A multi-time scale computing power load management method and system

Abstract

The application relates to the technical field of power systems, and discloses a multi-time-scale computing power load management method and system, wherein the method comprises the following steps: obtaining historical operation data of various types of data centers; based on the historical operation data, the computing power load characteristics of each type of data center under a multi-time scale and corresponding processing adaptation rules are analyzed in sequence; in response to a load optimization signal, real-time operation data of a target data center obtained is integrated and processed based on the computing power load characteristics and the corresponding processing adaptation rules, so that multi-source real-time data under a multi-time scale are obtained; the multi-source real-time data are input into a computing power load dynamic prediction model, so that computing power load prediction results of the target data center under different time scales are obtained; and based on the computing power load prediction results, time-adjustable tasks of the target data center are migrated and scheduled, so that load management optimization of the target data center is realized. The method improves the utilization efficiency of the computing power load.

Description

Technical Field

[0001] This invention relates to the field of data center energy efficiency management technology, and in particular to a multi-timescale computing load management method and system. Background Technology

[0002] With the rapid development of the digital economy, data centers, as the core of computing infrastructure, have seen continuous growth in scale and number, leading to increasingly prominent high energy consumption issues.

[0003] Existing data center multi-timescale computing load management relies on experience-based scheduling or single-timescale monitoring, which is difficult to cope with complex and ever-changing task loads and power supply fluctuations. Especially with the increasing penetration of new energy sources, existing methods often ignore the differences in load characteristics of different types of data centers and the load fluctuation patterns under multiple timescales, resulting in large load peak-valley differences, reduced energy utilization efficiency, and difficulty in meeting the development needs of green and efficient data centers. Summary of the Invention

[0004] This invention provides a multi-timescale computing load management method and system to solve the technical problem of how to improve existing multi-timescale computing load management methods and achieve the effect of improving the energy utilization efficiency of data centers.

[0005] To address the aforementioned technical problems, this invention provides a multi-timescale computing load management method, comprising: Acquire historical operational data for various types of data centers, including server performance metrics, power consumption data, and task processing volume over a long period of time; Based on the historical operational data, we sequentially analyze the computing load characteristics and corresponding processing adaptation rules of each type of data center at multiple time scales. In response to the load optimization signal, based on the computing load characteristics and the processing adaptation rules, the real-time server performance indicators, real-time power consumption data and real-time task processing volume of the target data center are integrated and processed to obtain multi-source real-time data at multiple time scales. The multi-source real-time data are respectively input into the pre-trained dynamic prediction model of computing load to obtain the computing load prediction results of the target data center at different time scales; wherein, the dynamic prediction model of computing load is configured as a multi-branch network structure to dynamically adjust the time granularity of the input data to adapt to the load patterns at different time scales. Based on the computing load prediction results, time-adjustable tasks in the target data center are migrated and scheduled across scales to optimize the load management of the target data center.

[0006] As one preferred option, the data center includes a general-purpose data center, an intelligent computing data center, and a supercomputing data center; The analysis of computing load characteristics and corresponding processing adaptation rules for various types of data centers across multiple time scales based on the historical operational data includes: The historical operational data is preprocessed to obtain a standardized historical load dataset; Based on the classification of data centers, the server performance indicators, power consumption data and task processing volume in the historical load dataset are classified and statistically analyzed to obtain the basic load characteristic data of each type of data center. The load change data at multiple time scales in the basic load characteristic data are analyzed in layers to extract the load peak and valley periods, fluctuation amplitude and periodicity characteristics at each time scale, so as to obtain the computing load characteristics of each type of data center at multiple time scales, and generate corresponding processing adaptation rules based on the computing load characteristics of each type of data center.

[0007] As a preferred embodiment, based on the computing load characteristics and the processing adaptation rules, the obtained real-time server performance indicators, real-time power consumption data, and real-time task processing volume of the target data center are integrated and processed to obtain multi-source real-time data at multiple time scales, including: Acquire real-time operational data of the target data center, including real-time server performance metrics, real-time power consumption data, and real-time task processing volume; Based on the processing adaptation rules of the target data center, outlier detection and filtering are performed on the real-time server performance indicators, the real-time power consumption data, and the real-time task processing volume to obtain real-time filtered data. Based on the computing load characteristics of the target data center, the real-time filtered data is sequentially aligned with time granularity and dynamically corrected to obtain corrected real-time sub-item data; wherein, the dynamic correction is designed to match the changing trend of the real-time filtered data with the historical load characteristics. The real-time segmented data are correlated and integrated according to a preset format to obtain standardized multi-source real-time data at multiple time scales.

[0008] As one preferred embodiment, the step of inputting the multi-source real-time data into a pre-trained dynamic prediction model for computing load to obtain the computing load prediction results for the target data center at different time scales includes: Feature extraction is performed on the multi-source real-time data, and model input feature sets are constructed at each time scale based on the feature extraction results; wherein, the model input feature sets include real-time performance features, energy consumption features, and task load features related to computing power load; The model input feature sets at each time scale are respectively input into the dynamic prediction model of computing load to perform multi-time scale load prediction, so as to obtain the computing load prediction results of the target data center at different time scales.

[0009] As one preferred embodiment, the cross-scale migration scheduling of time-adjustable tasks in the target data center based on the computing load prediction results includes: Based on the computing load prediction results, analyze the time migration characteristics of computing tasks in the target data center and identify time-adjustable tasks in the target data center computing tasks. Based on the characteristics of the time-adjustable tasks and the green operation requirements of the target data center, a cross-time-scale scheduling model for computing tasks adapted to new energy output is constructed. Based on the aforementioned scheduling model, time-adjustable tasks are flexibly migrated and scheduled at different time scales to match the computing load prediction results with the fluctuations in the new energy power generation curve.

[0010] Another aspect of the present invention provides a multi-time-scale computing load management system, comprising: The acquisition module is used to acquire historical operating data of various types of data centers, including server performance indicators, power consumption data, and task processing volume over a long period of time. The analysis module is used to sequentially analyze the computing load characteristics of various types of data centers at multiple time scales based on the historical operating data. The integration module is used to respond to the load optimization signal and integrate the real-time server performance indicators, real-time power consumption data and real-time task processing volume of the target data center based on the computing power load characteristics to obtain multi-source real-time data. The prediction module is used to input the multi-source real-time data into a pre-trained dynamic prediction model of computing load to obtain the prediction results of computing load of the target data center at different time scales. The scheduling module is used to migrate and schedule time-adjustable tasks in the target data center based on the computing load prediction results, so as to optimize the load management of the target data center.

[0011] As one preferred embodiment, the analysis module is specifically used for: The historical operational data is preprocessed to obtain a standardized historical load dataset; The server performance indicators, power consumption data, and task processing volume in the historical load dataset are classified and statistically analyzed to obtain the basic load characteristic data of each type of data center. The load change data at multiple time scales in the basic load characteristic data are analyzed hierarchically to extract the load peak and valley periods, fluctuation amplitude and periodicity characteristics at each time scale, so as to obtain the computing load characteristics of each type of data center at multiple time scales.

[0012] As one preferred embodiment, the integration module is specifically used for: Acquire real-time operational data of the target data center, including real-time server performance metrics, real-time power consumption data, and real-time task processing volume; Anomaly detection and filtering are performed on the real-time server performance indicators, the real-time power consumption data, and the real-time task processing volume to obtain real-time filtered data. Based on the computing load characteristics, the real-time filtered data is sequentially aligned with time granularity and dynamically corrected to match the changing trend of the real-time filtered data with the historical load characteristics, thereby obtaining corrected real-time sub-item data. The real-time sub-data is linked and integrated according to a preset format to obtain standardized multi-source real-time data.

[0013] As one preferred embodiment, the prediction module is specifically used for: Feature extraction is performed on the multi-source real-time data to extract real-time performance features, energy consumption features, and task load features related to computing power load, in order to construct a model input feature set; The model input feature set is input into the computing load dynamic prediction model to perform load prediction at multiple time scales, and the computing load prediction results of the target data center at different time scales are obtained. The dynamic prediction model for the central computing load is trained based on historical load characteristic data and machine learning algorithms, and has the ability to adapt to predictions at multiple time scales.

[0014] As one preferred embodiment, the scheduling module is specifically used for: Based on the computing load prediction results, analyze the time migration characteristics of computing tasks in the target data center and identify time-adjustable tasks in the target data center computing tasks. Based on the characteristics of the time-adjustable tasks and the green operation requirements of the target data center, the task completion rate and green electricity procurement cost are used as multi-dimensional objective optimization constraints to construct a cross-time scale scheduling model for computing tasks that adapt to the output of new energy sources. Based on the aforementioned scheduling model, time-adjustable tasks are flexibly migrated and scheduled at different time scales to match the computing load prediction results with the fluctuations in the new energy power generation curve.

[0015] Compared with the prior art, the beneficial effects of the present invention are at least one of the following: 1) The multi-timescale computing load management method of the present invention analyzes historical operating data at multiple time scales to accurately capture the computing load characteristics of different types of data centers. Combined with the dynamic integration and prediction model of real-time multi-source data, it significantly improves the accuracy and timeliness of computing load prediction and provides a scientific basis for load optimization scheduling.

[0016] 2) Based on the prediction results, this invention performs cross-scale migration scheduling for time-adjustable tasks, which can effectively balance the peak and valley of computing load in data centers, increase the proportion of new energy consumption, reduce the cost of green electricity procurement while ensuring task completion rate, realize the synergistic optimization of computing resources and energy supply, and provide strong support for the green and efficient operation of data centers. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a multi-timescale computing load management method in one embodiment of the present invention. Figure 2 This is a schematic diagram of a multi-timescale computing load management system in one embodiment of the present invention; Figure label: Among them, 11. Acquisition module; 12. Analysis module; 13. Integration module; 14. Prediction module; 15. Scheduling module. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] In the description of this application, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first," "second," "third," etc., may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0020] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal communication between two components. The terms "vertical," "horizontal," "left," "right," "upper," "lower," and similar expressions used herein are for illustrative purposes only and do not indicate or imply that the device or component referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as limiting the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0021] In the description of this application, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is for the purpose of describing specific embodiments only and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0022] One embodiment of the present invention provides a multi-time-scale computing load management method. For details, please refer to [link / reference]. Figure 1 , Figure 1 The diagram shown illustrates a multi-timescale computing load management method according to one embodiment of the present invention, which includes steps S1-S5: S1: Obtain historical operating data for various types of data centers, including server performance metrics, power consumption data, and task processing volume over a long period of time; It is understandable that acquiring historical operational data from various types of data centers is a fundamental step in supporting subsequent load characteristic analysis, predictive model construction, and scheduling optimization. Specifically, this embodiment of the invention focuses on collecting historical data for three typical data center types: general-purpose data centers, intelligent computing data centers, and supercomputing data centers. General-purpose data centers primarily handle routine IT services such as enterprise office systems and website hosting, with stable computing loads (CPU utilization 30%-60%). Server and cooling system energy consumption account for approximately 50% and 30% of total energy consumption, respectively, and load fluctuations are affected by weekdays and holidays. Intelligent computing data centers are centered on AI model training and inference, requiring GPUs or dedicated AI chips. Computing demand is sudden (GPU utilization 20%-95%), with GPU energy consumption accounting for 40%-60% of total energy consumption. Energy consumption fluctuations are strongly correlated with the start and stop of AI tasks. Supercomputing data centers primarily handle high-density scientific computing (CPU utilization 80%-95%, high load duration ≥8 hours / day), exhibiting pulsed high-load characteristics, with an energy density ≥10kW / ㎡, requiring dedicated liquid cooling systems.

[0023] Different types of data centers have significantly different business scenarios, computing power requirements, and energy consumption characteristics. For example, supercomputing data centers mostly undertake high-density computing tasks, and their load fluctuation patterns are quite different from those of general data centers that mainly handle routine business. Collecting data in categories can ensure the relevance of subsequent analysis.

[0024] The long-term span referred to in this invention is specifically at least 12 months, covering the complete annual cycle (including four seasons, weekdays, weekends, and public holidays) to capture intraday (e.g., peak business hours in general data centers from 9:00-18:00 and off-peak hours from 0:00-6:00), intraweekly (load difference between weekdays and weekends is 15%-20%), and seasonal (summer load is 10%-25% higher than winter due to cooling demand), providing sufficient samples for subsequent multi-timescale analysis. Historical operating data must cover three key types of information across the long-term span: Server performance metrics include core parameters that reflect the computing power operation status, such as server CPU utilization, memory usage, disk I / O speed, and network bandwidth utilization, which can accurately characterize the computing power load intensity of the data center. Electricity consumption data: covering total electricity consumption, electricity consumption at different times, and electricity consumption in different areas (such as computer rooms and cooling systems), which is directly related to the energy consumption characteristics of data centers and provides a basis for green electricity consumption analysis; Task processing volume: This includes the number of tasks submitted per unit of time, task types (such as real-time tasks and batch processing tasks), and task completion time. It reflects the changing patterns of the data center's business load and is a key basis for identifying time-adjustable tasks.

[0025] Long-term data collection can capture load fluctuation patterns across different time dimensions, such as intraday peak-valley variations and seasonal load differences. This type of data provides sufficient samples for analyzing the periodic characteristics, trend changes, and random fluctuations of the load, serving as crucial support for subsequently constructing multi-timescale forecasting models.

[0026] S2: Based on the historical operating data, analyze the computing load characteristics and corresponding processing adaptation rules of each type of data center at multiple time scales.

[0027] Preferably, in one embodiment of the present invention, the historical operating data acquired in S1 is first preprocessed to obtain a standardized historical load dataset. Preprocessing refers to the operation of eliminating data interference and unifying data format through a fixed process, specifically including outlier detection and filtering, missing value filling and data standardization. Preprocessing can eliminate interference such as data acquisition errors and equipment model differences, laying a reliable data foundation for feature analysis.

[0028] The standardized historical load dataset is categorized by data center type (general data center, intelligent computing data center, supercomputing data center). Then, for each data center type, three core data categories are analyzed to obtain basic load characteristic data. The categorization statistics involve quantitative analysis based on both data dimensions and data center type, specifically including: Server performance metric statistics: By calculating the mean, peak, and duration of parameters such as CPU utilization and memory usage, the intensity characteristics of computing load are characterized. Electricity consumption data statistics: Statistics on electricity consumption, energy density, and the proportion of energy consumption of the cooling system in different time periods, reflecting the correlation between energy consumption and computing load; Task processing volume statistics: Classify and statistically analyze task throughput and processing time distribution by task type and priority to identify the time distribution characteristics of business load.

[0029] By classifying and statistically analyzing data, we can obtain basic load characteristic data for various types of data centers, clarify the differences in computing power requirements among different data centers, such as the duration of high load in supercomputing data centers and the burstiness of tasks in intelligent computing data centers.

[0030] Furthermore, based on the basic load characteristic data, the focus is on performing hierarchical decomposition analysis on load change data across multiple time scales. Specifically, multiple time scales refer to four dimensions: hourly (unit of hour), daily (unit of day), monthly (unit of month), and grade-level (unit of year). Hierarchical analysis involves decomposing the data according to different time dimensions and extracting key features. Three core types of features are extracted, including: Peak and off-peak period characteristics: Identify peak and off-peak periods of load at different time scales, such as intraday peak electricity consumption periods (e.g., 9:00-18:00) and seasonal peak months (e.g., increased load in summer due to increased cooling demand). Fluctuation amplitude characteristics: Quantify the degree of load fluctuation in different time periods, calculate indicators such as peak-to-valley difference and volatility, and distinguish the difference between the stable load of general data centers and the pulsed load of supercomputing data centers; Periodic characteristics: Explore the periodic patterns of load changes, including intraday periodicity (such as the load difference between weekdays and nighttimes), weekly periodicity (such as the load fluctuation between weekdays and weekends), and seasonal periodicity, to provide a basis for subsequent prediction models to adapt to multiple time scales.

[0031] Based on the aforementioned computing load characteristics, processing adaptation rules are generated for each type of data center. These rules are type-specific standards developed for subsequent S3 real-time data integration, ensuring that real-time data processing aligns with the load patterns of different data centers. Specific rules include: time granularity adaptation rules, where general data centers, due to stable loads, align real-time data to a 5-minute granularity; intelligent computing data centers, due to sudden task bursts, align to a 1-minute granularity; and supercomputing data centers, due to high load spikes, align key indicators to a 10-second granularity; outlier threshold adaptation rules, where the CPU utilization outlier threshold for general data centers is set to >90%; for intelligent computing data centers, due to large GPU load fluctuations, the outlier threshold is dynamically adjusted to "average ±25%"; and for supercomputing data centers, due to stable loads, the outlier threshold is set to >95% or <75%; and dynamic correction adaptation rules, where general data centers correct real-time data periodically within a day, intelligent computing data centers correct based on the characteristics of sudden task bursts, and supercomputing data centers correct based on the continuous pattern of pulsed loads. These rules provide clear operational guidelines for S3 real-time data integration, avoiding data deviations caused by a one-size-fits-all approach.

[0032] It should be noted that all operations in this step must be based on the three types of data centers (general, intelligent, and supercomputing) defined in S1 to conduct type differentiation analysis. For example, the processing adaptation rules for supercomputing data centers need to be adapted to the characteristics of pulsed high load, intelligent computing data centers need to be adapted to the characteristics of task bursts, and general data centers need to be adapted to the characteristics of stable periodicity. The final output of computing load characteristics + processing adaptation rules needs to be synchronously stored in a structured database (such as MySQL). The fields include data center type, time scale, feature name, feature value, and adaptation rule content to ensure that subsequent steps can directly call them.

[0033] S3: In response to the load optimization signal, based on the computing power load characteristics and the processing adaptation rules, the real-time server performance indicators, real-time power consumption data and real-time task processing volume of the target data center are integrated and processed to obtain multi-source real-time data under multiple time scales.

[0034] It should be noted that the load optimization signal is a digital instruction to initiate the data integration process in this step. The triggering condition must meet any of the following: real-time computing load ≥ 80% of the target data center server's rated load (overload warning); real-time green electricity output ≥ 60% of the target data center's total power demand (green electricity consumption optimization opportunity); single-period power consumption cost ≥ 120% of the data center's historical average for the same period (cost control requirement). The signal format is JSON, containing the fields trigger_type (trigger type), timestamp (trigger time accurate to the second), and target_center_type (target data center type) to ensure clear and identifiable triggering logic. Simultaneously, real-time operational data of the target data center is acquired, which is consistent with the dimensions of S1 historical data, specifically including: Real-time server performance metrics include parameters that reflect the real-time load of computing power, such as current CPU utilization, memory usage, and network throughput, corresponding to the technical requirements for real-time monitoring of distributed resource characteristics. Real-time electricity consumption data: This includes energy consumption data such as current total electricity consumption, real-time energy consumption in various regions, and real-time green electricity consumption, providing a basis for subsequent matching with new energy output; Real-time task processing volume: including the current task queue length, the type and priority of the tasks being executed, the task completion progress, etc., supporting the dynamic identification of time-adjustable tasks.

[0035] Real-time data acquisition must meet the requirements of high frequency and continuity to ensure the capture of instantaneous changes in computing load.

[0036] It should be noted that the core purpose of the real-time data acquired in this step is to serve real-time optimization and scheduling. When responding to load optimization signals, the integrated processing of real-time data provides "real-time input" to the dynamic prediction model of computing load, supports the generation of load prediction results at different time scales, and ultimately provides an immediate decision-making basis for the migration and scheduling of time-adjustable tasks.

[0037] The core purpose of the historical operational data in step S1 is to support the analysis of computing load characteristics and the training of prediction models. Through long-term data accumulation, it provides raw materials for analyzing the basic load characteristics of different types of data centers (such as peak and valley periods and periodic patterns), and serves as "training samples" for constructing historical load characteristic maps and training dynamic prediction models.

[0038] Furthermore, to eliminate interference factors in real-time data, outlier processing needs to be performed on the three types of collected data: Statistical analysis and machine learning algorithms are used to identify data points that deviate from the normal range, such as sudden increases or decreases in server performance indicators and jumps in power consumption data. Filter or interpolate outliers to prevent data distortion caused by equipment failure or acquisition errors from affecting subsequent analysis.

[0039] Specifically, outlier detection and filtering are performed on the three types of real-time data to obtain real-time filtered data. Outlier detection and filtering refers to identifying and removing data that deviates from the normal range using a fixed algorithm to avoid interference from equipment malfunctions and acquisition errors in subsequent analysis. The specific method is formulated based on the computing load characteristics of the target data center obtained from S2: For general data centers with stable loads, the 3σ principle (removing values ​​exceeding "data mean ± 3 times standard deviation") is used for detection, such as setting the CPU utilization anomaly threshold to >90% or <10%; for intelligent computing data centers with highly bursty tasks, dynamic... The threshold method (adjusting the threshold based on the average of data over the past hour, such as ±25% of the average) is used. If the GPU utilization suddenly increases by more than 50%, it is considered abnormal. Due to the high load and stability of the supercomputing data center, the box plot method is used (removing values ​​that exceed the range of "Q1-1.5IQR to Q3+1.5IQR"). If the total energy consumption jumps by more than 15%, it is considered abnormal. For the detected abnormal values, short-term missing values ​​(≤2 sampling periods) are repaired by linear interpolation of the data before and after. Long-term abnormal values ​​(>2 sampling periods) trigger equipment failure alarms and use the normal data of the previous period to ensure the reliability of real-time filtered data.

[0040] Furthermore, based on the computing load characteristics obtained from previous analysis (such as fluctuation patterns and periodic characteristics across multiple time scales), the real-time filtered data is calibrated, specifically including: Based on the target data center's computing load characteristics obtained from S2, the real-time filtered data is aligned to a specific time granularity. This alignment unifies the three types of real-time data to the same sampling time unit, resolving the issue of different data source collection frequencies. The granularity selection must match the load characteristics of the target data center: general data centers, due to their stable daily load, are aligned to a 5-minute granularity (e.g., taking the 5-minute average of CPU utilization data collected once per minute); intelligent computing data centers, due to sudden task bursts, are aligned to a 1-minute granularity (ensuring the capture of short-term computing power fluctuations); and supercomputing data centers, due to their pulsed high load, are aligned to a 10-second granularity (focusing on monitoring instantaneous changes during high-load periods). The aligned data must be labeled with a unified timestamp (accurate to the second), such as 2024-05-20 14:30:00, to ensure that the three types of data correspond one-to-one in the time dimension.

[0041] Dynamic correction is performed on the real-time data after time granularity alignment. Dynamic correction refers to correcting the deviation data by comparing the real-time data with the historical computing load characteristics obtained by S2 to ensure that the trend of real-time data changes matches the historical pattern. The specific method is as follows: If the target data center is a general data center, refer to the "intraday periodic characteristics" extracted by S2 (such as linear increase in load from 9:00 to 18:00). If the real-time CPU utilization rate shows a non-linear decrease during this period (such as a sharp drop from 60% to 30%), then smooth correction is performed based on the upward slope of the historical data for the same period (such as an increase of 5% per hour). If it is an intelligent computing data center, refer to the "task burst characteristics" (such as GPU utilization rate rising from 30% to 80% within 10 minutes when an AI training task starts). If the real-time GPU utilization rate only rises to 50%, then correction is performed based on the historical growth rate of similar tasks. If it is a supercomputing data center, refer to the "pulsating high load characteristics" (such as high load lasting for 8 hours). If the real-time load only lasts for 3 hours and then drops, it is necessary to check whether it is a temporary task interruption and correct it based on the historical duration. Finally, real-time sub-item data consistent with the historical load characteristic trend is obtained.

[0042] The corrected real-time sub-item data is linked and integrated according to a preset format to obtain standardized multi-source real-time data. The preset format refers to unified data storage and field rules, which must include core fields such as timestamp, data center type, real-time performance characteristics (average CPU utilization, memory usage, etc.), real-time energy consumption characteristics (total power consumption, green electricity consumption ratio, etc.), and real-time task characteristics (task queue length, adjustable task ratio, etc.). The linkage and integration must establish a "task-computing power-energy consumption" mapping relationship, such as binding "increase of batch processing tasks by 10" with "CPU utilization increased by 8%" and "total energy consumption increased by 5kW", clarifying the quantitative relationship among the three. At the same time, the data units are unified, such as CPU utilization and green electricity consumption ratio retaining 1 decimal place (%), and total energy consumption retaining 2 decimal places (kWh), to ensure that the data format is completely matched with the input requirements of the S4 computing power load dynamic prediction model.

[0043] S4: Input the multi-source real-time data into the pre-trained dynamic prediction model of computing load to obtain the computing load prediction results of the target data center at different time scales; wherein, the dynamic prediction model of computing load is configured as a multi-branch network structure to dynamically adjust the time granularity of the input data to adapt to the load patterns at different time scales.

[0044] It should be noted that the "multi-source real-time data" used in this embodiment, namely the "standardized multi-source real-time data" finally output by S3, needs to be divided into three sub-data categories according to the time scale: short-term (corresponding to predictions of the next 1-24 hours), medium-term (corresponding to predictions of the next 1-7 days), and long-term (corresponding to predictions of the next 1-3 months). The division is based on the target data center computing load characteristics obtained in S2: For general data centers with stable loads, short-term data uses data with a granularity of about 2 hours and 5 minutes, medium-term data uses data with a granularity of about 24 hours, and long-term data uses data with a granularity of about 7 days; for intelligent computing data centers with sudden task bursts, short-term data uses data with a granularity of about 1 hour and 1 minute, medium-term data uses data with a granularity of about 12 hours, and long-term data uses data with a granularity of about 3 days; for supercomputing data centers with pulsed loads, short-term data uses data with a granularity of about 30 minutes and 10 seconds, medium-term data uses data with a granularity of about 8 hours, and long-term data uses data with a granularity of about 15 days, to ensure that the time granularity of the input data matches the prediction scale.

[0045] In this embodiment, feature extraction is performed on the segmented multi-source real-time data to construct a model input feature set. Feature extraction refers to screening and quantifying parameters that are strongly correlated with computing power load from the real-time data, which needs to cover the three dimensions of "computing power-energy-task". The feature definitions and extraction methods for each dimension are as follows: Real-time performance characteristics are parameters that reflect the intensity and stability of computing load, including CPU utilization average / peak (the arithmetic mean and maximum value within the corresponding time scale, such as 24 five-minute averages for short-term use), memory utilization fluctuation coefficient (the ratio of standard deviation to mean; for general data centers, this coefficient ≤0.15 is considered stable), and network bandwidth saturation time (the cumulative time with bandwidth utilization ≥90%; for supercomputing data centers, this time accounts for ≥60%). Energy consumption characteristics are parameters that quantify the relationship between energy and computing power, including energy consumption per unit of computing power (the ratio of total energy consumption to the average CPU utilization rate, in kWh / %, where this value is 30%-50% higher than that of general-purpose centers due to the high energy consumption of GPUs in intelligent computing data centers), real-time green electricity consumption ratio (real-time green electricity consumption / total energy consumption, in %), and energy consumption sensitivity (the ratio of changes in energy consumption to changes in task volume, reflecting the degree of impact of tasks on energy consumption). Task load features are parameters that characterize the business scheduling potential, including real-time task queue length, proportion of high-priority tasks (number of high-priority tasks / total number of tasks), and number of time-adjustable tasks (batch processing determined by S2 features / number of low-priority tasks). After extraction, they are integrated in the format of "feature name-feature value-time scale-data center type" to form a structured model input feature set.

[0046] The computing load dynamic prediction model in this embodiment of the invention is a multi-branch network structure. This model is a pre-trained model based on time series deep learning, and its core design is a "one main and three branches" architecture: The main network layer uses LSTM (Long Short-Term Memory) or GRU (Gated Recurrent Unit). LSTM controls information transmission through input gates, forget gates, and output gates, which can effectively capture long-term dependencies of computing load (such as intraday periodicity and seasonality), solving the problem that traditional ARIMA models cannot handle long-term time series dependencies. The three branches correspond to short-term, medium-term, and long-term forecasts, respectively. The input time granularity and network parameters of each branch are configured differently: the short-term branch inputs 5-minute / 10-second granular data, with 2 LSTM layers and 64 neurons, focusing on hourly load fluctuations; the medium-term branch inputs hourly granular data, with 3 LSTM layers and 128 neurons, focusing on daily load patterns; and the long-term branch inputs daily granular data, with 3 LSTM layers and 256 neurons, focusing on monthly seasonal changes. Each branch adapts to different scales of load patterns by dynamically adjusting the time granularity of the input data (e.g., the short-term branch aggregates 10-second data into a 5-minute average) to ensure forecast accuracy.

[0047] The training data for this dynamic prediction model of computing load comes from the historical operating data of S1 and the computing load characteristics of various types of data centers extracted by S2 (such as peak and valley periods and periodic parameters). The training algorithm adopts the time series deep learning algorithm (LSTM / GRU), and the specific training parameters are as follows: the optimizer is Adam (initial learning rate of 0.001, decaying by 50% every 20 rounds), the loss function is mean squared error (MSE, which measures the deviation between the predicted value and the true value), the training set and the test set are divided in an 8:2 ratio (the training set uses 10 months of historical data and the test set uses 2 months of data), the number of iterations is 100 rounds (early stopping mechanism: stop when the MSE of the test set increases for 3 consecutive rounds), and the training objective is that the MSE of the test set is ≤0.08 (short-term branch ≤0.05, long-term branch ≤0.1), to ensure that the model has the ability to predict and adapt to multiple time scales, that is, it can adapt to the stable load of general data centers, the burst load of intelligent computing centers, and the pulse load of supercomputing centers.

[0048] The model input feature set for each time scale is input into the corresponding branch to perform multi-time-scale load forecasting and output the results. The forecasting process needs to be combined with the characteristics of the target data center type. Specifically: For general data centers, short-term forecasts should focus on the hourly average load and peak / valley periods for the next 24 hours (e.g., peak from 9:00 to 18:00). Medium-term forecasts should show the 7-day daily average load and weekly periodic differences (weekdays are 15% higher than weekends). Long-term forecasts should show the monthly peak load and seasonal trends for the next 3 months (summer is 20% higher than winter). For intelligent computing data centers, short-term forecasts should indicate peak load periods that may be caused by sudden task interruptions (e.g., AI training starts from 10:00 to 12:00). Medium-term forecasts should be correlated with peak green energy output days (e.g., if photovoltaic output is high the following day, it is recommended to concentrate tasks). For supercomputing data centers, short-term forecasts should specify the duration of pulsed high loads (e.g., load ≥ 90% from 14:00 to 22:00). Long-term forecasts should adapt to load increases caused by seasonal cooling demand. The final forecast results should include the time scale, forecast period, peak / valley / average load, fluctuation range (±MSE values), and renewable energy output matching suggestions to ensure complete information.

[0049] In this embodiment, the output prediction results at different time scales need to be stored in a structured database (such as PostgreSQL). The fields are associated with the computing load characteristics of S2 and the real-time data of S3, so that they can be called during S5 scheduling. At the same time, the accuracy needs to be verified by backtesting with real-time data, including correcting the short-term prediction results with the latest real-time data every hour, correcting the medium-term results every day, and correcting the long-term results every month, to ensure that the prediction deviation is within an acceptable range (short-term deviation ≤8%, medium-term ≤10%, long-term ≤12%).

[0050] S5: Based on the computing load prediction results, perform cross-scale migration scheduling for time-adjustable tasks in the target data center to optimize the load management of the target data center.

[0051] Preferably, in one embodiment of the present invention, based on the computing load prediction results output by S4, time-adjustable tasks in the target data center are accurately identified. These time-adjustable tasks refer to computing tasks that meet time elasticity thresholds, are non-core priority, and have no exclusive resource requirements. The criteria for determining these tasks need to be formulated by combining the target data center type (general / intelligent / supercomputing as defined by S1) with the load peak and valley characteristics predicted by S4. Specifically: The adjustable tasks in a general data center are batch data computation (such as user log analysis) and non-real-time backup (such as daily cold backup). The criteria for judgment are: the maximum delay time is ≥2 hours, the task priority is lower than user interaction tasks, and the computing power consumption of a single task is ≤10% of the server's rated computing power. The adjustable tasks of the intelligent computing data center are offline training of non-core AI models (such as model iteration in the test environment). The criteria for judgment are: maximum delayed execution time ≥ 12 hours, support for breakpoint resumption, and no exclusive use of GPU resources. The adjustable tasks in the supercomputing data center are low-priority scientific computing subtasks (such as local simulation verification), and the criteria for judgment are: the maximum delay in execution is ≥4 hours and the computing power consumption of a single task is ≤5% of the rated computing power of the cluster. The identification process requires calling S3's real-time task processing volume data (including task priority and estimated duration), filtering it to form a time-adjustable task pool, and labeling fields such as task ID, pre-computational power requirement, and latency threshold to provide target objects for subsequent scheduling.

[0052] Based on the characteristics of time-adjustable tasks and the green operation requirements of data centers, a multi-objective optimization scheduling model is constructed. Specifically, the multi-objective scale scheduling model refers to a multi-constraint optimization model with the objectives of "computing load smoothing + maximizing green energy consumption". Its core includes multi-dimensional objective optimization constraints and new energy output matching logic, including: Task completion rate: Ensure that migration scheduling does not affect the final delivery time of tasks. For example, the delay time of high-priority adjustable tasks does not exceed the preset threshold to avoid the risk of business interruption. Green electricity procurement cost: Taking into account the fluctuations in time-of-use green electricity prices (such as when green electricity prices are lower during peak periods of renewable energy output), the goal is to minimize costs.

[0053] The scheduling model needs to be dynamically matched with the distributed renewable energy generation curves (such as peak daytime output of photovoltaics and nighttime fluctuations of wind power) to ensure that the direction of task migration is consistent with the periods when green electricity supply is sufficient. For example, the model will prioritize migrating tasks to periods with higher photovoltaic / wind power output to maximize the proportion of green electricity consumption.

[0054] Furthermore, based on the optimization scheme output by the scheduling model, task migration is performed at different time scales to achieve load management optimization.

[0055] Specifically, for the short-term (1-24 hours), medium-term (1-7 days), and long-term (1-3 months) timescales predicted by S4, differentiated cross-scale migration and scheduling strategies are formulated: Short-term scheduling focuses on smoothing out intraday load peaks, with general data centers migrating adjustable tasks during predicted peak periods (e.g., 9:00-18:00) to valley periods (0:00-6:00), and supercomputing data centers migrating tasks expected to experience pulse-like high loads (e.g., CPU utilization ≥90%) one hour later to a smoother period three hours later; Medium-term scheduling focuses on matching peak green power output. The intelligent computing data center centrally schedules offline training tasks for days with high green power output predicted by S4 (such as the next day when the photovoltaic output accounts for ≥60%) to the same day. The general data center, taking into account the periodic characteristics (20% lower load on weekends), migrates adjustable tasks from weekdays to weekends. Long-term scheduling focuses on adapting to seasonal load changes. The supercomputing data center migrates non-urgent tasks predicted by S4 in summer (15% higher load due to cooling) to spring and autumn. All types of data centers must ensure that the seasonal load peak-valley difference is reduced by ≥25% after long-term scheduling to achieve a match between computing power load and seasonal energy supply.

[0056] Meanwhile, a clear real-time monitoring and anomaly rollback mechanism for scheduling execution is established to ensure the stability and controllability of the scheduling process: Scheduling instructions are issued through data center task scheduling platforms (such as Apache Airflow and the SLURM scheduling system of supercomputing centers). After execution, server performance indicators (CPU utilization, memory usage) and power consumption data (total energy consumption, green electricity consumption) are collected in real time to verify whether the optimization goals are met (such as a short-term load peak-valley difference reduction of ≥10% and a green electricity consumption ratio increase of ≥20%). If an anomaly occurs (such as the load exceeding the rated value by 10% after migration), a rollback mechanism is triggered: migration is suspended from low to high task priority, with priority given to retaining high-priority adjustable tasks (such as financial data backup tasks in general data centers) until the load returns to a safe range (≤90% of the rated value). Anomaly handling logs need to be stored synchronously in the database, including the anomaly time, cause, and rollback measures, to facilitate subsequent optimization of scheduling model parameters.

[0057] The migration scheduling mechanism proposed in this invention deeply integrates multi-timescale computing load management with green energy demand through a closed-loop process of "prediction-identification-modeling-execution". This not only ensures the stable operation of data center services (task completion rate constraint) but also increases the proportion of green electricity consumption and reduces energy costs (green electricity procurement cost constraint). Ultimately, it achieves the green and efficient operation goal of "power-computing power synergy" and provides a feasible implementation path for data centers to participate in the time-of-use green electricity market.

[0058] Another embodiment of the present invention provides a multi-time-scale computing load management system. For details, please refer to [link to relevant documentation]. Figure 2 , Figure 2The diagram shown illustrates a multi-timescale computing load management system according to one embodiment of the present invention, which includes: The acquisition module 11 is used to acquire historical operating data of various types of data centers, wherein the historical operating data includes server performance indicators, power consumption data and task processing volume over a long period of time. Analysis module 12 is used to analyze the computing load characteristics and corresponding processing adaptation rules of various types of data centers at multiple time scales based on the historical operation data. Integration module 13 is used to respond to the load optimization signal and, based on the computing load characteristics and the processing adaptation rules, integrate and process the real-time server performance indicators, real-time power consumption data and real-time task processing volume of the target data center to obtain multi-source real-time data at multiple time scales. The prediction module 14 is used to input the multi-source real-time data into the pre-trained dynamic prediction model of computing load to obtain the computing load prediction results of the target data center at different time scales; wherein, the dynamic prediction model of computing load is configured as a multi-branch network structure to dynamically adjust the time granularity of the input data to adapt to the load patterns at different time scales. The scheduling module 15 is used to perform cross-scale migration scheduling of time-adjustable tasks in the target data center based on the computing load prediction results, so as to achieve load management optimization of the target data center.

[0059] As one preferred embodiment, the data center includes a general-purpose data center, an intelligent computing data center, and a supercomputing data center; the analysis module 12 is specifically used for: The historical operational data is preprocessed to obtain a standardized historical load dataset; Based on the classification of data centers, the server performance indicators, power consumption data and task processing volume in the historical load dataset are classified and statistically analyzed to obtain the basic load characteristic data of each type of data center. The load change data at multiple time scales in the basic load characteristic data are analyzed in layers to extract the load peak and valley periods, fluctuation amplitude and periodicity characteristics at each time scale, so as to obtain the computing load characteristics of each type of data center at multiple time scales, and generate corresponding processing adaptation rules based on the computing load characteristics of each type of data center.

[0060] As one preferred embodiment, the integration module 13 is specifically used for: Acquire real-time operational data of the target data center, including real-time server performance metrics, real-time power consumption data, and real-time task processing volume; Based on the processing adaptation rules of the target data center, outlier detection and filtering are performed on the real-time server performance indicators, the real-time power consumption data, and the real-time task processing volume to obtain real-time filtered data. Based on the computing load characteristics of the target data center, the real-time filtered data is sequentially aligned with time granularity and dynamically corrected to obtain corrected real-time sub-item data; wherein, the dynamic correction is designed to match the changing trend of the real-time filtered data with the historical load characteristics. The real-time segmented data are correlated and integrated according to a preset format to obtain standardized multi-source real-time data at multiple time scales.

[0061] As one preferred embodiment, the prediction module 14 is specifically used for: Feature extraction is performed on the multi-source real-time data, and model input feature sets are constructed at each time scale based on the feature extraction results; wherein, the model input feature sets include real-time performance features, energy consumption features, and task load features related to computing power load; The model input feature sets at each time scale are respectively input into the dynamic prediction model of computing load to perform multi-time scale load prediction, so as to obtain the computing load prediction results of the target data center at different time scales.

[0062] As one preferred embodiment, the scheduling module 15 is specifically used for: Based on the computing load prediction results, analyze the time migration characteristics of computing tasks in the target data center and identify time-adjustable tasks in the target data center computing tasks. Based on the characteristics of the time-adjustable tasks and the green operation requirements of the target data center, a cross-time-scale scheduling model for computing tasks adapted to new energy output is constructed. Based on the aforementioned scheduling model, time-adjustable tasks are flexibly migrated and scheduled at different time scales to match the computing load prediction results with the fluctuations in the new energy power generation curve.

[0063] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A multi-timescale computing load management method, characterized in that, include: Acquire historical operational data for various types of data centers, including server performance metrics, power consumption data, and task processing volume over a long period; wherein, the data centers include general-purpose data centers, intelligent computing data centers, and supercomputing data centers; Based on the historical operational data, the computing load characteristics and corresponding processing adaptation rules of various types of data centers at multiple time scales are analyzed sequentially, specifically including: The historical operational data is preprocessed to obtain a standardized historical load dataset; Based on the classification of data centers, the server performance indicators, power consumption data and task processing volume in the historical load dataset are classified and statistically analyzed to obtain the basic load characteristic data of each type of data center. The load change data at multiple time scales in the basic load characteristic data are analyzed in layers to extract the load peak and valley periods, fluctuation amplitude and periodic characteristics at each time scale, so as to obtain the computing load characteristics of each type of data center at multiple time scales, and generate corresponding processing adaptation rules based on the computing load characteristics of each type of data center. In response to the load optimization signal, based on the computing load characteristics and the processing adaptation rules, the acquired real-time server performance indicators, real-time power consumption data, and real-time task processing volume of the target data center are integrated and processed to obtain multi-source real-time data at multiple time scales, specifically including: Acquire real-time operational data of the target data center, including real-time server performance metrics, real-time power consumption data, and real-time task processing volume; Based on the processing adaptation rules of the target data center, outlier detection and filtering are performed on the real-time server performance indicators, the real-time power consumption data, and the real-time task processing volume to obtain real-time filtered data. Based on the computing load characteristics of the target data center, the real-time filtered data is sequentially aligned with time granularity and dynamically corrected to obtain corrected real-time sub-item data; wherein, the dynamic correction is designed to match the changing trend of the real-time filtered data with the historical load characteristics. The real-time sub-data is correlated and integrated according to a preset format to obtain standardized multi-source real-time data at multiple time scales; The multi-source real-time data are respectively input into the pre-trained dynamic prediction model of computing load to obtain the computing load prediction results of the target data center at different time scales; wherein, the dynamic prediction model of computing load is configured as a multi-branch network structure to dynamically adjust the time granularity of the input data to adapt to the load pattern at different time scales. Based on the computing load prediction results, time-adjustable tasks in the target data center are migrated and scheduled across scales to optimize the load management of the target data center.

2. The multi-timescale computing load management method as described in claim 1, characterized in that, The step of inputting the multi-source real-time data into the pre-trained dynamic prediction model of computing load to obtain the computing load prediction results of the target data center at different time scales includes: Feature extraction is performed on the multi-source real-time data, and model input feature sets are constructed at each time scale based on the feature extraction results; wherein, the model input feature sets include real-time performance features, energy consumption features, and task load features related to computing power load; The model input feature sets at each time scale are respectively input into the dynamic prediction model of computing load to perform multi-time scale load prediction, so as to obtain the computing load prediction results of the target data center at different time scales.

3. The multi-timescale computing load management method as described in claim 1, characterized in that, The cross-scale migration scheduling of time-adjustable tasks in the target data center based on the computing load prediction results includes: Based on the computing load prediction results, analyze the time migration characteristics of computing tasks in the target data center and identify time-adjustable tasks in the target data center computing tasks. Based on the characteristics of the time-adjustable tasks and the green operation requirements of the target data center, a cross-time-scale scheduling model for computing tasks adapted to new energy output is constructed. Based on the aforementioned scheduling model, time-adjustable tasks are flexibly migrated and scheduled at different time scales to match the computing load prediction results with the fluctuations in the new energy power generation curve.

4. A multi-timescale computing load management system, characterized in that, include: The acquisition module is used to acquire historical operational data of various types of data centers, including server performance indicators, power consumption data, and task processing volume over a long period of time; the data centers include general data centers, intelligent computing data centers, and supercomputing data centers. The analysis module is used to sequentially analyze the computing load characteristics and corresponding processing adaptation rules of various types of data centers at multiple time scales based on the historical operating data; specifically, the analysis module is used for: The historical operational data is preprocessed to obtain a standardized historical load dataset; Based on the classification of data centers, the server performance indicators, power consumption data and task processing volume in the historical load dataset are classified and statistically analyzed to obtain the basic load characteristic data of each type of data center. The load change data at multiple time scales in the basic load characteristic data are analyzed in layers to extract the load peak and valley periods, fluctuation amplitude and periodic characteristics at each time scale, so as to obtain the computing load characteristics of each type of data center at multiple time scales, and generate corresponding processing adaptation rules based on the computing load characteristics of each type of data center. The integration module, in response to the load optimization signal, integrates the acquired real-time server performance indicators, real-time power consumption data, and real-time task processing volume of the target data center based on the computing load characteristics and the processing adaptation rules to obtain multi-source real-time data at multiple time scales; the integration module is specifically used for: Acquire real-time operational data of the target data center, including real-time server performance metrics, real-time power consumption data, and real-time task processing volume; Based on the processing adaptation rules of the target data center, outlier detection and filtering are performed on the real-time server performance indicators, the real-time power consumption data, and the real-time task processing volume to obtain real-time filtered data. Based on the computing load characteristics of the target data center, the real-time filtered data is sequentially aligned with time granularity and dynamically corrected to obtain corrected real-time sub-item data; wherein, the dynamic correction is designed to match the changing trend of the real-time filtered data with the historical load characteristics. The real-time sub-data is correlated and integrated according to a preset format to obtain standardized multi-source real-time data at multiple time scales; The prediction module is used to input the multi-source real-time data into the pre-trained dynamic prediction model of computing load to obtain the computing load prediction results of the target data center at different time scales; wherein, the dynamic prediction model of computing load is configured as a multi-branch network structure to dynamically adjust the time granularity of the input data to adapt to the load patterns at different time scales. The scheduling module is used to perform cross-scale migration scheduling of time-adjustable tasks in the target data center based on the computing load prediction results, so as to optimize the load management of the target data center.

5. The multi-timescale computing load management system as described in claim 4, characterized in that, The prediction module is specifically used for: Feature extraction is performed on the multi-source real-time data, and a model input feature set is constructed based on the feature extraction results; wherein, the model input feature set includes real-time performance features, energy consumption features, and task load features related to computing power load; The model input feature sets at each time scale are respectively input into the dynamic prediction model of computing load to perform multi-time scale load prediction, so as to obtain the computing load prediction results of the target data center at different time scales.

6. The multi-timescale computing load management system as described in claim 4, characterized in that, The scheduling module is specifically used for: Based on the computing load prediction results, analyze the time migration characteristics of computing tasks in the target data center and identify time-adjustable tasks in the target data center computing tasks. Based on the characteristics of the time-adjustable tasks and the green operation requirements of the target data center, a cross-time-scale scheduling model for computing tasks adapted to new energy output is constructed. Based on the aforementioned scheduling model, time-adjustable tasks are flexibly migrated and scheduled at different time scales to match the computing load prediction results with the fluctuations in the new energy power generation curve.

Images