Electric calculation center load characteristic evaluation method based on macroscopic quantitative analysis

Abstract

The invention relates to the technical field of power data processing, in particular to a computing center load characteristic evaluation method based on macroscopic quantitative analysis, which comprises the following steps: acquiring a historical operation data set of a computing center, including computing power task scheduling logs, load demand records, intermittent energy output monitoring data and energy storage system state information; analyzing the data set, calculating a dynamic coupling degree index, identifying a high-priority task starting and energy attenuation overlapping time period, and generating a coupling event mark sequence with a time window identifier; loading a sequence to construct a load behavior dynamic model, and outputting a load curve probability envelope band through probability fluctuation simulation; and analyzing the envelope band to extract the load sudden change intensity, the fluctuation duration coefficient and the elastic adjustment margin index, and generating an optimization strategy report associated with the time window identifier. According to the method, the problem of load curve microscopic distortion identification misalignment is solved by quantifying the coupling effect of tasks and energy fluctuation, and the evaluation precision and decision reliability in a dynamic fluctuation environment are improved.

Description

Technical Field

[0001] This invention relates to the field of power data processing technology, and in particular to a method for evaluating the load characteristics of a computer center based on macroscopic quantitative analysis. Background Technology

[0002] The application of macro-level quantitative analysis methods in data center load characteristic assessment involves processing large amounts of historical operational data to identify overall load trends and cyclical patterns. Key variables and correlations are extracted through data aggregation to construct load behavior models, supporting facility resource optimization decisions and improving energy efficiency while controlling operating costs. For example, the analysis results reveal peak load characteristics and off-peak cycle dynamics, guiding capacity planning adjustments, promoting the rational allocation of computing resources, and avoiding unnecessary expenditures. Long-term applications enhance system stability, provide input for intelligent monitoring frameworks, and ultimately serve sustainable development goals.

[0003] During the load curve characteristic assessment of intelligent computing centers, a strong spatiotemporal coupling effect exists between the dynamic scheduling mechanism of computing tasks and the random fluctuations in renewable energy output, leading to nonlinear abrupt changes in the load curve. Specifically, high-priority, uninterruptible computing tasks are concentrated and launched within a specific time window, while renewable energy output is subject to unpredictable attenuation due to natural conditions. The superposition of these two factors forces the system to become unbalanced between rigid load demand and flexible adjustment capabilities. For example, an intelligent computing center in northern Hebei deployed an artificial intelligence training cluster. At 15:00 on a certain day, a real-time inference task required 8MW of constant power support. At the same time, the output of photovoltaic power in the region decreased due to sudden cloud cover, and wind power output also decreased synchronously. At this time, the system had to deal with three contradictions simultaneously: uninterruptible tasks had to run at full power, adjustable data analysis tasks were limited by a two-hour time window and could not be delayed, and the energy storage system's state of charge was at a low point in the charging cycle. Under multiple constraints, the actual load curve showed a dramatic fluctuation at 15:30, dropping sharply from 9.2MW to 4.1MW and then jumping back to 7.5MW, exceeding the prediction range of historical statistical models. Existing macroscopic analysis methods can only identify overall trends, but cannot analyze microscopic distortions caused by the combined effects of task urgency, new energy fluctuation cycles, and energy storage response lags, resulting in load characteristic assessments deviating from the actual operating conditions. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method for evaluating the load characteristics of computer centers based on macroscopic quantitative analysis, which solves the problem that the characteristics of the load curve of computer centers are difficult to accurately evaluate due to the superposition of the dynamic changes in computing power tasks and the intermittent fluctuations in energy.

[0005] To solve the above-mentioned technical problems, the specific details of the present invention are as follows: The present invention provides a method for evaluating the load characteristics of computer centers based on macroscopic quantitative analysis, comprising: Step 1: Collect historical operation datasets from the computer center. The historical operation datasets include computing task scheduling logs, load demand records, intermittent energy output monitoring data, and energy storage system status information. Step 2: Standardize the historical operation dataset collected in Step 1 to generate a load feature dataset with a uniform time granularity. The load feature dataset distinguishes between rigid load feature subsets and flexible load feature subsets. Step 3: Analyze the load characteristic dataset to calculate the dynamic coupling index. When the start time of a high-priority task overlaps with the period of intermittent energy output decay, generate a coupling event label sequence that includes time window identifiers and imbalance characteristics. Step 4: Load the dynamic coupling degree index and coupling event label sequence generated in Step 3, construct a dynamic model of load behavior, fit the long-term baseline macro trend of the load characteristic dataset, add the imbalance feature perturbation of the coupling event label sequence to the macro trend, and output the probability envelope of the load curve through probability fluctuation simulation. Step 5: The load behavior dynamic model is used to analyze the probability envelope of the load curve, extract the load curve characteristic indicators and generate a comprehensive load characteristic assessment report. The load curve characteristic indicators include load mutation intensity, fluctuation duration coefficient and elastic adjustment margin. The comprehensive load characteristic assessment report includes a visualized risk matrix and optimization strategies. The optimization strategies are associated with the time window markers in the event tag sequence.

[0006] Furthermore, in the computer center load characteristic assessment method based on macroscopic quantitative analysis described in this invention, step 2 includes: Perform sliding window outlier detection on the historical dataset collected in step 1 to obtain the dataset processed by outlier detection. Time alignment technology was applied to the outlier detection-processed dataset to unify different collection frequencies to a 5-minute granularity, generating a standardized dataset. The standardized dataset is divided into a rigid load feature subset and a flexible load feature subset. The rigid load feature subset includes GPU cluster power data, while the flexible load feature subset includes storage device power consumption data.

[0007] Furthermore, in the computer center load characteristic assessment method based on macroscopic quantitative analysis described in this invention, step 3 includes: The dynamic imbalance index DI is defined based on the rigid load values ​​in the rigid load feature subset, where DI = (actual rigid load value - predicted value) × energy shortage rate; The load feature dataset is processed by scanning with a moving time window. When the start of a high-priority task overlaps with energy decay, the DI value calculated in the current time window is inserted into the event marker. Output a sequence of event markers with a DI index, where each event marker includes a time window identifier and a corresponding imbalance feature value.

[0008] Furthermore, the computer center load characteristic assessment method based on macroscopic quantitative analysis described in this invention assesses the probability envelope of the load curve using a dynamic load behavior model, including: Load the standardized dataset generated in step 2 and the rigid load feature subset and flexible load feature subset obtained by its partitioning, and apply the ARIMA algorithm to fit the macro trend baseline on the standardized dataset; Parse the event-tagged sequence with DI index output from step 3, and extract the event time window identifier and the corresponding DI value; The event time window identifier and DI value are injected into the two-layer probabilistic framework: the macro trend baseline is fixed in the upper framework, and the intermittent energy random fluctuation scenario and task urgency probability distribution are incorporated into the lower framework. The two-layer probabilistic framework is processed by Monte Carlo simulation to output the load curve probability envelope.

[0009] Furthermore, in the computer center load characteristic assessment method based on macroscopic quantitative analysis described in this invention, step 5 includes: The load curve characteristic index is extracted from the probability envelope of the load curve, and the dynamic imbalance index is obtained from the event tag sequence with DI index output in step 3. Based on time window identification, load mutation intensity and dynamic imbalance index are matched to load curve characteristic indicators. High-risk periods that simultaneously meet the criteria of dynamic imbalance index DI > 0.8 and load mutation intensity > 500 kW / min are identified, and a period marker set is generated. A three-color risk matrix is ​​generated based on the mapping of time period marker sets to identify the risk level in the time dimension.

[0010] Furthermore, the optimization strategy of the computer center load characteristic assessment method based on macroscopic quantitative analysis described in this invention includes: Recommendations for task migration windows and energy storage power configuration thresholds; The task migration window recommendation includes: associating time window identifiers in the event tag sequence with DI index to guide flexible load offset operations to avoid overlapping peaks; The energy storage power configuration thresholds include: setting power dispatch boundary values ​​for different risk stages based on the risk level of the three-color risk matrix.

[0011] Furthermore, the computer center load characteristic assessment method based on macroscopic quantitative analysis described in this invention includes the following parameters for configuring the moving time window scanning: Set the scan step size to 1 minute and the time window width to 15 minutes, and detect high-frequency event mutation regions by continuously overlapping moving windows.

[0012] Furthermore, the computer center load characteristic assessment method based on macroscopic quantitative analysis described in this invention, with a Monte Carlo simulation-processed two-layer probabilistic framework, includes: Perform 500 random scenario calculations to obtain 500 sets of load fluctuation value output sequences; A probability density statistical algorithm was applied to the output sequence of 500 sets of load fluctuation values, and the 50th percentile value of the calculation result distribution was taken to generate the P50 prediction line. The standard deviation σ is calculated based on the output sequence of 500 sets of load fluctuation values, and the upper and lower boundaries of the power fluctuation range are generated based on the standard deviation σ value. The P50 prediction line is combined with the upper and lower boundaries of the power fluctuation range to form the probability envelope of the final output load curve.

[0013] Furthermore, the computer center load characteristic assessment method based on macroscopic quantitative analysis described in this invention also includes: The task time constraint data of the flexible load feature subset obtained in step 2 is analyzed, and the flexible task time window parameter representing the task delayability is extracted. Based on the flexible task time window parameter, the task migration offset time range is set to a range of 0.5-1.5 hours. The time range value is input into the task migration window suggested operation to control the flexible load offset time margin.

[0014] Furthermore, the computer center load characteristic assessment method based on macroscopic quantitative analysis described in this invention further includes: parsing the task time constraint data of the flexible load characteristic subset obtained in step 2, extracting the flexible task time window parameter characterizing the task's delayability, setting the task migration offset time range based on the flexible task time window parameter, limiting it to a range of 0.5-1.5 hours, inputting the time range value into the task migration window suggested operation, and controlling the flexible load offset time margin.

[0015] Beneficial effects of this invention; This invention quantifies the spatiotemporal coupling effect of computing power task scheduling and intermittent energy fluctuations through a dynamic coupling degree index, generating a coupled event marker sequence with time window identifiers. This captures the micro-scale distortion features of the load curve that existing methods cannot identify, solving the core problem in the background technology where drastic load fluctuations exceed the prediction range of historical statistical models. A dynamic load behavior model using a two-layer probabilistic framework integrates macro-trend baselines and micro-disturbance factors, outputting a probability envelope band through Monte Carlo simulation. This significantly improves the load prediction accuracy under scenarios where the randomness of new energy output and the urgency of tasks overlap, overcoming the defect in the background technology where assessments deviate from the actual operating state. Based on a decision-making closed-loop mechanism using a risk matrix and optimization strategy, the time identifier and risk level of the event marker sequence are associated with task migration window suggestions and energy storage power configuration thresholds. This achieves dynamic matching between resource scheduling strategies and load characteristic assessment results, effectively reducing the operational risks and energy losses of computing centers in dynamic fluctuation environments. It forms a seamless technical chain from feature identification and model prediction to strategy execution, directly addressing the pain point of system regulation failure under multiple constraints in the background technology. Attached Figure Description

[0016] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on the drawings without creative effort.

[0017] Figure 1 A flowchart of a method for evaluating the load characteristics of a computer center based on macroscopic quantitative analysis, provided for embodiments of the present invention. Detailed Implementation

[0018] To make the technical solution of the present invention clearer, the present invention will be clearly and completely described below with reference to specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. The present invention provided by various embodiments will be described in detail below with reference to the accompanying drawings. To better understand the purpose of the present invention, the present invention will be described in further detail below.

[0019] Please see Figure 1 The present invention provides a method for evaluating the load characteristics of computer centers based on macroscopic quantitative analysis, comprising: Step 1: Collect historical operation datasets from the computer center. The historical operation datasets include computing task scheduling logs, load demand records, intermittent energy output monitoring data, and energy storage system status information. Step 2: Standardize the historical operation dataset collected in Step 1 to generate a load feature dataset with a uniform time granularity. The load feature dataset distinguishes between rigid load feature subsets and flexible load feature subsets. Step 3: Analyze the load characteristic dataset to calculate the dynamic coupling index. When the start time of a high-priority task overlaps with the period of intermittent energy output decay, generate a coupling event label sequence that includes time window identifiers and imbalance characteristics. Step 4: Load the dynamic coupling degree index and coupling event label sequence generated in Step 3, construct a dynamic model of load behavior, fit the long-term baseline macro trend of the load characteristic dataset, add the imbalance feature perturbation of the coupling event label sequence to the macro trend, and output the probability envelope of the load curve through probability fluctuation simulation. Step 5: The load behavior dynamic model is used to analyze the probability envelope of the load curve, extract the load curve characteristic indicators and generate a comprehensive load characteristic assessment report. The load curve characteristic indicators include load mutation intensity, fluctuation duration coefficient and elastic adjustment margin. The comprehensive load characteristic assessment report includes a visualized risk matrix and optimization strategies. The optimization strategies are associated with the time window markers in the event tag sequence.

[0020] Collecting historical operational datasets from the computing center is the starting point of the technical solution. This step achieves comprehensive coverage by integrating data from multiple sources, including task execution logs from the computing power task scheduling system, load demand records from power monitoring equipment, intermittent energy output data collected from new energy monitoring devices, and state-of-charge information recorded by the energy storage system controller. These datasets cover the core elements of the computing center's operation, providing raw input for subsequent analysis. The data acquisition process relies on standardized interface protocols to achieve seamless access from various data sources, avoiding analytical biases caused by format differences.

[0021] Standardizing historical datasets is crucial for improving data quality. First, data cleaning is performed, using a sliding window technique to detect and remove outliers, generating a dataset with outlier removal. Next, time alignment is applied to unify data from different frequencies to a fixed time granularity, forming a standardized dataset. Based on this, a rigid load feature subset and a flexible load feature subset are defined. The rigid load feature subset focuses on power consumption data for uninterrupted tasks, such as GPU cluster power; the flexible load feature subset covers energy consumption characteristics for adjustable tasks, such as storage device power consumption. This process addresses data heterogeneity and lays the foundation for feature analysis.

[0022] Analyzing load characteristic datasets to calculate dynamic coupling indices aims to identify critical events. A dynamic imbalance index is defined as a coupling quantification tool, combining the deviation between actual and predicted rigid load values ​​and the energy shortage rate. A moving time window scanning technique is employed to traverse the dataset. When a high-priority task start-up time overlaps with a period of intermittent energy output decline, the system automatically inserts the current window's dynamic imbalance index value into the event tag. The output is a sequence of coupled event tags with time window identifiers and imbalance characteristics. This step captures the spatiotemporal coupling effect of tasks and energy fluctuations, revealing the root causes of microscopic distortions in the load curve.

[0023] Constructing a dynamic model of load behavior is the core innovation of the technical solution. After loading a standardized dataset and its subsets, time series analysis algorithms are applied to fit the long-term baseline macroeconomic trend. Event identifiers and dynamic imbalance index values ​​are extracted from coupled event-labeled sequences, and these disturbance factors are injected into a two-layer probabilistic framework. The upper framework fixes the macroeconomic trend, while the lower framework incorporates intermittent energy random fluctuation scenarios and the probability distribution of task urgency. Through probabilistic fluctuation simulation processing, a probability envelope band for the load curve is generated. This model integrates macroeconomic trends and microeconomic disturbances, significantly improving the accuracy of load forecasting in dynamic environments.

[0024] The evaluation analyzes the probability envelope of the load curve to generate the final decision support output. Load curve characteristic indicators are extracted from the envelope, including load mutation intensity, fluctuation duration coefficient, and resilience margin. Simultaneously, a dynamic imbalance index is obtained from the coupled event marker sequence. Based on time window identifiers, characteristic indicators and index values ​​are matched to identify high-risk periods and generate a time period marker set. This time period marker set is mapped to a three-color risk matrix to visually display the risk level across time dimensions. Optimization strategies are associated with time window identifiers; for example, task migration windows suggest flexible load offsets, and energy storage power configuration thresholds set scheduling boundaries. This step forms a closed-loop decision mechanism, transforming the evaluation results into executable strategies.

[0025] Specifically, the method for evaluating the load characteristics of a computer center based on macroscopic quantitative analysis as described in this invention includes step 2 as follows: Perform sliding window outlier detection on the historical dataset collected in step 1 to obtain the dataset processed by outlier detection. Time alignment technology was applied to the outlier detection-processed dataset to unify different collection frequencies to a 5-minute granularity, generating a standardized dataset. The standardized dataset is divided into a rigid load feature subset and a flexible load feature subset. The rigid load feature subset includes GPU cluster power data, while the flexible load feature subset includes storage device power consumption data.

[0026] Outlier detection is the first step in data standardization. A sliding window mechanism is used to traverse the historical dataset and identify power outliers based on statistical distribution characteristics. When data points within the window exceed a preset threshold, the system automatically marks them as outliers and performs interpolation replacement. This operation eliminates noisy data caused by sensor malfunctions or transient interference, providing a clean data foundation for subsequent analysis. The processed dataset retains the original data distribution pattern while eliminating the interference of abnormal fluctuations on feature extraction.

[0027] Time alignment technology addresses the issue of frequency heterogeneity in multi-source data. A timestamp resampling mechanism is applied to the outlier-processed dataset, uniformly mapping the mixed second- and minute-level data to a fixed time axis. Linear interpolation is used to fill in missing time points, and high-frequency sampled data is aggregated through weighted averaging. The output is a standardized dataset with completely consistent time granularity, ensuring comparability of load records, energy output, and energy storage status from different sources across the same time dimension. This process achieves spatiotemporal synchronization of data across systems, supporting subsequent feature correlation analysis.

[0028] Load feature subset partitioning is based on differences in task schedulability. A rigid load feature subset is separated from a standardized dataset; its data source is the power consumption records of uninterruptible computing tasks, typically such as the power curves of a GPU cluster. Simultaneously, a flexible load feature subset is extracted; its data source is the energy consumption characteristics of adjustable tasks, such as power fluctuations corresponding to read / write operations on storage devices. The partitioning operation is achieved through task type label matching, allowing the two types of loads to be processed independently in subsequent analysis. This step establishes a mapping relationship between task attributes and load features, providing a classification basis for identifying coupled events.

[0029] Specifically, step 3 of the computer center load characteristic assessment method based on macroscopic quantitative analysis described in this invention includes: The dynamic imbalance index DI is defined based on the rigid load values ​​in the rigid load feature subset, where DI = (actual rigid load value - predicted value) × energy shortage rate; The load feature dataset is processed by scanning with a moving time window. When the start of a high-priority task overlaps with energy decay, the DI value calculated in the current time window is inserted into the event marker. Output a sequence of event markers with a DI index, where each event marker includes a time window identifier and a corresponding imbalance feature value.

[0030] The dynamic imbalance index focuses on the real-time matching status between rigid loads and energy supply. Actual power values ​​are extracted from a subset of rigid load characteristics, reflecting the real-time energy consumption demand of uninterrupted computing tasks. Simultaneously, power forecasts based on historical operating data are acquired, reflecting load trends. The energy shortage ratio is calculated by combining intermittent energy output monitoring data, quantifying the degree of energy supply gap. By fusing power deviation values ​​and the energy shortage ratio, a dynamic imbalance index is generated. This index quantifies the intensity of supply-demand imbalance at a specific point in time, providing a core measurement tool for identifying coupled events. This step establishes an objective evaluation benchmark, revealing the inherent conflict between rigid task demands and fluctuations in renewable energy sources.

[0031] A moving time window scanning mechanism enables spatiotemporal capture of critical events. A fixed-width time window traverses the load characteristic dataset, continuously sliding to cover different time periods at preset steps. Within the window, high-priority task start identifiers and intermittent energy output decay characteristic parameters in the task scheduling log are analyzed simultaneously. When an overlap is detected between a task start timestamp and an energy decay time interval, an event marker implantation operation is triggered, binding the dynamic imbalance index value calculated in the current window to a new event marker. This scanning process efficiently identifies microscale coupling effects and avoids feature omissions due to data dispersion.

[0032] The event-tagged sequence is constructed to form a structured feature library output. Each event tag encapsulates two key fields: a time window identifier that records the start timestamp and duration of the event, and an imbalance feature value field that stores the corresponding dynamic imbalance index value. The output event-tagged sequence with the dynamic imbalance index fully preserves the spatiotemporal distribution characteristics of the coupled events, and this sequence serves as the input for subsequent models as a quantified disturbance factor. This step completes the closed-loop processing from event detection to feature encapsulation, supporting the accurate simulation of the dynamic model of load behavior.

[0033] Specifically, the computer center load characteristic assessment method based on macroscopic quantitative analysis described in this invention assesses the probability envelope of the load curve using a dynamic load behavior model, including: Load the standardized dataset generated in step 2 and the rigid load feature subset and flexible load feature subset obtained by its partitioning, and apply the ARIMA algorithm to fit the macro trend baseline on the standardized dataset; Parse the event-tagged sequence with DI index output from step 3, and extract the event time window identifier and the corresponding DI value; The event time window identifier and DI value are injected into the two-layer probabilistic framework: the macro trend baseline is fixed in the upper framework, and the intermittent energy random fluctuation scenario and task urgency probability distribution are incorporated into the lower framework. The two-layer probabilistic framework is processed by Monte Carlo simulation to output the load curve probability envelope.

[0034] Data loading and macroeconomic trend fitting form the basic framework of the model. The standardized dataset generated in step 2, along with its rigid load characteristic subset and flexible load characteristic subset, are loaded. The ARIMA time series analysis algorithm is then applied to establish a long-term load change trend model. This algorithm identifies periodic patterns and trends based on historical load characteristic data, outputting a macroeconomic trend baseline as a steady-state reference for load behavior. This process isolates the impact of short-term disturbances and captures the essential laws governing the evolution of the computer center's load.

[0035] Event feature extraction enables the quantification and mapping of micro-level disturbances. The event-marked sequence with the DI index output from step 3 is analyzed, separating the time window identifier field and the corresponding dynamic imbalance index value from each event marker. The time window identifier pinpoints the specific time period in which the coupled event occurs, and the dynamic imbalance index value quantifies the intensity of the imbalance between task demand and energy supply within that time period. The extraction results form a mapping table between spatiotemporal coordinates and disturbance intensity, providing structured input for probabilistic simulations. A two-layer probabilistic framework integrates steady-state and disturbance characteristics to generate an envelope. The upper-layer framework fixes the macroeconomic trend baseline, maintaining the long-term patterns of load evolution. The lower-layer framework injects dynamic imbalance index values ​​corresponding to event time windows, while incorporating probability distributions of intermittent energy output random fluctuations and task urgency. Multiple rounds of random scenario calculations are performed using Monte Carlo simulation to comprehensively output a set of probability distributions for load fluctuation values. By extracting quantile boundaries from this set of probability distributions, a curved probability envelope representing the range of load fluctuations is generated.

[0036] Specifically, step 5 of the computer center load characteristic assessment method based on macroscopic quantitative analysis described in this invention includes: The load curve characteristic index is extracted from the probability envelope of the load curve, and the dynamic imbalance index is obtained from the event tag sequence with DI index output in step 3. Based on time window identification, load mutation intensity and dynamic imbalance index are matched to load curve characteristic indicators. High-risk periods that simultaneously meet the criteria of dynamic imbalance index DI > 0.8 and load mutation intensity > 500 kW / min are identified, and a period marker set is generated. A three-color risk matrix is ​​generated based on the mapping of time period marker sets to identify the risk level in the time dimension.

[0037] Feature index extraction enables structured analysis of assessment elements. Three characteristic indicators—load mutation intensity, fluctuation duration coefficient, and elasticity margin—are separated from the probability envelope of the load curve, with each indicator stored in time-series format. Simultaneously, historical dynamic imbalance index values ​​are obtained from event-tagged sequences with dynamic imbalance indices. This operation achieves complete separation of feature data and event data through a timestamp indexing mechanism, forming standardized feature index datasets and dynamic imbalance index datasets. This step constructs the foundational data pool for the assessment report.

[0038] A time window matching mechanism links multi-source feature data. Based on the time window identifier field in the event-marked sequence, the dynamic imbalance index dataset and the load curve characteristic index dataset are spatiotemporally aligned and matched. A bivariate correlation table is established by mapping the load mutation intensity value to the corresponding dynamic imbalance index value within the time window through the identifier field. The matching process employs a time interval overlap algorithm to accurately pinpoint the spatiotemporal coupling relationship between indicators and events. This step addresses the cross-source data fusion problem and provides a foundation for correlation analysis in risk identification.

[0039] High-risk periods are identified and visualized to generate decision-making support. The system scans the combined values ​​of the dynamic imbalance index and load mutation intensity in a bivariate correlation table. When both exceed a preset risk threshold, a period marker set is generated. Each period marker includes a start timestamp, duration, and risk level identifier. The period marker set is input into a matrix mapping engine, which expands it along the time dimension to generate a three-color risk matrix. This matrix uses color saturation to indicate risk level intensity, forming a risk heatmap on the time axis. This step completes the transformation from data quantitative analysis to visual decision support.

[0040] Specifically, the optimization strategy of the computer center load characteristic assessment method based on macroscopic quantitative analysis described in this invention includes: Recommendations for task migration windows and energy storage power configuration thresholds; The task migration window recommendation includes: associating time window identifiers in the event tag sequence with DI index to guide flexible load offset operations to avoid overlapping peaks; The energy storage power configuration thresholds include: setting power dispatch boundary values ​​for different risk stages based on the risk level of the three-color risk matrix.

[0041] The task migration window suggestion generation mechanism is closely linked to event characteristics. It analyzes the time window identifier field in the event-tagged sequence with the DI index, extracting the time period corresponding to the high dynamic imbalance index as the risk peak interval. It maps the task schedulable attribute identifiers in the flexible load feature subset, filtering computational task types that allow time offsets. A task migration window suggestion table is generated, including recommended offset time intervals and priority ranking parameters, guiding the operation and maintenance system to avoid energy shortages and high load overlap periods for flexible load operations such as data backup and non-real-time analysis. This strategy alleviates the system's supply and demand imbalance through spatiotemporal resource reallocation.

[0042] The energy storage power configuration threshold is set based on the risk visualization results. The risk level distribution in the three-color risk matrix is ​​analyzed, dividing the time axis into red, yellow, and green risk phase zones. For the high-risk red phase zone, a lower limit threshold for the forced discharge power of the energy storage system is set to ensure power supply stability for critical tasks; for the medium-risk yellow phase zone, an upper limit threshold for the charging power is set to prevent overcharging; for the low-risk green phase zone, the hard constraints are removed, and an adaptive adjustment mode is enabled. An energy storage power configuration threshold mapping table is output, establishing a dynamic correspondence between the time coordinates and power boundary values.

[0043] The two optimization strategies work together to form a synergistic mechanism through technical feature linkage. The task migration window suggestion table is input into the computing power scheduling system to dynamically adjust the execution time of deferred tasks. The energy storage power configuration threshold mapping table is input into the energy management system to constrain charging and discharging behavior boundaries in real time. During strategy execution, the time window identifiers of the event-marked sequence with DI index and the risk level parameters of the three-color risk matrix are shared, enabling dynamic matching between resource scheduling strategies and load characteristic assessment results. Ultimately, a closed-loop decision-making chain is constructed from risk identification and strategy generation to system execution, enhancing the operational resilience of the computing center in a dynamically fluctuating environment.

[0044] Specifically, the parameters for configuring the moving time window scanning in the computer center load characteristic assessment method based on macroscopic quantitative analysis described in this invention include: Set the scan step size to 1 minute and the time window width to 15 minutes, and detect high-frequency event mutation regions by continuously overlapping moving windows.

[0045] The time window width for moving time window scanning needs to cover the duration period of typical events. The time window width is determined based on the average duration characteristics of historical coupled events, ensuring the window can fully capture the complete evolution of events where high-priority task initiation and energy decay overlap. This width value is greater than the normal load fluctuation period, effectively distinguishing between sudden coupled events and normal load fluctuations. The time window width parameter directly affects the spatiotemporal accuracy of event marking, providing an accurate disturbance timescale for subsequent probabilistic simulations.

[0046] The scan step size parameter controls the granularity of time window movement. Setting a scan step size value smaller than the time window width ensures overlapping coverage areas between adjacent time windows. By traversing the load feature dataset through continuously overlapping moving windows, event detection omissions due to excessively large step sizes are avoided. The step size value is kept on the same order of magnitude as the data acquisition frequency to prevent the loss of high-frequency data features during scanning. The continuous overlap mechanism ensures the complete capture of microscale coupling effects.

[0047] The detection of high-frequency event abrupt change regions relies on a sliding window mechanism. Within each time window, priority identifiers and energy output attenuation characteristic parameters in the task scheduling log are analyzed synchronously. When both conditions are met, a potential event window is marked. The detection status is updated in real time as the window slides in steps, and the start and end points of load abrupt changes are identified by the rate of change of states in adjacent windows. This mechanism accurately locates regions of load curve distortion, supporting the spatiotemporal binding of the dynamic imbalance index.

[0048] Specifically, the computer center load characteristic assessment method based on macroscopic quantitative analysis described in this invention, with a two-layer probabilistic framework processed by Monte Carlo simulation, includes: Perform 500 random scenario calculations to obtain 500 sets of load fluctuation value output sequences; A probability density statistical algorithm was applied to the output sequence of 500 sets of load fluctuation values, and the 50th percentile value of the calculation result distribution was taken to generate the P50 prediction line. The standard deviation σ is calculated based on the output sequence of 500 sets of load fluctuation values, and the upper and lower boundaries of the power fluctuation range are generated based on the standard deviation σ value. The P50 prediction line is combined with the upper and lower boundaries of the power fluctuation range to form the probability envelope of the final output load curve.

[0049] A probability distribution set is constructed through multiple rounds of random scenario calculations. A predetermined number of random scenario calculations are performed, each injecting different intermittent energy output random fluctuation samples and task urgency probability distribution samples into the lower-level probabilistic framework. By simulating the combined effects of dynamic imbalance index disturbances and random factors, a load fluctuation value sequence under the corresponding scenario is output. The accumulated results from multiple rounds of calculations form a load fluctuation value output sequence set, which covers the load state space that the system may exhibit under dynamic fluctuation environments.

[0050] Probability density statistics are used to generate a baseline prediction line. A probability density statistical algorithm is applied to the set of load fluctuation value output sequences to aggregate the load value distribution characteristics of all sequences along the time axis. The load values ​​at the 50th percentile of the distribution set are extracted and connected in chronological order to form the P50 prediction line. This prediction line characterizes the most likely load evolution trajectory of the system in a dynamic fluctuation environment, serving as the core benchmark reference for the probability envelope.

[0051] Standard deviation calculation quantifies the range of load fluctuation boundaries. Based on the set of load fluctuation value output sequences at the same time point, the standard deviation of the load value at each time point is calculated. This metric reflects the degree of dispersion of load fluctuations from the baseline prediction line. Using the P50 prediction line as the baseline, a standard deviation of a set multiple is superimposed upwards to generate the upper boundary of power fluctuation, and a standard deviation of a set multiple is subtracted downwards to generate the lower boundary of power fluctuation. The upper and lower boundaries together define the reasonable range of load fluctuation variation.

[0052] The probability envelope integrates forecasting and fluctuation range. The P50 forecast line and the upper and lower boundaries of power fluctuations are aligned and integrated along the time axis to form a three-line combination load curve probability envelope. This envelope includes both the baseline forecast trajectory and the fluctuation range boundary, visually displaying the probability space of load evolution at different confidence levels. The final output is a strip-shaped region graph in a time-load coordinate system, supporting the quantitative assessment of operational risks in data centers.

[0053] Specifically, the computer center load characteristic assessment method based on macroscopic quantitative analysis described in this invention further includes: The task time constraint data of the flexible load feature subset obtained in step 2 is analyzed, and the flexible task time window parameter representing the task delayability is extracted. Based on the flexible task time window parameter, the task migration offset time range is set to a range of 0.5-1.5 hours. The time range value is input into the task migration window suggested operation to control the flexible load offset time margin.

[0054] Task timeliness constraint data parsing focuses on the schedulable characteristics of flexible loads. The task timeliness constraint field is extracted from the flexible load feature subset obtained in step 2. This field records the latest completion time requirement for delayable computation tasks. By calculating the difference between the task submission timestamp and the latest completion timestamp, the maximum allowable delay duration of the task is quantified. The data parsing process associates unique task identifiers with the data, establishing a mapping table between task instances and timeliness constraints, providing structured input for time margin control.

[0055] The flexible task timeliness window parameter defines a quantitative benchmark for task delayability. Based on the timeliness constraint mapping table, the time interval from task submission time to the latest completion time is extracted as the basic window parameter. By introducing a task priority weight coefficient to correct the basic window value, a timeliness window parameter characterizing task delayability is generated. The value of this parameter is positively correlated with the allowable delay time of the task; the larger the value, the higher the task scheduling flexibility, providing a benchmark for elasticity measurement of offset operations.

[0056] The system sets the migration offset duration range and implements closed-loop control for strategy execution. A finite offset interval is set based on the flexible task time window parameter value to constrain the allowable time margin of the task migration operation. The offset interval value is input into the task migration window's suggested operation module and dynamically matched with the time window identifier in the event tag sequence with the DI index. When the suggested offset period exceeds the set interval, the system automatically truncates to the nearest valid boundary point, achieving closed-loop execution of the flexible load offset operation within a controllable time margin.

[0057] Specifically, the computer center load characteristic assessment method based on macroscopic quantitative analysis described in this invention further includes: parsing the task time constraint data of the flexible load characteristic subset obtained in step 2, extracting the flexible task time window parameter characterizing the task's delayability, setting the task migration offset time range based on the flexible task time window parameter, limiting it to a range of 0.5-1.5 hours, inputting the time range value into the task migration window suggested operation, and controlling the flexible load offset time margin.

[0058] Task timeliness constraint data parsing focuses on the schedulable characteristics of flexible loads. The task timeliness constraint field is extracted from the flexible load feature subset obtained in step 2. This field records the latest completion time requirement for delayable computation tasks. By calculating the difference between the task submission timestamp and the latest completion timestamp, the maximum allowable delay duration of the task is quantified. The data parsing process associates unique task identifiers with the data, establishing a mapping table between task instances and timeliness constraints, providing structured input for time margin control.

[0059] The flexible task timeliness window parameter defines a quantitative benchmark for task delayability. Based on the timeliness constraint mapping table, the time interval from task submission time to the latest completion time is extracted as the basic window parameter. By introducing a task priority weight coefficient to correct the basic window value, a timeliness window parameter characterizing task delayability is generated. The value of this parameter is positively correlated with the allowable delay time of the task; the larger the value, the higher the task scheduling flexibility, providing a benchmark for elasticity measurement of offset operations.

[0060] The system sets the migration offset duration range and implements closed-loop control for strategy execution. A finite offset interval is set based on the flexible task time window parameter value to constrain the allowable time margin of the task migration operation. The offset interval value is input into the task migration window's suggested operation module and dynamically matched with the time window identifier in the event tag sequence with the DI index. When the suggested offset period exceeds the set interval, the system automatically truncates to the nearest valid boundary point, achieving closed-loop execution of the flexible load offset operation within a controllable time margin.

[0061] This invention addresses the problem of load assessment distortion caused by the superposition of computing power tasks and energy fluctuations by establishing a dynamic coupling effect identification mechanism, a probabilistic load prediction model, and a closed-loop strategy execution system. The specific technical logic is as follows: A precise quantification mechanism for dynamic coupling effects is employed: A moving time window scanning technique is used to detect the spatiotemporal overlap of high-priority task initiation and intermittent energy decay events in real time. When both conditions are met, a dynamic imbalance index is generated based on the actual demand of rigid loads and the energy supply gap, and an event-marked sequence with time window identifiers is embedded. This sequence quantifies the coupling strength between task scheduling and energy fluctuations, capturing microscopic distortion features of the load curve that existing methods cannot identify. For example, the event-marked sequence can locate regions of sudden load changes caused by a sudden drop in photovoltaic output when the GPU cluster is running at full power.

[0062] When constructing a dynamic load behavior model, the upper-level framework fixes the macroeconomic trend baseline, while the lower-level framework injects disturbance factors based on event-marked sequences and random energy fluctuation scenarios. A probability envelope band for the load curve is generated through Monte Carlo simulation, reflecting both steady-state evolution and random fluctuation boundaries. For example, in a scenario of emergency task startup superimposed with wind power output fluctuations, the envelope band can display the probability distribution range of load abrupt changes, significantly improving prediction accuracy under dynamic conditions.

[0063] The probability envelope is analyzed to extract characteristic indicators such as the intensity of load mutations, and a dynamic imbalance index is associated with the event-marked sequence. High-risk periods are identified based on spatiotemporal matching results, generating a three-color risk matrix. Optimization strategies are directly linked to risk levels and time markers: task migration windows are suggested to associate event sequence time markers to guide flexible loads such as storage backups to avoid peak periods; energy storage power configuration thresholds are set to set charging and discharging boundaries based on the risk matrix level. For example, during red-risk periods, energy storage is forced to discharge to maintain GPU cluster operation, while charging power is limited during yellow-risk periods.

[0064] The specific implementation of this invention is based on a typical operating scenario of a smart computing center in the Hebei North region, and solves the problem of inaccurate load assessment through the following technical solutions: Collect computing task scheduling logs, load demand records, intermittent energy output monitoring data, and energy storage system status information. Perform sliding window outlier detection on multi-source heterogeneous data to eliminate sensor noise interference; apply time alignment technology to unify to a fixed time granularity and generate a standardized dataset. Divide rigid load characteristic subsets (such as GPU cluster power) and flexible load characteristic subsets (such as storage device power consumption) and establish a mapping relationship between task attributes and energy consumption characteristics.

[0065] A dynamic imbalance index is defined based on a subset of rigid load characteristics to quantify the real-time deviation between task demand and energy supply. A moving time window is used to scan the standardized dataset. When a high-priority task initiation period (e.g., the 15:00 real-time inference task) overlaps with an energy decay period (e.g., a sudden drop in photovoltaic output), the dynamic imbalance index is embedded into the event marker. The output is a coupled event marker sequence with time window identifiers, accurately capturing the load surge event described in the background technology (a sudden drop from 9.2MW to 4.1MW at 15:30).

[0066] A standardized dataset is loaded, and the ARIMA algorithm is applied to fit a macroeconomic trend baseline. Event-marked sequences are analyzed to extract spatiotemporal identifiers and dynamic imbalance indices, which are then incorporated into a two-layer probabilistic framework: the upper layer fixes the macroeconomic trend, while the lower layer integrates random fluctuations in new energy sources (such as the probability distribution of wind power output) and task urgency parameters. Monte Carlo simulation outputs the probability envelope of the load curve, covering fluctuations in the background technology that exceed historical statistics (a 7.5MW jump).

[0067] Indicators such as load mutation intensity are extracted from the probability envelope and associated with the dynamic imbalance index of the event-labeled sequence. High-risk periods are identified based on spatiotemporal matching (e.g., DI > threshold and mutation intensity > threshold), and a three-color risk matrix is ​​generated. Optimization strategies are bound to time window identifiers: task migration window suggestions guide flexible loads (e.g., data analysis tasks) to avoid overlapping peaks; energy storage power configuration thresholds set charge and discharge boundaries based on risk levels (e.g., forced discharge during red-risk periods).

[0068] In the implementation case at the Hebei North Intelligent Computing Center, the event tagging sequence successfully captured the 15:00 task-energy decay coupling event; the probability envelope accurately represented the load fluctuation range (measured values ​​of 4.1-7.5MW were within the envelope boundary); the optimization strategy drove the energy storage system to discharge during high-risk periods to support the GPU cluster, while flexible loads were delayed to low-risk periods. The load curve fluctuation amplitude was reduced, verifying that the technical solution effectively solved the problems of inaccurate assessment and scheduling failure described in the background technology.

Claims

1. A method for evaluating the load characteristics of computer centers based on macroscopic quantitative analysis, characterized in that, include: Step 1: Collect historical operation datasets from the computer center. The historical operation datasets include computing task scheduling logs, load demand records, intermittent energy output monitoring data, and energy storage system status information. Step 2: Standardize the historical operation dataset collected in Step 1 to generate a load feature dataset with a uniform time granularity. The load feature dataset distinguishes between rigid load feature subsets and flexible load feature subsets. Step 3: Analyze the load characteristic dataset to calculate the dynamic coupling index. When the start time of a high-priority task overlaps with the period of intermittent energy output decay, generate a coupling event label sequence that includes time window identifiers and imbalance characteristics. Step 4: Load the dynamic coupling degree index and coupling event label sequence generated in Step 3, construct a dynamic model of load behavior, fit the long-term baseline macro trend of the load characteristic dataset, add the imbalance feature perturbation of the coupling event label sequence to the macro trend, and output the probability envelope of the load curve through probability fluctuation simulation. Step 5: The load behavior dynamic model is used to analyze the probability envelope of the load curve, extract the load curve characteristic indicators and generate a comprehensive load characteristic assessment report. The load curve characteristic indicators include load mutation intensity, fluctuation duration coefficient and elastic adjustment margin. The comprehensive load characteristic assessment report includes a visualized risk matrix and optimization strategies. The optimization strategies are associated with the time window markers in the event tag sequence.

2. The method for evaluating the load characteristics of a computer center based on macroscopic quantitative analysis as described in claim 1, characterized in that, Step 2 includes: Perform sliding window outlier detection on the historical dataset collected in step 1 to obtain the dataset processed by outlier detection. Time alignment technology was applied to the outlier detection-processed dataset to unify different collection frequencies to a 5-minute granularity, generating a standardized dataset. The standardized dataset is divided into a rigid load feature subset and a flexible load feature subset. The rigid load feature subset includes GPU cluster power data, while the flexible load feature subset includes storage device power consumption data.

3. The method for evaluating the load characteristics of a computer center based on macroscopic quantitative analysis as described in claim 2, characterized in that, Step 3 includes: The dynamic imbalance index DI is defined based on the rigid load values ​​in the rigid load feature subset, where DI = (actual rigid load value - predicted value) × energy shortage rate; The load feature dataset is processed by scanning with a moving time window. When the start of a high-priority task overlaps with energy decay, the DI value calculated in the current time window is inserted into the event marker. Output a sequence of event markers with a DI index, where each event marker includes a time window identifier and a corresponding imbalance feature value.

4. The method for evaluating the load characteristics of a computer center based on macroscopic quantitative analysis as described in claim 3, characterized in that, The assessment uses a dynamic load behavior model to analyze the probability envelope of the load curve, which includes: Load the standardized dataset generated in step 2 and the rigid load feature subset and flexible load feature subset obtained by its partitioning, and apply the ARIMA algorithm to fit the macro trend baseline on the standardized dataset; Parse the event-tagged sequence with DI index output from step 3, and extract the event time window identifier and the corresponding DI value; The event time window identifier and DI value are injected into the two-layer probabilistic framework: the macro trend baseline is fixed in the upper framework, and the intermittent energy random fluctuation scenario and task urgency probability distribution are incorporated into the lower framework. The two-layer probabilistic framework is processed by Monte Carlo simulation to output the load curve probability envelope.

5. The method for evaluating the load characteristics of a computer center based on macroscopic quantitative analysis as described in claim 4, characterized in that, Step 5 includes: The load curve characteristic index is extracted from the probability envelope of the load curve, and the dynamic imbalance index is obtained from the event tag sequence with DI index output in step 3. Based on time window identification, load mutation intensity and dynamic imbalance index are matched to load curve characteristic indicators. High-risk periods that simultaneously meet the criteria of dynamic imbalance index DI > 0.8 and load mutation intensity > 500 kW / min are identified, and a period marker set is generated. A three-color risk matrix is ​​generated based on the mapping of time period marker sets to identify the risk level in the time dimension.

6. The method for evaluating the load characteristics of a computer center based on macroscopic quantitative analysis as described in claim 5, characterized in that, Optimization strategies include: Recommendations for task migration windows and energy storage power configuration thresholds; The task migration window recommendation includes: associating time window identifiers in the event tag sequence with DI index to guide flexible load offset operations to avoid overlapping peaks; The energy storage power configuration thresholds include: setting power dispatch boundary values ​​for different risk stages based on the risk level of the three-color risk matrix.

7. The method for evaluating the load characteristics of a computer center based on macroscopic quantitative analysis as described in claim 6, characterized in that, The parameters for configuring moving time window scanning include: Set the scan step size to 1 minute and the time window width to 15 minutes, and detect high-frequency event mutation regions by continuously overlapping moving windows.

8. The method for evaluating the load characteristics of a computer center based on macroscopic quantitative analysis as described in claim 7, characterized in that, The two-layer probabilistic framework processed by Monte Carlo simulation includes: Perform 500 random scenario calculations to obtain 500 sets of load fluctuation value output sequences; A probability density statistical algorithm was applied to the output sequence of 500 sets of load fluctuation values, and the 50th percentile value of the calculation result distribution was taken to generate the P50 prediction line. The standard deviation σ is calculated based on the output sequence of 500 sets of load fluctuation values, and the upper and lower boundaries of the power fluctuation range are generated based on the standard deviation σ value. The P50 prediction line is combined with the upper and lower boundaries of the power fluctuation range to form the probability envelope of the final output load curve.

9. The method for evaluating the load characteristics of a computer center based on macroscopic quantitative analysis as described in claim 8, characterized in that, Also includes: The task time constraint data of the flexible load feature subset obtained in step 2 is analyzed, and the flexible task time window parameter representing the task delayability is extracted. Based on the flexible task time window parameter, the task migration offset time range is set to a range of 0.5-1.5 hours. The time range value is input into the task migration window suggested operation to control the flexible load offset time margin.

10. The method for evaluating the load characteristics of a computer center based on macroscopic quantitative analysis as described in claim 9, characterized in that, Also includes: The task time constraint data of the flexible load feature subset obtained in step 2 is analyzed, and the flexible task time window parameter representing the task delayability is extracted. Based on the flexible task time window parameter, the task migration offset time range is set to a range of 0.5-1.5 hours. The time range value is input into the task migration window suggested operation to control the flexible load offset time margin.

Images