of the deviation within the window. A deviation trajectory graph is plotted, with the horizontal axis representing time (hours) and the vertical axis representing the deviation level (0-1 normalized), showing the temporal evolution of the deviation. Key event moments are marked on the trajectory graph, including the time points when the deviation first appears, reaches its peak, and begins to alleviate. The morphological pattern of the trajectory graph is analyzed to identify different types such as periodic fluctuations, monotonically increasing fluctuations, and random fluctuations. The geometric characteristics of the trajectory are measured, including the total trajectory length, average slope, and maximum deviation amplitude. The trajectory graph is smoothed using a 3-point moving average to filter out high-frequency noise and retain the main trends.

[0045] The location coordinates of calibration adjustment points are extracted from the deviation change trajectory graph. Inflection points where the deviation degree changes significantly are identified on the deviation change trajectory as candidate calibration adjustment points. Extreme points of the rate of change are identified using the first derivative of the trajectory curve; points with an absolute derivative value greater than 0.05 / hour are marked as candidate adjustment points. The second derivative of the trajectory curve is used to detect curvature changes and identify inflection points in the deviation change trend. Candidate adjustment points are screened, retaining inflection points where the deviation change amplitude exceeds 0.1 and the duration exceeds 10 minutes. The time coordinates of the adjustment points on the trajectory graph are converted to spatial coordinates, and combined with the spatial information from the deviation analysis window, the actual location of the adjustment points in the data center is determined. The distribution density of calibration adjustment points is measured to identify spatial patterns of dense and sparse areas. Adjustment points are ranked according to their importance based on the amplitude and direction of deviation change, with points showing greater deviation improvement being more important. In operational practice, it has been found that certain locations can significantly improve the matching degree of the surrounding area after power consumption adjustments; these key adjustment points are usually located at the intersection of air conditioning airflows or near main heat dissipation channels. Cluster analysis was performed on adjacent adjustment points, grouping adjustment points less than 3 meters apart into adjustment point groups, and selecting a representative point for each group. The location coordinates of the calibration adjustment points were compiled into a list format, including four fields: X coordinate, Y coordinate, timestamp, and importance weight.

[0046] The calibration effect of position coordinates is ranked to determine the boundary calibration optimization zone. The calibration potential of each calibration adjustment point is analyzed, and the calibration effect of each point is analyzed through historical deviation improvement data. The calibration effect index E=(D_after-D_before) / D_before is calculated, where D_before is the deviation degree before calibration and D_after is the deviation degree after calibration. All adjustment points are ranked in descending order according to the calibration effect index, with the most significant effect at the top. The spatial distribution pattern of high-effect adjustment points is analyzed, and it is found that these points are mostly located near airflow channels and heat dissipation equipment. The adjustment points with the top 60% calibration effect are selected as the key areas for boundary calibration. The spatial envelope of the key adjustment points is calculated using the convex hull algorithm to determine the smallest convex polygon containing all key points. The convex hull boundary is extended outward by 3 meters to form a buffer zone, generating the complete calibration optimization zone. The geometric parameters of the boundary calibration optimization zone are measured, including the total area, boundary perimeter, and shape regularity. The optimization zone is further subdivided according to the calibration effect, with the 20% of the area with the best effect designated as the core optimization zone, and the rest as the general optimization zone.

[0047] The dynamic power management boundary is determined based on the boundary calibration optimization zone. This zone is designated as the key control area for dynamic power management, where fine-grained power consumption adjustment is implemented. The power distribution characteristics of each grid within the calibration optimization zone are analyzed to identify the spatial layout of high-power, medium-power, and low-power grids. Power control strategies are adjusted according to the spatial location of the calibration optimization zone, with stricter power limits and faster adjustment responses applied to the core optimization zone. Control parameters for the dynamic power management boundary are set, including a power limit 15% lower than the conventional area and a reduction in adjustment response time to 2 minutes. A power allocation strategy within the boundary is formulated, prioritizing the power consumption needs of core service nodes, while allowing dynamic adjustment of power consumption for non-core nodes. An adaptive adjustment mechanism for the dynamic boundary is constructed: when the average matching degree within the boundary is above 0.8 for 30 consecutive minutes, the boundary range shrinks by 10%; when it is below 0.6, the boundary expands by 15%. Triggering conditions and constraints for boundary adjustments are set, with a single adjustment not exceeding 20% ​​of the boundary area to avoid over-adjustment. The determined dynamic power management boundary parameters are compiled into a configuration file containing complete information such as boundary coordinates, control strategies, adjustment rules, and monitoring indicators.

[0048] Step S130: Based on the dynamic power management boundary, identify the quality of service requirements to generate a quality requirement matrix, generate hierarchical power consumption parameters based on the quality requirement matrix, and use the hierarchical power consumption parameters to construct a power level allocation configuration.

[0049] In some embodiments, the step of identifying quality of service requirements and generating a quality of service requirement matrix based on the dynamic power management boundary includes: identifying quality of service requirements based on the dynamic power management boundary to determine a requirement mismatch region; extracting mismatch compensation parameters from the requirement mismatch region; using the mismatch compensation parameters to correct the requirement mismatch region and generate requirement correction adjustment points; and generating a quality of service requirement matrix according to the requirement correction adjustment points.

[0050] Service quality requirements are identified and mismatched regions are determined based on the dynamic power management boundary. The power consumption cap parameters of each grid within the dynamic power management boundary are read and compared with the actual power consumption requirements of services running within the grid. The actual power consumption requirements of services are measured, including power consumption corresponding to CPU computing load, memory access frequency, and network communication bandwidth. Grid regions where the power consumption cap cannot meet service requirements are identified; regions where the actual power consumption requirement exceeds 90% of the grid's power consumption cap are marked as potential mismatched regions. The matching relationship between service quality indicators and power supply is analyzed, focusing on three key indicators: response time, throughput, and availability. A region is identified as a demand mismatched region when the average response time of services within the grid exceeds 150% of the target value or the availability is below a preset threshold. Adjacent mismatched grids are merged through connected component analysis to form continuous demand mismatched regions. The geometric characteristics and impact range of each demand mismatched region are measured, including region area, boundary shape, and the number of service types included. The severity of the mismatched regions is analyzed and graded according to the importance of the service type and the magnitude of the mismatch. Even slight mismatches in critical business service regions are marked as high severity, while moderate mismatches in general business service regions are marked as medium severity. The number and total area of ​​mismatched areas at each severity level are counted to determine the priority order for subsequent compensation processing.

[0051] Extract mismatch compensation parameters from demand mismatch areas. For each identified demand mismatch area, analyze the specific causes and gap size. Measure the power consumption demand gap and calculate the difference between actual service demand and current supply; a larger difference indicates a more severe mismatch. Analyze the temporal characteristics of the mismatch, identifying different patterns of persistent and intermittent mismatches. Persistent mismatches require a permanent increase in power consumption, while intermittent mismatches can be resolved through dynamic adjustments. Extract the power consumption distribution pattern of normal areas surrounding the mismatch area as a reference benchmark for compensation parameters. Calculate the compensation intensity coefficient, the magnitude of which is proportional to the degree of mismatch and service importance. Set a compensation intensity coefficient of 1.5-2.0 for mismatch areas of critical business services, 1.2-1.5 for important business services, and 1.0-1.2 for general business services. Consider the impact of compensation operations on adjacent areas and set compensation range limits to avoid affecting the normal operation of other areas during compensation. Generate a compensation parameter vector containing key information such as compensation intensity, compensation range, compensation duration, and compensation source. In actual data center operations, it has been found that video processing services experience a surge in power consumption when encoding high-definition content, requiring rapid allocation of power resources from surrounding lightly loaded areas to compensate for this.

[0052] The process utilizes mismatch compensation parameters to correct demand mismatch areas, generating demand correction adjustment points. The extracted mismatch compensation parameters are applied to the corresponding demand mismatch areas to perform power allocation correction. The power allocation weight of the mismatch area is adjusted according to the compensation intensity coefficient, with the adjustment magnitude proportional to the compensation intensity. A progressive correction method is adopted, gradually increasing the power supply of the mismatch area in multiple steps, with each step's increase controlled within 20% to avoid over-adjustment and system oscillation. During the correction process, changes in service quality indicators (SHI) within the area are monitored in real time, including the degree of improvement in response time and throughput. Correction stops when the SHI reaches the preset target value, and the power allocation status at this point is recorded as the correction completion status. Key adjustment points are identified during the correction process, corresponding to the moments and locations where the power adjustment effect is most significant. By analyzing the changes in power distribution before and after correction, the spatial coordinates and adjustment magnitude of the correction adjustment points are determined. The influence radius of each correction adjustment point is measured, and the impact of power change at the adjustment point on the service quality of the surrounding area is calculated. The correction adjustment points are ranked according to their importance based on their adjustment effect, with adjustment points having significant effects and a wide impact range being of higher importance. Record detailed information for each required correction adjustment point, including the adjustment point coordinates, the power consumption difference before and after adjustment, the range of influence, and the adjustment time.

[0053] A quality demand matrix is ​​generated based on demand correction adjustment points. Using the spatial coordinates and adjustment parameters of these points, a matrix structure reflecting the distribution of service quality demands is constructed. The data center space is discretized according to a grid coordinate system, with the grid size consistent with the grid of the dynamic power management boundary. Each grid is assigned a corresponding row and column position in the matrix; the X-coordinate of the grid corresponds to the matrix row number, and the Y-coordinate corresponds to the matrix column number. The values ​​of matrix elements are determined based on the adjustment magnitude of the demand correction adjustment points, with larger adjustment magnitudes corresponding to higher matrix element values. For grids containing multiple adjustment points, a weighted average method is used to calculate the comprehensive matrix element values, with weights determined according to the importance of the adjustment points. Blank areas in the matrix are addressed; grids without adjustment points are filled using interpolation of values ​​from surrounding grids. A bilinear interpolation algorithm is used to calculate the service quality demand values ​​for blank grids, ensuring the spatial continuity of the matrix. The generated quality demand matrix is ​​normalized, mapping all element values ​​to the standard 0-1 range for easier subsequent parameter comparison and calculation. The quality demand matrix is ​​represented hierarchically according to service type, with critical business services, important business services, and general business services corresponding to different levels of the matrix.

[0054] Tiered power consumption parameters are generated based on the quality demand matrix. Service quality demand values ​​for each grid are extracted from the matrix, and the demand levels are converted into corresponding power consumption tier parameters. For critical business service areas, a first-level power consumption parameter is set with the highest priority and a power guarantee coefficient of 1.2, indicating that 120% of the baseline power supply must be guaranteed even under power constraints. For important business service areas, a second-level power consumption parameter is set with a medium priority and a power guarantee coefficient of 1.0, indicating that 100% of the baseline power supply is guaranteed. For general business service areas, a third-level power consumption parameter is set with a low priority and a power guarantee coefficient of 0.8, indicating that operation can be reduced to 80% of the baseline power supply when power is insufficient. Considering the different sensitivities of different service types to power fluctuations, the adjustment range and response time for each level of power consumption parameter are calculated. The adjustment range for the first-level power consumption parameter is controlled within ±5%, with a response time of within 1 minute. The second-level power consumption parameter allows an adjustment range of ±10% with a response time of 3 minutes. The third-level power consumption parameter allows an adjustment range of ±20% with a response time of 5 minutes. Based on the power coordination relationship between grids, a conversion coefficient for cross-level power allocation is set. When the demand in the high-level power region increases, power resources can be called from the low-level power region.

[0055] A power allocation configuration is constructed using tiered power parameters. Priority weights and guarantee coefficients from these parameters are used to construct a power allocation scheme for the data center. All available power is pre-allocated to Tier 1, Tier 2, and Tier 3 in a 5:3:2 ratio to ensure sufficient power supply for critical services. Within each tier, secondary allocation is performed based on the specific needs of the grid, with grids with higher demand receiving a larger share of power. A power pool mechanism is introduced to aggregate unused power from each tier into the corresponding pool to handle sudden increases in power demand. Power borrowing rules are set between tiers; when a tier's power pool is insufficient, power resources can be borrowed from other tiers according to borrowing priority. Tier 1 power areas can borrow power from Tier 2 and Tier 3, Tier 2 power areas can only borrow from Tier 3, and Tier 3 power areas cannot borrow power from other tiers. A real-time power allocation adjustment strategy is developed to dynamically adjust the power allocation ratio based on real-time changes in service load within the grid. When the service load of a grid exceeds a preset threshold, the power allocation for that grid is automatically increased; when the load decreases, the power allocation is reduced accordingly. Set constraints for power allocation to ensure that any adjustments will not cause the power consumption of critical business services to fall below the minimum guaranteed threshold. Convert the constructed power level allocation configuration into executable control instructions, including parameters such as target power consumption, adjustment range, and execution timing for each grid.

[0056] Step S140: For the power level allocation configuration, execute load-power correlation to obtain adaptive parameters, identify power redundant components from the adaptive parameters to generate redundant adjustment parameters, use the redundant adjustment parameters to convert the power redundant components into a power enhancement buffer, and use the power enhancement buffer to adaptively adjust the adaptive parameters to form adjusted power consumption data.

[0057] Specifically, load-power correlation is performed to obtain adaptive parameters for power level allocation configuration. The power level allocation configuration is used as input to read the target power consumption, adjustment range, and execution timing parameters for each grid. Real-time load data of servers within each grid is collected, including key indicators such as CPU utilization, memory usage, disk I / O frequency, and network traffic. The load data is correlated with the corresponding grid's power allocation parameters to establish the response relationship between load changes and power demand. The load-power correlation coefficient is determined through linear regression analysis, and the correlation strength is calculated using the formula R=ΔP / ΔL, where R is the correlation coefficient, ΔP is the power change, and ΔL is the load change. Correlation coefficients are determined for different power allocation levels. The correlation coefficient for the first-level power allocation level is typically between 1.2 and 1.5, indicating that a 1% increase in load corresponds to a 1.2%-1.5% increase in power consumption; for the second-level power allocation level, it is between 0.8 and 1.2; and for the third-level power allocation level, it is between 0.6 and 0.9. The time delay characteristics of the load-power correlation are analyzed, measuring the time interval between load changes and the effective power adjustment. The extracted adaptive parameters include four dimensions: correlation strength, response delay, adjustment range, and settling time.

[0058] In some embodiments, the step of identifying power-redundant components from the adaptive parameters and generating redundancy adjustment parameters includes: constructing a redundant time axis using the adaptive parameters; identifying redundant peak points from the adaptive parameters and mapping them to the redundant time axis to form peak time markers; dividing the redundant time axis into active periods and sleep periods according to the peak time markers; and comparing the power distribution characteristics corresponding to the active periods and the sleep periods to generate redundancy adjustment parameters.

[0059] A redundancy timeline is constructed using adaptive parameters. Time-series data from the adaptive parameters is extracted, including changes in correlation strength, response latency, and adjustment range at each time point. The parameter data is arranged chronologically on a time axis to form a continuous parameter evolution curve. The time axis is set to a 5-minute interval to ensure that subtle changes in power redundancy are captured. The comprehensive redundancy index is determined for each time point; the redundancy index equals 1 minus power utilization efficiency, with a higher value indicating a higher degree of redundancy. The redundancy index curve is plotted on the time axis to create an intuitive redundancy timeline chart. Key time points on the time axis are marked, including the start, end, and transition times of redundancy. The periodic characteristics of the redundancy timeline are analyzed to identify regular patterns in redundancy changes. Redundancy changes in most data centers exhibit a 24-hour periodicity, with significantly higher redundancy levels at night than during the day. The statistical characteristics of the redundancy timeline are measured, including average redundancy level, redundancy fluctuation amplitude, and redundancy duration.

[0060] Redundant peak points are identified from adaptive parameters and mapped onto a redundancy time axis to form peak moment markers. Local maxima of the redundancy index are scanned on the redundancy time axis; when the redundancy index at a given moment is higher than both the preceding and following moments, it is marked as a candidate peak point. A threshold condition for peak identification is set; only candidate points with a redundancy index exceeding 1.5 times the average are confirmed as true peak points. The causes of peak point formation are analyzed, including sudden drops in service load, failure to adjust power allocation in a timely manner, and temporary increases in equipment operating efficiency. The peak intensity of each peak point is measured; the intensity equals the difference between the redundancy index at the peak and the baseline redundancy index. Peak points are classified according to their intensity: those greater than 0.3 are strong peak points, those between 0.1 and 0.3 are medium peak points, and those less than 0.1 are weak peak points. The identified peak points are mapped onto the redundancy time axis according to the time coordinate, and peak symbols are marked at the corresponding positions. The distribution density of peak moment markers is analyzed to identify time periods with concentrated peaks and sparse time periods. During nighttime operation in the data center, due to decreased user access, power redundancy peak points frequently occur, providing important opportunities for power optimization. Peak time markers are sorted according to peak intensity, with stronger peaks having higher adjustment priority. A complete peak time marker database is formed, containing information such as peak time, peak intensity, and peak type.

[0061] For example, dividing the redundant time axis into an active period and a dormant period based on the peak time marker includes: identifying a period transition failure point based on the peak time marker; extracting a transition delay parameter from the period transition failure point; performing time-series reconstruction on the period transition failure point using the transition delay parameter to obtain a period coordination optimization region; and dividing the redundant time axis into an active period and a dormant period based on the period coordination optimization region.

[0062] Identifying periodic transition failure points based on peak time markers. Anomalies on the redundancy time axis are detected near peak time markers; points where the redundancy index fluctuates drastically within a short period are marked as candidate transition anomalies. The characteristic patterns of transition anomalies are analyzed, including three failure types: too-fast transition, too-slow transition, and transition interruption. Too-fast transition refers to a redundancy index change exceeding 50% within 10 minutes; too-slow transition refers to a transition process lasting more than 2 hours; and transition interruption refers to a reverse change during the transition. The deviation of transition failure points is measured, quantifying the severity of deviation by the area difference between the actual and ideal transition curves. Causes of transition failures are identified, including inaccurate load forecasting, delayed power consumption adjustment response, and insufficient system adjustment capabilities. The frequency and distribution of various transition failure points are statistically analyzed; transition failures typically occur at the beginning and end of peak business periods. Identified periodic transition failure points are categorized according to failure type and severity, with severe failures requiring priority handling. The impact of transition failures on system performance is analyzed; service quality near failure points typically fluctuates. Detailed files for each transition failure point are created, including failure time, failure type, scope of impact, and processing priority.

[0063] Extract conversion delay parameters from cycle conversion failure points. For each identified cycle conversion failure point, analyze the time delay characteristics of the conversion process. Measure the difference between the ideal conversion time and the actual conversion time; a positive difference indicates conversion lag, and a negative difference indicates conversion lead. Calculate the delay time using the formula T=|t_actual-t_ideal|, where T is the conversion delay time, t_actual is the actual conversion time, and t_ideal is the ideal conversion time. Analyze the relationship between delay parameters and load variation; larger load variations generally correspond to longer conversion delays. Identify influencing factors of delay parameters, including system response capability, regulation mechanism sensitivity, and external environmental changes. Set different weights for delay parameters according to the type of conversion failure; the delay parameter weight for slow conversion is set to the highest, with a weighting coefficient of 1.5. Extract the spatial distribution characteristics of delay parameters; delay parameters may differ significantly in different grid regions.

[0064] A periodic coordinated optimization region is obtained by temporally reconstructing the periodic conversion failure points using conversion delay parameters. The delay duration and delay characteristics in the conversion delay parameters are used to temporally correct the conversion failure points. The time coordinates of the conversion failure points are adjusted according to the delay parameters, moving lagging conversion points forward and leading conversion points backward. The ideal curve of the conversion process is redrawn using a temporal reconstruction algorithm; the corrected curve should smoothly connect the activation and dormancy cycles. The degree of difference between the conversion curves before and after reconstruction is measured; a smaller difference indicates a better reconstruction effect. Regions requiring special optimization during the reconstruction process are identified; these regions correspond to more severe conversion failures and larger delay parameters. The regions requiring optimization are expanded into a coordinated optimization region with a certain spatial and temporal range; the size of the optimization region is proportional to the severity of the failure. Boundary parameters of the coordinated optimization region are set, including key elements such as time span, spatial range, and optimization intensity.

[0065] The redundant time axis is divided into active and dormant periods based on the periodic coordination optimization region. The boundary information of the periodic coordination optimization region is used to redetermine the division points of the active and dormant periods. The time-series reconstruction results within the coordination optimization region are used as the new division basis, replacing the original peak time markers. The true periodic transition time is identified based on the reconstructed transition curve; the transition time corresponds to the position with the largest change rate of the redundancy index. In actual data center operation, this transition time usually corresponds to a critical switching point of business load, such as the transition from peak daytime web services and database queries to nighttime data backup and log archiving processing. The time range of the active and dormant periods is redefined using the new periodic transition time as the dividing point. The duration and redundancy characteristics of each period after re-division are measured to ensure that the division results conform to the actual changes in power consumption. The division boundaries are fine-tuned to eliminate abrupt changes and discontinuities in the redundancy index near the boundaries. This fine-tuning is common in data center operation; for example, when a storage system starts a large-scale data synchronization task late at night, the boundaries of the dormant period need to be adjusted to adapt to the sudden changes in the power consumption requirements of the storage cluster.

[0066] Redundancy adjustment parameters are generated by comparing the power distribution characteristics of the active and sleep cycles. Power distribution data for both cycles are statistically analyzed, including average power consumption, power fluctuation range, and power utilization efficiency. The power consumption characteristics of the active cycle are determined: average power consumption is typically 80%-95% of the allocated power consumption, power fluctuation range is 10%-15%, and utilization efficiency reaches 85%-95%. The power consumption characteristics of the sleep cycle are also determined: average power consumption is typically 50%-70% of the allocated power consumption, power fluctuation range is 5%-10%, and utilization efficiency is 55%-75%. The degree of difference in power distribution between the two cycles is analyzed; a greater difference indicates higher redundancy adjustment potential. The power transfer between cycles is measured; the transfer amount equals the redundant power consumption of the sleep cycle minus the power gap of the active cycle. Redundancy adjustment parameters are generated based on the power distribution differences, including four dimensions: adjustable power consumption, adjustment response time, adjustment frequency, and adjustment stability. The numerical range of the redundancy adjustment parameters is set: adjustable power consumption should not exceed 80% of the average redundancy of the sleep cycle, and the adjustment response time should be controlled within 15 minutes. Considering the impact of cycle switching on the regulation effect, the regulation intensity should be appropriately reduced during the cycle switching period to avoid the regulation operation interfering with the normal cycle switching process.

[0067] Redundant power components are converted into power-enhanced buffers using redundancy adjustment parameters. These components are reconfigured using the adjustable range and priority of the redundancy adjustment parameters. Excess power consumption from the redundant components is extracted and aggregated into a power-enhanced buffer pool. The buffer pool is partitioned according to the spatial location and service type of the redundant components, with each partition having a service radius of 50 meters to ensure rapid response to power demands from adjacent grids. Upper and lower limits are set for the buffer pool's capacity. The upper limit, set at 90% of the total redundant power consumption, prevents excessive caching from affecting system stability. The lower limit, set at 30% of the total redundant power consumption, ensures sufficient adjustment capability for the buffer. A power consumption allocation strategy for the buffer pool is implemented, allowing for rapid acquisition of additional power support from the buffer pool when a grid's power demand surges, with a response time controlled within 5 minutes. A priority allocation mechanism is established, prioritizing power support for adjacent and same-level grids; cross-level calls require an approval process. A refresh mechanism for the power-enhanced buffer pool is set up, updating the buffer pool's capacity and available power consumption every 2 hours. Buffer utilization efficiency is monitored; the buffer pool size is reduced when utilization remains below 30%, and its capacity is increased when utilization exceeds 90%.

[0068] An adaptive parameter adjustment data is generated by utilizing a power-enhanced cache to dynamically optimize the adaptive parameters. Additional power resources provided by the cache are used to dynamically optimize the parameters. When the load-power correlation coefficient of a grid changes, power resources are retrieved from the cache for parameter compensation. Target values ​​for parameter adjustment are determined to keep the adjusted correlation coefficient within a reasonable range, avoiding over- or under-response. A gradual parameter adjustment strategy is implemented, breaking down large parameter changes into multiple small steps, with each adjustment increment controlled within 10%. Service quality indicators within the grid, including response time, throughput, and availability, are monitored in real time during adjustment to ensure that the adjustment operation does not degrade service performance. A feedback mechanism for parameter adjustment is established to automatically correct subsequent adjustment strategies based on the adjustment effect; the adjustment intensity is reduced when the effect is poor, and the adjustment frequency is appropriately increased when the effect is good. Boundary conditions for parameter adjustment are set to prevent parameters from exceeding safe operating ranges during adjustment; the adjustment range of the correlation coefficient is limited to ±30%. The adjusted adaptive parameters are compared with the original parameters to generate a parameter change report, including the adjustment magnitude, adjustment duration, and effect evaluation. Organize all power consumption data generated by adjustment operations, including four dimensions: power consumption before adjustment, power consumption after adjustment, power consumption contributed by the buffer, and net power consumption change.

[0069] Step S150: Construct a power control matrix based on the power consumption data and power event triggering threshold; detect power imbalance nodes in the power consumption data within the power control matrix; pair the power imbalance nodes with the power consumption data to obtain a power optimization control group; and generate power management instructions based on the power optimization control group.

[0070] In some embodiments, constructing a power control matrix based on the adjusted power consumption data and the power consumption event triggering threshold includes: identifying a control deviation region based on the adjusted power consumption data and the power consumption event triggering threshold; extracting deviation correction parameters from the control deviation region; adjusting the control deviation region to a control stability adjustment point using the deviation correction parameters; and constructing a power control matrix using the control stability adjustment point.

[0071] Control deviation areas are identified based on power consumption data and power event trigger thresholds. The power consumption status of each rack in the power consumption data is compared one by one with the corresponding power event trigger threshold to identify rack areas with deviations. When the power consumption of a Web server rack still exceeds the trigger threshold range after adjustment, that rack is marked as a candidate control deviation area. The direction and magnitude of the deviation are analyzed: database server racks with power consumption above the upper trigger threshold are considered positive deviations, while file storage racks with power consumption below the lower trigger threshold are considered negative deviations. The degree of deviation is measured; the degree of deviation equals the difference between the actual power consumption and the trigger threshold divided by the threshold range. Adjacent racks with deviations are merged through connected component analysis to form continuous control deviation areas. The causes of deviation areas are identified, including three main situations: insufficient adjustment, improper threshold setting, and external interference. Insufficient adjustment means that the amplitude or speed of rack power consumption adjustment cannot meet control requirements; improper threshold setting means that the trigger threshold does not match the actual needs of the rack; and external interference means that unexpected load changes affect the control effect. The spatial distribution characteristics of control deviation areas are statistically analyzed, including the geometry and coverage area of ​​the deviation rack areas. Rack areas with deviations exceeding 30% for more than one hour are marked as severe deviation areas. During nighttime maintenance in the data center, the backup storage system suddenly initiates a large-scale data migration task, causing the power consumption demand of the storage rack area to far exceed the preset threshold, forming a typical control deviation area.

[0072] Extract deviation correction parameters from the control deviation area. For each identified control deviation rack area, analyze the specific characteristics of the deviation and the correction requirements. Measure the correction gap in the deviation rack area; the correction gap equals the difference between the current deviation level and the target control state. Analyze the time requirements for deviation correction: deviations in critical business server racks need to be corrected within 10 minutes, while deviations in auxiliary service racks can be corrected within 30 minutes. Extract the spatial range of deviation correction, determining the number of racks affected and the radius of influence of the correction operation. Identify the constraints on deviation correction, including available power resource limitations, rack operation safety constraints, and power supply system capacity constraints. Generate deviation correction parameters including four core elements: correction intensity, correction direction, correction speed, and correction duration. Correction intensity reflects the magnitude of the rack power adjustment operation; correction direction indicates the direction of increase or decrease in power adjustment; correction speed indicates the speed of the correction process; and correction duration indicates the length of the correction operation. Consider the impact of deviation correction on surrounding racks, set boundary values ​​for the correction parameters to prevent the correction operation from causing new deviations in adjacent racks. For server racks with positive deviations, power consumption reduction correction parameters are required; for storage server racks with negative deviations, power consumption increase correction parameters are required. The extracted deviation correction parameters are then sorted according to their correction priority, with the correction parameters for critical business racks having the highest priority.

[0073] The deviation correction parameters are used to adjust the control deviation area to a control stability adjustment point. These parameters are then applied to the corresponding control deviation rack area, and deviation correction operations are performed. Based on the correction intensity and direction in the correction parameters, the power distribution of the deviation racks is adjusted accordingly. A progressive correction strategy is adopted, breaking down large deviation corrections into multiple small steps, with each step's correction magnitude controlled within 25% of the total correction amount. During the correction process, the power consumption status of the servers within the racks is monitored, and the correction speed is slowed when the power consumption level approaches the target control range. A feedback control mechanism is used to adjust the correction parameters in real time, dynamically correcting subsequent correction operations based on the rack power consumption adjustment effect. A stable equilibrium point is identified during the correction process; when the rack deviation decreases to an acceptable range and remains stable, the rack location is marked as the control stability adjustment point. The stability characteristics of the control stability adjustment point are measured, including rack power consumption fluctuation amplitude, stability duration, and anti-interference capability. The control stability adjustment point is spatially located to accurately determine its rack coordinates within the data center.

[0074] A power consumption control matrix is ​​constructed by controlling stable adjustment points. The matrix framework for power consumption control is constructed using the rack location information and stability characteristics of these stable adjustment points. Each stable adjustment point serves as the control base point of the matrix, and corresponding control units are set within the matrix. The weight values ​​of matrix elements are determined based on the stability strength of the racks at the adjustment points; web server racks with higher stability correspond to larger weight values. The control influence relationships between the racks at the adjustment points are analyzed and transformed into off-diagonal elements in the matrix. A row-column structure of the matrix is ​​constructed, where row coordinates represent the rack area of ​​the control input and column coordinates represent the target rack of the control output. The matrix elements are filled with values, and the specific values ​​of the elements are determined using the parameter characteristics of the stable adjustment points. Blank areas in the matrix are addressed; for rack areas not covered by adjustment points, interpolation methods are used to fill in the matrix elements. The power consumption control matrix is ​​divided into blocks according to data center areas, with the control characteristics of server room areas, storage room areas, and network equipment areas being relatively independent. A complete power consumption control matrix is ​​formed, containing the power consumption control relationships and adjustment mechanisms for each rack area in the data center.

[0075] The system detects power imbalance nodes within the power consumption data in the power consumption control matrix. It scans for abnormal cabinets in the power consumption control matrix, marking cabinets with power consumption status values ​​exceeding a set threshold as potential imbalance candidates. The imbalance characteristics of these candidate cabinets are analyzed, including three main types: power supply-demand mismatch, uneven power distribution, and power regulation lag. Power supply-demand mismatch manifests as a significant difference between the actual power consumption demand and the allocated power consumption for database server cabinets; uneven power distribution manifests as excessively large differences in power distribution between adjacent web server cabinets; and power regulation lag manifests as excessively long power regulation response times for video processing server cabinets. The degree of imbalance is measured, equal to the deviation between the actual power consumption state and the expected power consumption state. The spatial distribution characteristics of various types of imbalanced cabinets are statistically analyzed to identify clustering patterns and diffusion trends. Severely imbalanced database server cabinets may affect the power consumption stability of the entire data processing area, while slightly imbalanced file storage cabinets only affect local storage performance. Analyzing the temporal evolution characteristics of unbalanced server racks reveals that persistently unbalanced compute server racks require long-term adjustments, while intermittently unbalanced cache server racks can be resolved through short-term intervention.

[0076] In some embodiments, the step of pairing the power imbalance node with the regulated power consumption data to obtain a power optimization control group includes: identifying and locating the imbalance source of the power imbalance node and the regulated power consumption data to generate transmission imbalance and processing imbalance; using the transmission imbalance to evaluate the impact of the processing imbalance to form an imbalance matrix; generating a dominant imbalance vector through eigenvalue decomposition of the imbalance matrix; and implementing balance correction according to the dominant imbalance vector to obtain a power optimization control group.

[0077] This study identifies and locates the sources of power imbalances, including transmission and processing imbalances, by analyzing the imbalance characteristics of power-imbalanced cabinets and the power consumption change patterns in the regulated power consumption data. It tracks the root cause of the imbalance by correlating the imbalance characteristics of cabinets with power consumption imbalances with the power consumption change patterns in the regulated power consumption data. Power flow analysis identifies the propagation path of the imbalance, distinguishing between the direct source cabinet and intermediate propagation links. Imbalance sources are categorized into two basic types: transmission imbalance and processing imbalance. Transmission imbalance stems from power line losses or transmission blockages during power transfer between cabinets, while processing imbalance arises from a mismatch between the power processing capacity and demand of servers within the cabinet. Analysis of transmission imbalance characteristics reveals that it typically manifests as attenuation or delay of power consumption along the power transmission path. Key transmission cabinets with transmission imbalances are identified, as they are usually located at power distribution bottlenecks. Analysis of processing imbalance characteristics reveals that processing imbalance typically manifests as a discrepancy between the power supply within the cabinet and the actual power consumption of the servers. Key processing cabinets with processing imbalances are identified, corresponding to areas with insufficient or excessive server power processing capacity. The intensity distribution of transmission and processing imbalances is measured, and the greater the imbalance, the more severe the impact on the data center. In the storage cluster area of ​​the data center, when cache servers transmit large amounts of data to disk arrays, the capacity limitation of the power transmission channel causes transmission imbalance, while the insufficient processing capacity of the disk array racks causes processing imbalance.

[0078] An imbalance matrix is ​​generated by performing an impact analysis on processing imbalance using transmission imbalance. The mutual influence between transmission and processing imbalances is analyzed, quantifying the degree of influence of transmission imbalance on processing imbalance. The impact degree formula I = α × T_loss × P_demand is used for calculation, where I is the impact degree, α is the impact coefficient, T_loss is the intensity of transmission imbalance, and P_demand is the demand for processing imbalance. An impact propagation model from transmission imbalance to processing imbalance is established, analyzing the time delay and spatial attenuation characteristics of the propagation. The impact intensity of different types of transmission imbalance on various types of processing imbalance is measured, forming a numerical matrix of impact intensity. The basic framework of the imbalance matrix is ​​constructed, with rows corresponding to the source cabinets of transmission imbalance and columns corresponding to the target cabinets of processing imbalance. The elements of the imbalance matrix are filled according to the impact degree analysis results, with cabinets with higher impact degrees corresponding to larger matrix element values. The diagonal elements in the processing imbalance matrix reflect the self-influence degree of the same imbalanced cabinet. The off-diagonal elements in the processing imbalance matrix reflect the cross-influence between different imbalanced cabinets. Normalize the unbalanced matrix to ensure its numerical stability and convergence.

[0079] The dominant imbalance vector is generated through eigenvalue decomposition of the imbalance matrix. Eigenvalue decomposition is performed on the imbalance matrix to obtain eigenvalues ​​and corresponding eigenvectors. The principal eigenvalues ​​of the imbalance matrix are identified; the largest eigenvalue reflects the dominant influence mode of the imbalance system. The eigenvectors corresponding to the principal eigenvalues ​​are extracted; these are called the dominant imbalance vectors. The element distribution of the dominant imbalance vector is analyzed; the magnitude of the vector elements reflects the contribution of each imbalanced cabinet to the dominant imbalance mode. Key elements in the dominant imbalance vector are identified; these key elements correspond to the imbalanced cabinets that have the greatest impact on the dominant imbalance mode. The directional characteristics of the dominant imbalance vector are analyzed; the vector direction indicates the main direction of imbalance propagation. The magnitude of the dominant imbalance vector reflects the overall strength of the dominant imbalance mode. The dominant imbalance vector is normalized to facilitate subsequent balancing and correction operations. The eigenvectors corresponding to the secondary eigenvalues ​​are analyzed; these secondary eigenvectors reflect the secondary influence modes of the imbalance system. The dominant imbalance vector is mapped to the original imbalanced cabinets to determine the position of each imbalanced cabinet in the dominant imbalance mode.

[0080] A power consumption optimization control group is obtained by implementing balance correction according to the dominant imbalance vector. A targeted balance correction strategy is formulated using the direction and intensity information of the dominant imbalance vector. Based on the numerical values ​​of each element in the dominant imbalance vector, the key rack areas and priority order for balance correction are determined. Key elements in the dominant imbalance vector are targeted for correction, with the correction intensity proportional to the element value. A balance correction operation plan is designed, including three basic methods: power redistribution, transmission path optimization, and processing capacity adjustment. Power redistribution achieves balance by adjusting the power distribution ratio of each rack; transmission path optimization reduces transmission imbalance by improving the power transmission efficiency between racks; and processing capacity adjustment resolves processing imbalance by enhancing or weakening the power consumption processing capacity of servers within a rack. Balance correction operations are implemented, gradually eliminating imbalances in each rack according to the guidance of the dominant imbalance vector. The effectiveness of the balance correction is monitored, and the effectiveness of the correction operation is judged by changes in the degree of imbalance. The intensity and direction of the balance correction are adjusted, and subsequent correction operations are optimized based on feedback from the correction effect. Successfully corrected imbalanced racks are combined with corresponding adjustment methods to form a power consumption optimization control group. Each power optimization control group consists of four parts: target unbalanced cabinet, correction operation type, correction parameter settings, and expected correction effect.

[0081] Power management instructions are generated based on the power optimization control group. Using the target imbalanced cabinet information from the power optimization control group, the specific cabinet location and type requiring power adjustment are identified. The basic operation mode of the instructions is determined based on the correction operation type: power redistribution operations correspond to power adjustment instructions, transmission path optimization operations correspond to transmission improvement instructions, and processing capacity adjustment operations correspond to processing adjustment instructions. Correction parameter settings are converted into executable operation command parameters, including specific values ​​such as adjustment range, adjustment speed, and adjustment duration. For the video processing server cabinet, which is the target imbalanced cabinet, a processing adjustment instruction is generated based on its processing capacity adjustment correction operation type. The adjustment parameters are: increase power allocation by 150 watts, adjustment speed of 25 watts per minute, and duration of 6 minutes. Success criteria for management instructions are set, and the target power state to be achieved after instruction execution is determined based on the expected correction effect. An instruction execution priority ranking is established, determining the execution order based on the importance level of the target imbalanced cabinet and the urgency of the correction operation type. An instruction execution effect monitoring mechanism is introduced, comparing the actual execution effect with the expected correction effect in real time. Instruction adjustment is triggered when the deviation exceeds 10%. The instructions include execution feedback requirements, demanding that the rack management unit report the status changes of the target unbalanced rack and the achievement of the expected correction effect after completing the instructions. The generated power management instructions are categorized according to correction operation type and execution sequence to form a structured instruction execution scheme.

[0082] To implement the data center power distribution optimization method corresponding to the above method embodiments, in order to achieve the corresponding functional and technical effects. See also Figure 2 , Figure 2 This diagram illustrates a structural block diagram of a data center power distribution optimization system 200 according to an embodiment of this application. For ease of explanation, only the parts relevant to this embodiment are shown. The data center power distribution optimization system 200 provided in this embodiment includes:

[0083] The correlation and fusion module 201 is used to monitor the power consumption signals of computing nodes and environmental sensor data in the data center, extract power demand information from the power consumption signals of computing nodes, and correlate and fuse the power demand information with the environmental sensor data to form an environment-power consumption correlation field.

[0084] The probability analysis module 202 is used to perform probability inference analysis using the environment-power consumption correlation field to obtain probability power consumption parameters, perform event evaluation using the probability power consumption parameters to determine power consumption event trigger thresholds, and perform power consumption configuration based on the power consumption event trigger thresholds and the probability power consumption parameters to determine dynamic power consumption management boundaries.

[0085] Quality identification module 203 is used to identify service quality requirements based on the dynamic power management boundary, generate a quality requirement matrix, generate hierarchical power consumption parameters based on the quality requirement matrix, and construct a power consumption level allocation configuration using the hierarchical power consumption parameters.

[0086] The redundancy conversion module 204 is used to perform load-power correlation to obtain adaptive parameters for the power level allocation configuration, identify power redundancy components from the adaptive parameters to generate redundancy adjustment parameters, use the redundancy adjustment parameters to convert the power redundancy components into power enhancement buffers, and use the power enhancement buffers to adaptively adjust the adaptive parameters to form adjusted power consumption data.

[0087] The control optimization module 205 is used to construct a power control matrix based on the adjusted power consumption data and the power consumption event triggering threshold, detect power imbalance nodes in the adjusted power consumption data in the power control matrix, perform power balance pairing between the power imbalance nodes and the adjusted power consumption data to obtain a power optimization control group, and generate power management instructions based on the power optimization control group.

[0088] The data center power distribution optimization system 200 described above can implement a data center power distribution optimization method according to the above method embodiments. The options in the above method embodiments are also applicable to this embodiment, and will not be detailed here. The remaining content of this application embodiment can be referred to the content of the above method embodiments, and will not be repeated in this embodiment.

[0089] The purpose of the above embodiments is to reproduce and derive the technical solution of the present invention by way of example, and to fully describe the technical solution, purpose and effect of the present invention. The purpose is to enable the public to have a more thorough and comprehensive understanding of the disclosure of the present invention, and not to limit the scope of protection of the present invention.

[0090] The above embodiments are not an exhaustive list based on the present invention, and there may be many other embodiments not listed. Any substitutions and improvements made without departing from the concept of the present invention are within the protection scope of the present invention.

Claims

1. A method for optimizing power distribution in a data center, characterized in that, include: Monitor the power consumption signals of computing nodes and environmental sensor data in the data center, extract power demand information from the computing node power consumption signals, and correlate and fuse the power demand information with the environmental sensor data to form an environment-power consumption correlation field; The probabilistic power consumption parameters are obtained by performing probabilistic inference analysis using the environment-power consumption correlation field. Event evaluation is performed using the probabilistic power consumption parameters to determine the power event triggering threshold. Based on the power event triggering threshold and the probabilistic power consumption parameters, power configuration is performed to determine the dynamic power management boundary. Based on the dynamic power management boundary, service quality requirements are identified to generate a quality requirement matrix. Based on the quality requirement matrix, hierarchical power consumption parameters are generated, and power level allocation configuration is constructed using the hierarchical power consumption parameters. The process involves: performing load-power correlation to obtain adaptive parameters for the power level allocation configuration; identifying power redundancy components from the adaptive parameters to generate redundancy adjustment parameters; constructing a redundancy time axis using the adaptive parameters; identifying redundant peak points from the adaptive parameters and mapping them to the redundancy time axis to form peak time markers; dividing the redundancy time axis into active periods and sleep periods based on the peak time markers; comparing the power distribution characteristics corresponding to the active periods and sleep periods to generate redundancy adjustment parameters; using the redundancy adjustment parameters to convert the power redundancy components into power enhancement buffers; and using the power enhancement buffers to adaptively adjust the adaptive parameters to form adjusted power consumption data. Specifically, dividing the redundancy time axis into active periods and sleep periods based on the peak time markers includes: identifying period transition failure points based on the peak time markers; extracting transition delay parameters from the period transition failure points; performing time-series reconstruction on the period transition failure points using the transition delay parameters to obtain a period coordination optimization region; and dividing the redundancy time axis into active periods and sleep periods based on the period coordination optimization region. A power control matrix is ​​constructed based on the regulated power consumption data and the power consumption event trigger threshold. Power imbalance nodes within the regulated power consumption data are detected in the power control matrix. The power imbalance nodes are paired with the regulated power consumption data to obtain a power optimization control group. Power management instructions are generated based on the power optimization control group.

2. The method according to claim 1, characterized in that, The process of obtaining probabilistic power consumption parameters through probabilistic inference analysis using the environment-power consumption correlation field includes: The environment-power consumption correlation field is used to identify sparse regions of probability distribution; Sparse enhancement parameters are generated based on the sparse regions of the probability distribution; The sparse region is reconfigured using the aforementioned sparsity enhancement parameters to form a probability inference optimization region. The probabilistic power consumption parameters are obtained based on the probabilistic inference optimization region.

3. The method according to claim 1, characterized in that, The step of determining the dynamic power management boundary by performing power configuration based on the power event trigger threshold and the probabilistic power parameters includes: A matching analysis is performed between the power consumption event trigger threshold and the probabilistic power consumption parameter to generate a matching degree distribution; From the matching degree distribution, identify the configuration deviation region and generate deviation identification parameters; The configuration deviation region is transformed into a boundary calibration optimization region using the deviation identification parameter. The dynamic power management boundary is determined based on the boundary calibration optimization region.

4. The method according to claim 1, characterized in that, The step of identifying quality of service requirements and generating a quality of service requirement matrix based on the dynamic power management boundary includes: Service quality requirements are identified and mismatched areas are determined based on the dynamic power management boundary. Extract mismatch compensation parameters from the demand mismatch area; The mismatch compensation parameters are used to correct the demand mismatch area and generate demand correction adjustment points. A quality requirement matrix is ​​generated based on the required adjustment points.

5. The method according to claim 1, characterized in that, The step of constructing a power control matrix based on the adjusted power consumption data and the power consumption event triggering threshold includes: Based on the adjusted power consumption data and the power consumption event trigger threshold, the control deviation region is identified; Extract deviation correction parameters from the control deviation region; The control deviation range is adjusted to the control stability adjustment point using the aforementioned deviation correction parameters. A power consumption control matrix is ​​constructed using the control stability adjustment point.

6. The method according to claim 1, characterized in that, The step of pairing the power imbalance node with the adjusted power consumption data to obtain a power optimization control group includes: The imbalance source is located and identified for the power imbalance node and the adjusted power consumption data, resulting in transmission imbalance and processing imbalance. An imbalance matrix is ​​formed by evaluating the impact of the transmission imbalance on the processing imbalance. The dominant imbalance vector is generated by eigenvalue decomposition of the imbalance matrix. The power consumption optimization control group is obtained by performing balance correction according to the dominant imbalance vector.

7. The method according to claim 3, characterized in that, The step of converting the configuration deviation region into a boundary calibration optimization region using the deviation identification parameter includes: Based on the deviation identification parameters and the configured deviation region, the deviation intensity characteristics are identified to determine the deviation analysis window, wherein the deviation intensity characteristics include deviation frequency, duration intensity, and range of influence. The deviation change process is traced along the deviation analysis window to form a deviation change trajectory diagram; Extract the position coordinates of the calibration adjustment point from the deviation change trajectory diagram; The calibration effect of the position coordinates is arranged to determine the boundary calibration optimization zone.

8. A data center power distribution optimization system, characterized in that, include: The correlation and fusion module is used to monitor the power consumption signals of computing nodes and environmental sensor data in the data center, extract power demand information from the power consumption signals of computing nodes, and correlate and fuse the power demand information with the environmental sensor data to form an environment-power consumption correlation field. The probability analysis module is used to perform probability inference analysis using the environment-power consumption correlation field to obtain probability power consumption parameters, perform event evaluation using the probability power consumption parameters to determine power consumption event trigger thresholds, and perform power consumption configuration based on the power consumption event trigger thresholds and the probability power consumption parameters to determine dynamic power consumption management boundaries. The quality identification module is used to identify service quality requirements based on the dynamic power management boundary, generate a quality requirement matrix, generate hierarchical power consumption parameters based on the quality requirement matrix, and construct a power level allocation configuration using the hierarchical power consumption parameters. A redundancy conversion module is used to perform load-power correlation to obtain adaptive parameters for the power level allocation configuration, identify power redundancy components from the adaptive parameters, and generate redundancy adjustment parameters. This includes: constructing a redundancy time axis using the adaptive parameters; identifying redundant peak points from the adaptive parameters and mapping them to the redundancy time axis to form peak time markers; dividing the redundancy time axis into active periods and sleep periods based on the peak time markers; comparing the power distribution characteristics corresponding to the active periods and sleep periods to generate redundancy adjustment parameters; using the redundancy adjustment parameters to convert the power redundancy components into power enhancement buffers; and using the power enhancement buffers to adaptively adjust the adaptive parameters to form adjusted power consumption data. Specifically, dividing the redundancy time axis into active periods and sleep periods based on the peak time markers includes: identifying period conversion failure points based on the peak time markers; extracting conversion delay parameters from the period conversion failure points; performing time-series reconstruction on the period conversion failure points using the conversion delay parameters to obtain a period coordination optimization region; and dividing the redundancy time axis into active periods and sleep periods based on the period coordination optimization region. The control optimization module is used to construct a power control matrix based on the adjusted power consumption data and the power consumption event triggering threshold, detect power imbalance nodes in the adjusted power consumption data in the power control matrix, perform power balance pairing between the power imbalance nodes and the adjusted power consumption data to obtain a power optimization control group, and generate power management instructions based on the power optimization control group.

Images