Calculation center load prediction method and system based on data driving

Abstract

The invention provides a computing power center load prediction method and system based on data driving, and the method comprises the steps: predicting a reference load from a minute level to a day level in the future and an adjustable margin of the reference load based on the internal operation parameters and historical load data of a computing power center; based on historical and planning information of the number of racks, a saturation-adaptive growth rate decline model is adopted to predict the number of future racks, and based on historical and development trends of power consumption of a single rack, a linear regression model is adopted to predict average power consumption of the single rack in the future, and then long-term total electric quantity prediction in the future is obtained through calculation. The output of the long-term-medium-and-long-term prediction module is used as a reference line, and the prediction range of the ultra-short-term-short-term prediction module is constrained; meanwhile, the deviation between the actual operation data accumulated by the ultra-short term-short term prediction module and the prediction data is used as a feedback signal, and parameters of the long term-medium and long term evolution prediction model are dynamically calibrated. According to the method, the full-cycle load prediction from the minute grade to the ten grade is realized.

Description

Technical Field

[0001] This invention relates to the fields of information technology and energy management technology, and more specifically, to a data-driven method and system for predicting the load of computing centers. Background Technology

[0002] With the rapid development of cloud computing, artificial intelligence, and big data technologies, the scale and energy consumption of computing centers are increasing daily. Their power load (including IT equipment load and cooling system load) is a core component of operating costs. Accurate load forecasting is key to achieving efficient and energy-saving operation, dynamic resource scheduling, and optimized power costs for computing centers.

[0003] Existing technologies include load forecasting methods for traditional buildings or industrial users, but these methods typically rely on macroscopic factors such as weather and calendars, making it difficult to capture the dynamic impact of the operating status of IT equipment within the computing center on the load. Furthermore, most existing research on computing center load forecasting focuses on a single time scale, either using short-term forecasts for real-time scheduling or long-term forecasts for capacity planning, with the two being independent of each other.

[0004] Therefore, there is an urgent need for a comprehensive forecasting solution that can span different time scales, make full use of internal and external data, and dynamically assess load regulation capabilities. Summary of the Invention

[0005] The problem addressed by this invention is to overcome the shortcomings of existing technologies and provide a method for panoramic prediction of computing center load that can seamlessly span from minutes to ten years.

[0006] To address the above problems, this invention provides a data-driven method for predicting the load of computing centers, the method comprising the following steps: Step S1: Using the ultra-short-term to short-term forecasting module, based on the internal operating parameters of the computing center and historical load data, a deep neural network model is used to predict the future baseline load and its adjustable margin from minutes to days. Step S2: Using the long-term and medium-term forecasting module, based on the historical and planning information of the number of racks, a saturation-adaptive growth rate decreasing model is used to predict the future number of racks, and based on the historical and development trend of single rack power consumption, a linear regression model is used to predict the future average power consumption of a single rack, thereby calculating the total power consumption forecast for the long term. Step S3: Using the coupled feedback and dynamic calibration module, the output of the long-term to medium-term prediction module is used as a baseline to constrain the prediction range of the ultra-short-term prediction module; at the same time, the deviation between the actual operating data and the prediction data accumulated by the ultra-short-term prediction module is used as a feedback signal to dynamically calibrate the parameters of the long-term to medium-term evolution prediction model.

[0007] As a preferred embodiment, in step S1, the internal operating parameters of the computing center include IT equipment resource indicators, business scheduling indicators, and infrastructure indicators; wherein, The IT equipment resource metrics include one or more of the following: CPU utilization, GPU utilization, memory usage, storage IOPS, and network traffic. The business scheduling metrics include one or more of the following: task queue length and number of real-time computing jobs; The infrastructure metrics include one or more of the following: cooling system power consumption, power efficiency, and ambient temperature.

[0008] Furthermore, in step S1, the internal operating parameters further include an adjustable capability driving factor; the adjustable capability driving factor includes: Real-time interruptible task power is an estimated total power consumed by low-priority or interruptible computing tasks obtained by querying the task management system. The inertial adjustment potential of the refrigeration system is the temperature adjustment range calculated based on the current refrigeration system temperature setpoint and the allowable upper and lower safety limits, which is mapped to the adjustable refrigeration power. The energy storage system's charge and discharge status includes the current state of charge (SOC) and maximum dischargeable power.

[0009] As a preferred embodiment, in step S1, the deep neural network model is a time-series prediction model based on an attention mechanism. The deep neural network model adopts a multi-task learning architecture, which includes a shared hidden layer and three parallel-connected output layers, namely: The first output layer is used to output the baseline expected load values ​​for future time series. The second output layer is used to output the adjustable lower limit of the load for future time series. The third output layer is used to output the adjustable upper limit of the load for future time series. The baseline load expectation, the lower limit of load adjustment, and the upper limit of load adjustment together constitute the baseline load and its adjustment margin.

[0010] As a preferred embodiment, in step S2, the formula for the saturation adaptive decreasing growth rate model is as follows: , in, Let be the predicted number of racks in year t. Let k be the actual number of racks in year t-1, and k be the maximum number of racks the data center can accommodate. For internal growth rate, This is the curve shape coefficient.

[0011] Furthermore, the intrinsic growth rate A dynamic parameter modulated by external macroeconomic trends , The specific calculation formula is as follows: , in, Based on the growth rate, This is the modulation sensitivity coefficient. This is an external technology-market trend factor built based on industry reports and company strategic data.

[0012] As a preferred embodiment, the specific process for predicting the future average power consumption per rack is as follows: Computing center racks are classified into several types according to power density and function; For each rack type, a trend prediction model for single rack power consumption is established. Predict the future proportion of each type of rack in the total number of racks; The global average rack power consumption is calculated by weighting the predicted power consumption of each type of rack according to its predicted proportion.

[0013] As a preferred embodiment, in step S3, the long-term output is used as a baseline to constrain short-term predictions, and the parameters of the long-term-medium-term evolution prediction model are dynamically calibrated, specifically as follows: The annual total electricity consumption obtained from long-term forecasts is decomposed into daily baselines, and a forecast consistency confidence interval around the daily baselines is calculated based on the historical uncertainty of long-term forecasts. Based on confidence intervals and historical experience, a deviation threshold is obtained, and the output of the ultra-short-term to short-term prediction module is required to have a cumulative effect within the deviation threshold.

[0014] Furthermore, the parameters of the long-term and medium-term evolution prediction model are dynamically calibrated, specifically as follows: Monitor the cumulative deviation of short-term forecasts and trigger the calibration process when the deviation continues to exceed the allowable deviation threshold; Before calibration, a deviation source analysis is performed to decompose the total load deviation into the contributions of the rack quantity prediction deviation and the single rack power consumption prediction deviation, as shown in the following formula: , Where ∆L is the total load deviation, Na and Pa are the actual number of racks and the power consumption per rack calculated from recent actual data, respectively, and Nf and Pf are the predicted values ​​of the number of racks and the power consumption per rack for the corresponding period, respectively. Based on the source analysis results, the parameters of the long-term and medium-term evolution prediction models were selectively recalibrated.

[0015] This invention also provides a data-driven computing center load forecasting system to implement the above-described data-driven computing center load forecasting method. The system includes: The data acquisition and preprocessing module is used to collect internal operating parameters and historical data of the computing center, and to preprocess the collected data. The ultra-short-term forecasting module is used to predict the baseline load and its adjustable margin in the future, from minutes to days, based on the internal operating parameters and historical load data of the computing center and through a deep neural network model. The long-term and medium-term forecasting module is used to predict the future number of racks based on historical and planning information on the number of racks, using a saturation-adaptive growth rate decreasing model, and to predict the future average power consumption of a single rack based on historical and development trends of single rack power consumption, using a linear regression model, and then calculate the total power consumption forecast for the long term. The coupled feedback and dynamic calibration module is used to use the output of the long-term-medium-term prediction module as a baseline to constrain the prediction range of the ultra-short-term prediction module; at the same time, it uses the deviation between the actual operating data and the prediction data accumulated by the ultra-short-term prediction module as a feedback signal to dynamically calibrate the parameters of the long-term-medium-term evolution prediction model. The prediction results output and visualization module is used to output and display the prediction results. Beneficial effects

[0016] This invention provides a data-driven load forecasting method and system for computing centers, achieving full-cycle load forecasting from minute-level to ten-year-level, meeting different levels of needs from real-time operation to strategic planning. Through a unique coupling feedback mechanism, short-term forecasts utilize a long-term perspective, and long-term forecasts obtain high-frequency data calibration, fundamentally improving the forecasting accuracy of the system at different time scales. For the first time, the adjustable margin is quantified in load forecasting, opening up a new path for computing centers to participate in grid peak shaving and peak regulation as flexible loads, thereby reducing energy costs. Attached Figure Description

[0017] Figure 1 A flowchart of a data-driven computing center load prediction method provided in an embodiment of the present invention; Figure 2 A block diagram of a data-driven computing center load prediction system provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of a computing device provided in an embodiment of the present invention. Detailed Implementation

[0018] In embodiments of the present invention, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" or "for example" in embodiments of the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0019] To ensure clarity and conciseness in the description of the following embodiments, a brief introduction to the related technologies is given first: Unlike conventional high-energy-consuming loads, computing centers are the only new type of load that can achieve instantaneous transfer of power load without relying on the power grid. They have the potential for wide-area, second-level, time-series scheduling. Dispersed power loads urgently need centralized integration and optimized control, and further exploration of efficient collaborative models between power and computing power is needed.

[0020] Existing methods for predicting the load of computing centers have the following limitations: Time-scale fragmentation: Traditional forecasting methods typically handle forecasting problems at different time scales independently. Ultra-short-term forecasts (e.g., minute-level) rely on real-time data sequences but lack guidance on long-term trends; long-term forecasts (e.g., interannual) are based on macro-planning and fail to reflect short-term fluctuations and adjustability potential. This fragmentation leads to serious deficiencies in the consistency and synergy of forecast results.

[0021] Static model rigidity: Once most models are trained, their parameters remain fixed. However, the workload, hardware configuration, and energy efficiency strategies of computing centers evolve dynamically, and static models cannot adapt to these changes, causing prediction accuracy to decline over time.

[0022] Ignoring adjustability: Most existing methods treat computing load as an immutable, rigid demand for prediction. In reality, computing load has a significant "adjustability margin" through task scheduling, liquid cooling system regulation, and energy storage system charging and discharging. Ignoring this characteristic makes the prediction results unsupportive of advanced energy interaction applications (such as demand response).

[0023] Disconnected from the evolution of physical resources: Long-term forecasts often rely simply on the extrapolation of historical power consumption, failing to deeply connect with the most fundamental driving factors of computing centers—the growth in the number of racks and the technological evolution of power consumption per rack. This results in long-term forecasts lacking physical basis and having low reliability.

[0024] In view of this, embodiments of the present invention provide a data-driven method for predicting the load of a computing center, which can be executed by a processing device. This processing device can be a terminal or a server. Terminals include, but are not limited to, smartphones, tablets, laptops, personal digital assistants, or smart wearable devices. Servers can be cloud servers, such as central servers in a central cloud computing cluster or edge servers in an edge cloud computing cluster. Of course, servers can also be servers in a local data center. A local data center refers to a data center directly controlled by the user.

[0025] To address the problems of fragmented time scales, static and rigid models, neglect of adjustability, and disconnect from physical resource evolution in traditional computing center load forecasting, this invention breaks down the barriers between independent long-term and short-term models in traditional forecasting methods through a multi-time-scale coupled forecasting architecture. By employing a "top-down constraint" and "bottom-up feedback" mechanism, it achieves a deep integration of macro-evolutionary trends and micro-fluctuation characteristics, ensuring the consistency and synergy of panoramic forecasting results from minutes to years in a dynamic process.

[0026] To make the technical solution of this application clearer and easier to understand, the following description, in conjunction with the accompanying drawings, introduces a data-driven computing center load prediction method provided by an embodiment of this application. Figure 1 As shown, this figure is a flowchart of a data-driven computing center load prediction method provided in an embodiment of this application. The method includes: Step S1: Using the ultra-short-term to short-term forecasting module, based on the internal operating parameters of the computing center and historical load data, a deep neural network model is used to predict the future baseline load and its adjustable margin from minutes to days.

[0027] Specifically, collecting historical load data means collecting historical total power consumption data, with a time resolution of at least 15 minutes, and preferably minutes.

[0028] Internal operating parameters include IT equipment resource indicators, business scheduling indicators, and infrastructure indicators; among them, The IT equipment resource metrics include one or more of the following: CPU utilization, GPU utilization, memory usage, storage IOPS, and network traffic; IOPS is an important metric for measuring storage device performance, representing the number of read and write operations that the storage device can complete per second. The business scheduling metrics include one or more of the following: task queue length and number of real-time computing jobs; The infrastructure metrics include one or more of the following: cooling system power consumption (chillers, pumps, cooling towers, computer room air conditioning), power usage efficiency (PUE), and ambient temperature (cold aisle / hot aisle).

[0029] External related data: date type (weekday, weekend, holiday), time period within a day (peak, average, valley).

[0030] The deep neural network model is a time-series prediction model based on an attention mechanism. The deep neural network model adopts a multi-task learning architecture, which includes a shared hidden layer and three parallel-connected output layers, namely: The first output layer is used to output the baseline expected load values ​​for future time series. The second output layer is used to output the adjustable lower limit of the load for future time series. The third output layer is used to output the adjustable upper limit of the load for future time series. The baseline load expectation, the lower limit of load adjustment, and the upper limit of load adjustment together constitute the baseline load and its adjustment margin.

[0031] When training the deep neural network model, its loss function is a weighted sum of the losses of the first output layer, the second output layer, and the third output layer.

[0032] The loss functions for the first, second, and third output layers are either Huber loss or quantile loss.

[0033] Step S2: Using the long-term and medium-term forecasting module, based on the historical and planning information of the number of racks, a saturation-adaptive growth rate decreasing model is used to predict the future number of racks, and based on the historical and development trend of single rack power consumption, a linear regression model is used to predict the future average power consumption of a single rack, thereby calculating the total power consumption forecast for the long term. The formula for the saturation-adaptive decreasing growth rate model is as follows: , in, Let be the predicted number of racks in year t. Let k be the actual number of racks in year t-1, and k be the maximum number of racks the data center can accommodate. For internal growth rate, This is the curve shape coefficient.

[0034] The specific process for predicting the future average power consumption per rack is as follows: Computing center racks are classified into several types according to power density and function; For each rack type, a trend prediction model for single rack power consumption is established. Predict the future proportion of each type of rack in the total number of racks; The global average rack power consumption is calculated by weighting the predicted power consumption of each type of rack according to its predicted proportion.

[0035] Step S3: Using the coupled feedback and dynamic calibration module, the output of the long-term to medium-term prediction module is used as a baseline to constrain the prediction range of the ultra-short-term prediction module; at the same time, the deviation between the actual operating data and the prediction data accumulated by the ultra-short-term prediction module is used as a feedback signal to dynamically calibrate the parameters of the long-term to medium-term evolution prediction model.

[0036] The long-term output is used as a baseline to constrain short-term predictions, and the parameters of the long-term-medium-term evolution prediction model are dynamically calibrated, specifically as follows: The annual total electricity consumption obtained from long-term forecasts is decomposed into daily baselines, and a forecast consistency confidence interval around the daily baselines is calculated based on the historical uncertainty of long-term forecasts. Based on confidence intervals and historical experience, a deviation threshold is obtained, and the output of the ultra-short-term to short-term prediction module is required to have a cumulative effect within the deviation threshold.

[0037] The parameters of the long-term and medium-term evolution prediction model are dynamically calibrated, specifically as follows: Monitor the cumulative deviation of short-term forecasts and trigger the calibration process when the deviation continues to exceed the allowable deviation threshold; Before calibration, a deviation source analysis is performed to decompose the total load deviation into the contributions of the rack quantity prediction deviation and the single rack power consumption prediction deviation, as shown in the following formula: , Where ∆L is the total load deviation, Na and Pa are the actual number of racks and the power consumption per rack calculated from recent actual data, respectively, and Nf and Pf are the predicted values ​​of the number of racks and the power consumption per rack for the corresponding period, respectively. Based on the source analysis results, the parameters of the long-term and medium-term evolution prediction models were selectively recalibrated.

[0038] In some embodiments, in step S1, the internal operating parameters further include an adjustable capability driving factor; the adjustable capability driving factor includes: Real-time interruptible task power is an estimated total power consumed by low-priority or interruptible computing tasks obtained by querying the task management system. Specifically, the task management system is used to statistically estimate the total power consumed by low-priority or interruptible computing tasks (such as batch processing jobs and model training tasks). This load is the primary target for "computational load migration" or "task peak shaving." Specifically, a temperature adjustment range is calculated based on the current chiller unit outlet water temperature setpoint and the allowable safety limits. This range represents the instantaneous cooling power saved by increasing the chilled water supply temperature by 1-2℃ (without affecting equipment safety). This is an easily obtainable and highly effective feature.

[0039] The inertial regulation potential of a cooling system is a temperature regulation range calculated based on the current cooling system temperature setpoint and permissible safety limits. This range maps to adjustable cooling power; specifically, it includes the current SOC (State of Charge) and maximum dischargeable power. This is directly related to the ability to offset IT loads through energy storage discharge.

[0040] In some embodiments, in step S2, the intrinsic growth rate A dynamic parameter modulated by external macroeconomic trends , The specific calculation formula is as follows: , in, Based on the growth rate, This is the modulation sensitivity coefficient. This is an external technology-market trend factor constructed based on industry reports and company strategic data. Among them: , in, , , All weighting coefficients are derived from historical experience. , , These are the annual growth rate index of AI computing power demand, the annual growth percentage of enterprise cloud adoption rate, and the coefficient of company strategic investment direction. All of these data come from publicly available industry reports or internal company strategic documents.

[0041] The above prediction method will be implemented using a large computing data center as an example below.

[0042] Preparation phase: Collect minute-level data from the monitoring system over the past year: total power consumption, power consumption of IT equipment in each area, cooling system power consumption, average CPU / GPU utilization, rack quantity change records, etc.

[0043] Obtain rack planning diagrams for the next five years from the infrastructure management department, including planned rack deployment times and estimated full rack capacity times, and determine the maximum capacity K=5000 racks.

[0044] We collect data from industry reports on the evolution of high-density server technology as a reference for single-rack power consumption trends.

[0045] Model initialization and pre-training phase: The ultra-short-term prediction module uses the Informer model. Input features include: total power consumption, CPU utilization, and outdoor dry-bulb temperature over the past 72 hours. Output is a baseline load value at 96 points at 15-minute intervals for the next 24 hours, and a 90% confidence interval obtained through quantile regression as the upper and lower boundaries of the adjustable margin. Model training is performed using data from the past 6 months.

[0046] The long-term and medium-term forecasting module includes rack quantity forecasting: based on the annual rack quantity data of the past 5 years, a saturation adaptive model is fitted, with initial estimates of α=0.3 and β=1.2; Single rack power consumption prediction: Based on the annual average single rack power consumption data, linear regression was performed to obtain an annual growth rate of approximately γ = 50 watts / rack / year.

[0047] Online operation and coupled feedback: The system automatically runs short-term forecasts daily, generating detailed load curves for the next 24 hours.

[0048] The system automatically runs long-term forecasts every quarter, generating annual total electricity consumption forecasts for the next 5 years and breaking them down into daily average electricity consumption as a reference baseline for short-term forecasts.

[0049] Example of coupling: Assume a long-term forecast of an average daily electricity consumption of 1.2 million kWh for the next year. On a certain day, a short-term forecast of daily electricity consumption is 1.25 million kWh, significantly exceeding the baseline. The system will issue an alert, indicating that the short-term trend may deviate from the long-term plan and requires business verification.

[0050] Example of feedback calibration: For a consecutive quarter, the daily average power consumption forecast in the short term was 5% lower than the long-term forecast baseline. The system automatically triggered the calibration process, using the latest actual rack growth data and power consumption data to refit the long-term model parameters, correcting the rack growth rate α from 0.28 to 0.26 and the annual power consumption growth rate γ per rack from 50 to 45. This makes the long-term forecast more in line with actual development.

[0051] Through the implementation of this embodiment, the computing center has achieved high-precision, full-time-scale dynamic load forecasting, which has been successfully applied to demand response projects with power grid companies and has reduced energy procurement costs through accurate electricity consumption forecasting.

[0052] Please see Figure 2This invention also provides a data-driven computing center load prediction system, the system comprising: The data acquisition and preprocessing module is used to collect internal operating parameters and historical data of the computing center, and to preprocess the collected data. The ultra-short-term forecasting module is used to predict the baseline load and its adjustable margin in the future, from minutes to days, based on the internal operating parameters and historical load data of the computing center and through a deep neural network model. The long-term and medium-term forecasting module is used to predict the future number of racks based on historical and planning information on the number of racks, using a saturation-adaptive growth rate decreasing model, and to predict the future average power consumption of a single rack based on historical and development trends of single rack power consumption, using a linear regression model, and then calculate the total power consumption forecast for the long term. The coupled feedback and dynamic calibration module is used to use the output of the long-term-medium-term prediction module as a baseline to constrain the prediction range of the ultra-short-term prediction module; at the same time, it uses the deviation between the actual operating data and the prediction data accumulated by the ultra-short-term prediction module as a feedback signal to dynamically calibrate the parameters of the long-term-medium-term evolution prediction model. The prediction results output and visualization module is used to output and display the prediction results.

[0053] The data-driven computing center load forecasting system according to the embodiments of this application can correspond to the execution of the methods described in the embodiments of this application, and the other operations and / or functions of each module / unit of the data-driven computing center load forecasting system are respectively for implementing Figure 1 For the sake of brevity, the corresponding processes of each method in the illustrated embodiments will not be described in detail here.

[0054] This application also provides a computing device. For example... Figure 3 As shown in the figure, this is a schematic diagram of a computing device provided in an embodiment of this application. The computing device 700 includes a bus 701, a processor 702, a communication interface 703, and a memory 704. The processor 702, the memory 704, and the communication interface 703 communicate with each other via the bus 701.

[0055] The 701 bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, Figure 3 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0056] The processor 702 can be any one or more of the following processors: central processing unit (CPU), graphics processing unit (GPU), microprocessor (MP), or digital signal processor (DSP).

[0057] The communication interface 703 is used for communication with external devices.

[0058] Memory 704 may include volatile memory, such as random access memory (RAM). Memory 704 may also include non-volatile memory, such as read-only memory (ROM), flash memory, hard disk drive (HDD), or solid state drive (SSD).

[0059] The memory 704 stores executable code, and the processor 702 executes the executable code to perform the aforementioned data-driven computing center load prediction method.

[0060] Specifically, in achieving Figure 2 In the case of the illustrated embodiment, and Figure 2 When the modules or units of the data-driven computing center load prediction system described in the embodiments are implemented in software, the execution... Figure 2 The software or program code required for the functions of each module / unit can be partially or entirely stored in memory 704. Processor 702 executes the program code corresponding to each unit stored in memory 704 to execute the aforementioned data-driven computing center load prediction method.

[0061] This application also provides a computer-readable storage medium. The computer-readable storage medium can be any available medium that a computing device can store, or a data storage device such as a data center containing one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive). The computer-readable storage medium includes instructions that instruct the computing device to execute the aforementioned data-driven computing center load prediction method.

[0062] This application also provides a computer program product comprising one or more computer instructions. When the computer instructions are loaded and executed on a computing device, all or part of the processes or functions described in this application are generated.

[0063] The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, or data center to another website, computer, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0064] When the computer program product is executed by a computer, the computer executes any of the aforementioned data-driven computing center load forecasting methods. The computer program product can be a software installation package; when any of the aforementioned data-driven computing center load forecasting methods needs to be used, the computer program product can be downloaded and executed on the computer.

[0065] The descriptions of the processes or structures corresponding to the above figures each have their own emphasis. For parts of a process or structure that are not described in detail, please refer to the relevant descriptions of other processes or structures.

[0066] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in this application should be covered within the scope of protection of this application.

Claims

1. A data-driven computing power center load prediction method, characterized in that, The method comprises the following steps: Step S1: using a super-short-term-short-term prediction module, based on the internal operating parameters and historical load data of the computing power center, predicting the benchmark load and its adjustable margin in the future minute to day level through a deep neural network model; Step S2: using a long-term-medium-long-term prediction module, based on the historical and planning information of the rack number, using a saturation-adaptive growth rate decreasing model to predict the future rack number, and based on the historical and development trend of single-rack power consumption, using a linear regression model to predict the future single-rack average power consumption, and then calculating the total power prediction in the future long term; Step S3: using a coupling feedback and dynamic calibration module, taking the output of the long-term-medium-long-term prediction module as the benchmark, constraining the prediction range of the super-short-term-short-term prediction module; at the same time, taking the deviation between the actual operation data and the prediction data accumulated by the super-short-term-short-term prediction module as a feedback signal, and dynamically calibrating the parameters of the long-term-medium-long-term evolution prediction model.

2. The data-driven computing center load prediction method of claim 1, wherein, In step S1, the internal operating parameters of the computing power center include IT device resource indicators, business scheduling indicators, and infrastructure indicators; wherein, The IT device resource indicators include one or more of CPU utilization, GPU utilization, memory occupancy, storage IOPS, and network traffic; The business scheduling indicators include one or more of task queue length and real-time computing job quantity; The infrastructure indicators include one or more of cooling system power consumption, power usage efficiency, and environmental temperature.

3. The data-driven based computing power center load prediction method of claim 2, wherein, In step S1, the internal operating parameters further include adjustable capacity driving factors; The adjustable capacity driving factors include: Real-time interruptable task power, which is the total power consumption estimate of low-priority or interruptable computing tasks obtained by querying the task management system; Refrigeration system inertia adjustment potential, which is the temperature adjustment interval calculated according to the current refrigeration system temperature set value and the allowed safety upper and lower limits, which is mapped to the adjustable refrigeration power; Energy storage system charge and discharge state, including current state of charge SOC and maximum dischargeable power.

4. The data-driven computing center load prediction method of claim 1, wherein, In step S1, the deep neural network model is a time series prediction model based on an attention mechanism, and the deep neural network model adopts a multi-task learning architecture, which includes shared hidden layers and three output layers connected in parallel, respectively: The first output layer is used to output the benchmark load expectation value of the future time series; The second output layer is used to output the lower limit of the adjustable load of the future time series; The third output layer is used to output the upper limit of the adjustable load of the future time series; Wherein, the benchmark load expectation value, the lower limit of the adjustable load, and the upper limit of the adjustable load together constitute the benchmark load and its adjustable margin.

5. The data-driven computing center load prediction method of claim 1, wherein, In step S2, the formula of the saturation-adaptive growth rate decreasing model is as follows: ， wherein, is the predicted number of racks in year t, is the actual number of racks in year t-1, k is the maximum number of racks that can be housed in a room, is the intrinsic growth rate, is the curve shape coefficient.

6. The data-driven computing center load prediction method of claim 5, wherein, The intrinsic growth rate is a dynamic parameter modulated by external macro trends , The specific formula is as follows: ， wherein, is the base growth rate, is the modulation sensitivity coefficient, is an external technology-market trend factor built on industry reports and company strategy data.

7. The data-driven computing center load prediction method of claim 1, wherein, The specific process of predicting the future single-rack average power consumption is as follows: Divide the computing power center racks into multiple types according to power density and function; For each type of rack, a trend prediction model of single-rack power consumption is established respectively; Predict the proportion of each type of rack in the total rack number in the future; The average single-rack power consumption is calculated by weighting and averaging the predicted power consumption of each type of rack according to its predicted proportion. 8.The data-driven computing center load prediction method of claim 1, wherein, In step S3, the long-term output is used as a baseline to constrain the short-term prediction, and the parameters of the long-term-Mid-long-term evolution prediction model are dynamically calibrated, specifically: The annual total power calculated by the long-term prediction is decomposed into a daily baseline, and a prediction consistency confidence interval around the daily baseline is calculated based on the historical uncertainty of the long-term prediction; According to the confidence interval and historical experience, a deviation threshold is obtained, and the output of the ultra-short-term-short-term prediction module is required to have a cumulative effect within the deviation threshold.

9. The data-driven computing center load prediction method of claim 8, wherein, In step S3, the parameters of the long-term-Mid-long-term evolution prediction model are dynamically calibrated, specifically: The cumulative deviation of the short-term prediction is monitored, and when the deviation continues to exceed the deviation threshold, the calibration process is triggered; Before calibration, deviation traceability analysis is performed to decompose the total load deviation into the contribution of rack quantity prediction deviation and single-rack power prediction deviation, as follows: ， Where ∆L is the total load deviation, Na and Pa are the actual rack quantity and single-rack power calculated from recent actual data, and Nf and Pf are the rack quantity prediction value and single-rack power prediction value for the corresponding period; According to the traceability analysis results, the parameters of the long-term-Mid-long-term evolution prediction model are selectively recalibrated.

10. A data-driven computing center load prediction system for implementing the data-driven computing center load prediction method of any one of claims 1-9, characterized in that, The system comprises: A data acquisition and preprocessing module for acquiring internal operating parameters and historical data of the computing power center, and preprocessing the acquired data; An ultra-short-term-short-term prediction module for predicting the baseline load and adjustable margin of the computing power center in the future minutes to days based on the internal operating parameters and historical load data of the computing power center through a deep neural network model; A long-term-Mid-long-term prediction module for predicting the future rack quantity using a saturation-adaptive growth rate decreasing model based on historical and planning information of rack quantity, and predicting the future single-rack average power consumption using a linear regression model based on historical and development trend of single-rack power consumption, and then calculating the total power prediction in the future long term; A coupling feedback and dynamic calibration module for using the output of the long-term-Mid-long-term prediction module as a baseline to constrain the prediction range of the ultra-short-term-short-term prediction module, and using the deviation between the cumulative actual operating data and predicted data of the ultra-short-term-short-term prediction module as a feedback signal to dynamically calibrate the parameters of the long-term-Mid-long-term evolution prediction model; A prediction result output and visualization module for outputting and displaying the prediction results.

Images