Intelligent management and control system for dynamic energy efficiency optimization of large-computing-power data center

Abstract

The application discloses a kind of big computing power data center dynamic energy efficiency optimization intelligent management and control systems, including data acquisition layer, edge pre-processing layer, global energy efficiency optimization decision layer and closed-loop execution layer;Data acquisition layer collects the running raw data of energy efficiency associated equipment;Edge pre-processing layer is excluded to the received running raw data, time series registration and normalization processing are carried out, and standardized time series data set is screened and formed Transmission to global energy efficiency optimization decision layer;Global energy efficiency optimization decision layer is based on standardized time series data set and two types of predicted time series sequence completes multi-device energy efficiency quantitative calculation;Closed-loop execution layer receives control instruction and carries out risk check, and the instruction that passes through check is sent to corresponding energy efficiency associated equipment.The application realizes the global collaborative control of computing power scheduling and power supply and distribution, refrigeration facilities, can significantly improve the overall energy efficiency level of big computing power data center, reduce idle energy consumption and refrigeration energy consumption, effectively avoid the operation risk caused by unreasonable control.

Description

Technical Field

[0001] This invention relates to the field of energy efficiency management and control technology for high-performance data centers, and in particular to a dynamic energy efficiency optimization and intelligent management and control system for high-performance data centers. Background Technology

[0002] High-performance computing data centers are the core digital infrastructure supporting artificial intelligence training and inference, cloud computing services, and industrial internet computing. In recent years, with the rapid development of general artificial intelligence and high-performance computing industries, the construction scale of high-performance computing data centers has continued to expand, and their total energy consumption accounts for a growing proportion of the total energy consumption of the whole society. Energy efficiency optimization has become a core requirement in the operation of high-performance computing data centers, and the industry urgently needs intelligent energy efficiency management solutions that can adapt to heterogeneous computing clusters and dynamically adjust according to operating conditions.

[0003] Currently, mainstream energy efficiency management solutions for high-performance computing data centers fall into two categories. The first is the static threshold-based management solution. Its core working principle involves operators pre-setting fixed cooling temperatures and equipment operating threshold rules, such as fixing the precision air conditioning supply temperature to 22 degrees Celsius and keeping all computing nodes online year-round. The infrastructure management system monitors the equipment's operating status, and manual adjustments are made after alarms are triggered. This solution has low deployment costs, simple operating logic, and no system conflict risks, and is currently widely used in traditional data centers with smaller scales and mild load fluctuations. However, it is difficult to adapt to the large load fluctuations of high-performance computing data centers, which can easily lead to long-term mismatch between cooling supply and computing heat load, high idle energy consumption of computing nodes, and limited space for energy efficiency optimization.

[0004] The second type is a single-dimensional energy efficiency management solution. Its core working principle involves deploying an optimization system separately for either computing power scheduling or cooling control. The computing power scheduling system collects cluster load data independently to optimize task allocation, while the cooling control system collects ambient temperature data independently to adjust cooling output. These two systems operate independently without data interaction. This solution can achieve a certain degree of energy efficiency improvement in a single dimension and has been implemented in most medium-to-large data centers. However, this solution fails to achieve coordinated control between heterogeneous computing power clusters and power supply and cooling infrastructure. Changes in heat load caused by computing power scheduling cannot be transmitted to the cooling end in real time, resulting in insufficient timeliness and accuracy in cooling adjustment. Furthermore, the management model only adapts to a single data center hardware architecture, exhibiting weak generalization and adaptability, making it difficult to meet the global energy efficiency management needs of large-scale data centers with multiple architectures and business models.

[0005] In summary, neither of the above two types of control solutions is suitable for the operational characteristics of large-scale computing data centers, which have a high proportion of heterogeneous computing power, large load fluctuations, and high energy efficiency optimization requirements. They cannot achieve optimal overall energy efficiency under the constraints of ensuring business service quality and equipment operation safety. This has become the core bottleneck restricting cost reduction and efficiency improvement of large-scale computing data centers. The industry urgently needs a more intelligent global energy efficiency control solution. Summary of the Invention

[0006] The present invention proposes a dynamic energy efficiency optimization intelligent management and control system for high-computing-power data centers to solve the problems mentioned in the prior art.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: The intelligent management and control system for dynamic energy efficiency optimization of high-computing-power data centers includes a data acquisition layer, an edge preprocessing layer, a global energy efficiency optimization decision layer, and a closed-loop execution layer; The data acquisition layer interfaces with energy efficiency-related devices to collect raw operational data from these devices, which include heterogeneous computing clusters, power supply and distribution equipment, refrigeration equipment, and environmental sensing terminals. The input of the edge preprocessing layer is connected to the output of the data acquisition layer. The edge preprocessing layer performs anomaly removal, time series registration and normalization on the received raw running data, filters and forms a standardized time series dataset and transmits it to the global energy efficiency optimization decision layer. The global energy efficiency optimization decision layer includes a time-series prediction unit, a computing power scheduling unit, a cooling capacity regulation unit, and a dynamic energy efficiency evaluation model. The time-series prediction unit receives a standardized time-series dataset and generates a computing power operation prediction time-series sequence and a power distribution cooling energy consumption prediction time-series sequence, respectively. The dynamic energy efficiency evaluation model completes the energy efficiency quantification calculation of multiple devices based on the standardized time-series dataset and the two types of prediction time-series sequences. The computing power scheduling unit and the cooling capacity regulation unit output control commands based on the calculation results. The input of the closed-loop execution layer is connected to the output of the global energy efficiency optimization decision layer. The closed-loop execution layer receives control commands and performs risk verification. Commands that pass the verification are sent to the corresponding energy efficiency related devices. At the same time, the operation feedback data of the devices is collected and transmitted to the global energy efficiency optimization decision layer to update the parameters of the dynamic energy efficiency assessment model.

[0008] Furthermore, the raw operational data collected by the data acquisition layer includes: real-time load, operating temperature, real-time power consumption, and task queue length of each node in the heterogeneous computing power cluster; input power, output power, and real-time PUE values ​​of power supply and distribution equipment; supply air temperature, return air temperature, cooling capacity distribution ratio, and operating power of precision air conditioners of cooling equipment; and ambient temperature and humidity parameters of each partition of the data center.

[0009] Furthermore, the edge preprocessing layer pre-stores device judgment thresholds. The edge preprocessing layer is also used to compare the data after anomaly removal, time-series registration, and normalization with the device judgment thresholds to identify overheating and power supply abnormality faults of computing nodes and generate fault alarm signals simultaneously. It also sends temporary safety control commands to the energy efficiency-related devices corresponding to the faults. At the same time, it filters out data from the processed data that has no missing content, continuous timestamps, and whose values ​​do not exceed the rated operating range of the devices and do not have distortion jumps, forming a standardized time-series dataset that is sent to the global energy efficiency optimization decision layer.

[0010] Furthermore, the time-series prediction unit generates a computing power operation prediction time-series sequence corresponding to the heterogeneous computing power cluster within the next hour based on a standardized time-series dataset. The time-series prediction unit also generates a power distribution and cooling energy consumption prediction time-series sequence corresponding to the power supply and distribution equipment and cooling equipment within the next 24 hours based on the standardized time-series dataset, externally acquired temperature information, and preset power grid time period identifiers. The computing power operation prediction time-series sequence and the power distribution and cooling energy consumption prediction time-series sequence are sent to the dynamic energy efficiency assessment model for multi-device energy efficiency quantification calculation.

[0011] Furthermore, the computing power scheduling unit of the global energy efficiency optimization decision layer receives the multi-device energy efficiency quantification calculation results output by the dynamic energy efficiency assessment model, and allocates computing power tasks to be allocated according to the calculation results. The computing power tasks are allocated to the heterogeneous computing power nodes with the best multi-device energy efficiency quantification calculation results. At the same time, for computing power nodes whose idle time exceeds the preset threshold, corresponding control instructions for hibernation and shutdown are generated. The control instructions are output from the global energy efficiency optimization decision layer to the closed-loop execution layer.

[0012] Furthermore, the cooling capacity regulation unit has a built-in infrastructure linkage optimization module. Based on the energy consumption change trend and the temperature distribution data of the computing power nodes, the infrastructure linkage optimization module dynamically adjusts the air supply speed, air supply angle, and cold aisle cooling capacity distribution parameters of the precision air conditioner, generates cooling control commands, and outputs them to the closed-loop execution layer.

[0013] Furthermore, the closed-loop execution layer has a built-in instruction risk verification module. Before issuing control instructions, the instruction risk verification module verifies the execution risk of the control instructions and eliminates abnormal instructions that may cause overheating of computing nodes or power overload. After the instruction is executed, energy efficiency change data and equipment operation status data are collected and sent back to the global energy efficiency optimization decision layer to iteratively correct the weight parameters of the dynamic energy efficiency assessment model.

[0014] Furthermore, the intelligent control system for dynamic energy efficiency optimization of high-performance data centers also includes a digital twin simulation and deduction unit that communicates with the global energy efficiency optimization decision-making layer. The digital twin simulation and deduction unit has a built-in full-element digital twin model of the high-performance data center. After the global energy efficiency optimization decision-making layer generates control instructions to be verified, it first pushes them to the digital twin simulation and deduction unit. The digital twin simulation and deduction unit then performs a pre-simulation of the execution effect of the control instructions, assesses the energy efficiency improvement and potential operational risks. When the assessment results meet the preset threshold requirements, the global energy efficiency optimization decision-making layer then pushes the corresponding control instructions to the closed-loop execution layer.

[0015] Furthermore, the dynamic energy efficiency optimization and intelligent management and control system for high-performance data centers also includes an energy efficiency visualization unit that communicates with the global energy efficiency optimization decision-making layer. The energy efficiency visualization unit connects to the data center operation and management platform, receives various energy efficiency statistics output by the global energy efficiency optimization decision-making layer, and displays the real-time PUE, computing power energy efficiency ratio, cumulative energy savings, and control command execution records of the entire data center in real time.

[0016] Compared with existing technologies, the beneficial effects of this invention are: The system collects raw operational data from various energy-efficient related devices through a unified data acquisition layer. An edge preprocessing layer then performs anomaly removal, time-series registration, and normalization to obtain a standardized time-series dataset. Finally, the global energy efficiency optimization decision layer's time-series prediction unit generates computing power operation prediction time-series sequences and power distribution cooling energy consumption prediction time-series sequences, replacing the existing static threshold-based control and manual post-event adjustments. This system allows for adjustments based on real-time operational status and load change trends, adapting to the large load fluctuations of high-performance data centers. It improves upon the mismatch between cooling capacity and computing power heat load, and the high idle energy consumption of computing power under static control, addressing the shortcomings of insufficient adaptability and lagging control in static solutions.

[0017] The dynamic energy efficiency assessment model integrates standardized time-series datasets and two types of predicted time-series sequences to conduct quantitative calculations of energy efficiency for multiple devices. The computing power scheduling unit and the cooling capacity regulation unit output control commands based on the same calculation results, overcoming the problems of independent operation and data incompatibility between existing computing power and refrigeration subsystems. Load and heat load changes caused by computing power scheduling can be synchronously fed back to the cooling capacity regulation link, realizing coordinated regulation among heterogeneous computing power clusters, power supply and distribution equipment, and refrigeration equipment. This solves the problems of lagging cooling capacity regulation and difficulty in achieving global energy efficiency optimization under single-dimensional optimization.

[0018] The closed-loop execution layer performs risk verification on control commands before issuing them to energy-efficient related devices, taking device operational safety as a constraint prerequisite for energy efficiency optimization. Simultaneously, it collects operational feedback data and transmits it back to the global energy efficiency optimization decision-making layer, continuously updating the parameters of the dynamic energy efficiency assessment model. This approach avoids equipment operational risks caused by energy efficiency control while allowing the energy efficiency calculation logic to continuously adjust according to actual operating conditions. It overcomes the shortcomings of existing solutions in simultaneously addressing business stability, device safety, and long-term optimization adaptability, meeting the global energy efficiency optimization needs of high-performance data centers in heterogeneous, high-load-fluctuation scenarios. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the architecture of a high-performance data center dynamic energy efficiency optimization intelligent management and control system according to an embodiment of this application. Detailed Implementation

[0020] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] This invention provides a dynamic energy efficiency optimization and intelligent management and control system for high-computing-power data centers, such as... Figure 1 As shown, the system 100 includes a data acquisition layer 110, an edge preprocessing layer 120, a global energy efficiency optimization decision layer 130, and a closed-loop execution layer 140.

[0022] The data acquisition layer 110 interfaces with energy efficiency-related devices to collect raw operational data from these devices, which include heterogeneous computing clusters, power supply and distribution equipment, refrigeration equipment, and environmental sensing terminals.

[0023] The heterogeneous computing cluster includes computing nodes with various architectures such as GPUs, CPUs, and FPGAs, undertaking AI training, inference, and high-performance computing tasks. Power supply and distribution equipment includes high-voltage switchgear, UPS, precision power distribution busbars, PDUs, and other power supply and distribution equipment. Cooling equipment includes in-row precision air conditioners, chillers, cooling towers, air-cooled units, and other cooling infrastructure. Environmental sensing terminals include temperature and humidity sensors, differential pressure sensors, and thermal imaging acquisition terminals, deployed in cold aisles, hot aisles, and inside cabinets to collect cabinet microenvironment parameters.

[0024] The raw data is real-time raw observation data directly collected by the data acquisition layer 110 from various energy efficiency-related devices without cleaning or standardization processing.

[0025] The input of the edge preprocessing layer 120 is connected to the output of the data acquisition layer 110. The edge preprocessing layer 120 performs anomaly removal, time series registration and normalization on the received raw running data, filters and forms a standardized time series dataset and transmits it to the global energy efficiency optimization decision layer 130.

[0026] The process includes several key steps: Anomaly removal identifies and removes broken or missing values, sensor jumps, and distorted data from equipment malfunctions in the raw data, distinguishing between random noise and persistent distortion. Time-series registration addresses the issue of asynchronous timestamps in data collected from different energy-efficient devices. The edge preprocessing layer 120 unifies the time reference, aligning the time axis of the dispersed data at a uniform time granularity to ensure time-series comparability across different devices. Normalization maps raw physical quantities such as power, temperature, and load rate, with different dimensions and numerical ranges, to a unified numerical range, eliminating dimensional differences and facilitating subsequent calculations by the time-series prediction unit and dynamic energy efficiency assessment model. After anomaly removal, time-series registration, and normalization, a standardized time-series dataset is obtained. This dataset meets the requirements of continuous timestamps, no missing values, values ​​within the normal operating range of the equipment, and no distortion jumps. It serves as the unified input data source for the global energy efficiency optimization decision layer 130 for prediction and multi-device energy efficiency quantification.

[0027] The global energy efficiency optimization decision layer 130 includes a time-series prediction unit, a computing power scheduling unit, a cooling capacity regulation unit, and a dynamic energy efficiency evaluation model. The time-series prediction unit receives a standardized time-series dataset and generates a computing power operation prediction time-series sequence and a power distribution cooling energy consumption prediction time-series sequence, respectively. The dynamic energy efficiency evaluation model completes the energy efficiency quantification calculation of multiple devices based on the standardized time-series dataset and the two types of prediction time-series sequences. The computing power scheduling unit and the cooling capacity regulation unit output control commands according to the calculation results.

[0028] The computing power operation prediction time series forecasts future load changes, task occupancy, and computing node heating trends in heterogeneous computing power clusters, supporting the prediction of heat load changes. The power distribution and cooling energy consumption prediction time series combines standardized time series datasets, external meteorological data, and power grid time period information to predict power supply and distribution losses and expected energy consumption of the cooling system. Both types of prediction series enable upfront load prediction, avoiding lagging adjustments based solely on current instantaneous data.

[0029] The dynamic energy efficiency assessment model integrates current real-time standardized time-series data with future predicted time-series data, and is no longer limited to instantaneous operating conditions. The quantitative calculation takes into account the computing power output of the computing cluster, power supply and distribution losses, cooling energy consumption, and equipment temperature safety boundaries, and outputs a comprehensive energy efficiency index to characterize the overall data center energy efficiency level under different control schemes.

[0030] The computing power scheduling unit optimizes and controls computing resources based on energy efficiency calculations; the cooling capacity regulation unit adjusts the operating parameters and cooling capacity allocation of cooling equipment based on energy efficiency calculations, optimizing and controlling the cooling capacity of infrastructure. Both units collaboratively output unified control commands, achieving coordinated control of computing power and cooling, overcoming the limitations of existing technologies that rely on single-dimensional, independent control.

[0031] The input of the closed-loop execution layer 140 is connected to the output of the global energy efficiency optimization decision layer 130. The closed-loop execution layer 140 receives control commands and performs risk verification. It sends commands that pass the verification to the corresponding energy efficiency related devices. At the same time, it collects device operation feedback data and transmits it to the global energy efficiency optimization decision layer 130 to update the parameters of the dynamic energy efficiency assessment model.

[0032] After receiving the control command, the closed-loop execution layer 140 simulates and predicts the consequences of command execution, and verifies whether there is a safety risk of equipment operating beyond limits. It filters high-risk abnormal control commands and only allows control commands that meet the equipment safety constraints to be issued to energy efficiency-related equipment. After the control command is issued and executed, it collects the actual operating status, power consumption, temperature, and load change data of each energy efficiency-related device after execution, and sends them back to the global energy efficiency optimization decision layer 130. The global energy efficiency optimization decision layer 130 uses the returned operating feedback data to continuously update the parameters of the dynamic energy efficiency assessment model, forming a complete closed-loop optimization link of data acquisition → edge preprocessing → energy efficiency decision → command issuance and execution → feedback transmission → model iteration.

[0033] In this invention, the raw operational data collected by the data acquisition layer 110 includes: real-time load, operating temperature, real-time power consumption, and task queue length of each node in the heterogeneous computing power cluster; input power, output power, and real-time PUE value of the power supply and distribution equipment; supply air temperature, return air temperature, cooling capacity allocation ratio, and operating power of the precision air conditioner of the cooling equipment; and ambient temperature and humidity parameters of each partition of the data center.

[0034] Real-time load characterizes the current computing resource utilization of a computing node; operating temperature characterizes the hardware operating temperature of the computing node; real-time power consumption characterizes the instantaneous power consumption of the computing node; and task queue length characterizes the number of computing tasks waiting to be allocated to the corresponding computing node. These parameters collectively reflect the service load, heat generation status, and energy consumption level of the heterogeneous computing cluster.

[0035] Input power is the total electrical power input upstream of the power supply and distribution equipment; output power is the electrical power delivered by the power supply and distribution equipment to the downstream load. The difference between input power and output power reflects the power loss of the power distribution system; real-time PUE value represents the ratio of the total energy consumption of the data center to the energy consumption of IT computing equipment, which is the core indicator for measuring the energy efficiency of the data center.

[0036] Supply air temperature is the temperature of the cold air supplied by the refrigeration equipment to the cold aisle; return air temperature is the temperature of the hot air recovered by the refrigeration equipment from the hot aisle. The difference between supply and return air temperatures can reflect the heat exchange load of the refrigeration system; the cold capacity distribution ratio represents the distribution ratio of the cold capacity output by the refrigeration equipment in each area; the precision air conditioner operating power represents the energy consumption of the refrigeration equipment.

[0037] The ambient temperature and humidity parameters of each partition in the data center are the ambient temperature and humidity collected in different physical partitions of the computer room. They reflect the micro-environmental conditions around the server racks and can help determine the heat dissipation conditions of computing equipment.

[0038] This implementation method simultaneously collects raw operational data from multiple dimensions, including computing power, power supply and distribution, cooling, and data center environment. This can completely restore the actual operating conditions of the data center, providing sufficient data support for subsequent data preprocessing, time series prediction, and multi-device energy efficiency quantification. It avoids deviations in operating condition assessment caused by single-dimensional data and ensures the accuracy of the collaborative control strategy for computing power and cooling.

[0039] In this invention, the edge preprocessing layer 120 pre-stores device judgment thresholds. The edge preprocessing layer 120 is also used to compare the data after anomaly removal, time registration, and normalization with the device judgment thresholds, identify overheating and power supply abnormality faults of computing nodes, and generate fault alarm signals simultaneously, and send temporary safety control commands to the energy efficiency related devices corresponding to the faults. At the same time, it filters out data from the processed data that has no missing content, continuous timestamps, and whose values ​​do not exceed the rated operating range of the devices and have no distortion jumps, forming a standardized time series dataset that is sent to the global energy efficiency optimization decision layer 130.

[0040] The device judgment thresholds are parameter limits pre-stored within the edge preprocessing layer 120, corresponding to the safe operation boundaries of various energy-efficient related devices. These limits cover hardware safety boundary parameters such as the upper limit of computing node temperature and the power or current limit of power supply circuits. The device judgment thresholds can be configured in advance based on the device hardware specifications and data center operation and maintenance standards.

[0041] The edge preprocessing layer 120 compares the processed time-series parameters with the corresponding device's preset safety limits one by one. When the collected parameters exceed the safety range corresponding to the device's judgment threshold, it determines that the device has experienced operational failures such as computing node overheating or power supply abnormalities. This fault identification is completed locally at the edge, without needing to upload the raw operational data to the global energy efficiency optimization decision layer 130.

[0042] The fault alarm signal is a prompt signal generated synchronously after the edge preprocessing layer 120 identifies the equipment over-limit fault. It can be pushed to the operation and maintenance management platform so that operation and maintenance personnel can keep abreast of abnormal operating conditions of the equipment on site.

[0043] Temporary safety control commands are emergency protection commands generated locally by the edge preprocessing layer 120 after identifying a fault. They are then sent to the energy efficiency-related equipment that has malfunctioned to enable local emergency safety handling.

[0044] No missing content indicates that there are no missing data collection points in the time series segment; continuous timestamps indicate that the data has a unified and coherent time series benchmark; all values ​​not exceeding the rated operating range of the equipment indicate that the parameters are within the safe operating range of the equipment; no distortion jumps indicate that there is no abnormal data collection with sudden distortions. Data that meets all the screening criteria are aggregated to form a standardized time series dataset, and transmitted upwards to the global energy efficiency optimization decision layer 130, serving as a unified input data source for time series prediction units and dynamic energy efficiency assessment models to carry out time series prediction and multi-device energy efficiency quantification calculations.

[0045] In this embodiment, the edge preprocessing layer 120 locally relies on the device's judgment threshold to complete fault identification and issue temporary safety control commands, thereby realizing on-site fault response and shortening the delay of device anomaly handling. Simultaneously, based on multiple constraints, valid data is screened to generate a standardized time series dataset, avoiding the upward transmission of unqualified data that interferes with upper-layer time series prediction and multi-device energy efficiency quantification calculation, and ensuring the accuracy of the system's overall energy efficiency decision results.

[0046] This invention also discloses three types of local preprocessing operations in the edge preprocessing layer 120: anomaly removal, temporal registration, and normalization. Specific implementation methods are as follows: First, the 3σ criterion is used to remove abnormal jump values ​​in the raw data and fill in missing bits. This is done for a time series sequence {x1, x2, ..., x...} with a specific parameter. n First, calculate the sequence mean μ and standard deviation σ. If a data point xᵢ satisfies |xᵢ|σ, then... i If -μ|>3σ, it is considered an outlier; only when a single data point is missing, the arithmetic mean of the two adjacent valid values ​​before and after the data point is used to fill the gap; if multiple consecutive data points are missing, the data segment is directly marked as missing, and no interpolation is performed.

[0047] Next, time registration is performed: multi-source data with different sampling frequencies are uniformly aligned to a time reference with a 10-second granularity. For parameters with a sampling period of less than 10 seconds (sampling frequency higher than 1 time / 10 seconds), downsampling alignment is completed by taking the average value of the 10-second time interval; for parameters with a sampling period of more than 10 seconds (sampling frequency lower than 1 time / 10 seconds), linear interpolation is used to fill in the sampling points and uniformly align them to a 10-second time granularity.

[0048] Then perform normalization: map all numerical parameters with different dimensions to the range of 0 to 1 using the normalization formula, which is: ; Where x is the original value. , These are the minimum and maximum historical values ​​of the parameter, respectively, as pre-calculated. If a parameter shows... = If the value is constant, the normalization result will be uniformly assigned a value of 0.5; normalization can reduce the convergence difficulty of subsequent time series prediction units and dynamic energy efficiency assessment models.

[0049] The edge preprocessing layer 120 can have a built-in local emergency rule base. The rule base is pre-written with emergency response logic that conforms to the data center operation specifications. When a minor fault that can be handled locally, such as overheating of a local computing node or overload of a power supply circuit, is detected, a preliminary emergency response instruction to reduce frequency and cut off non-core loads is directly issued. This emergency response instruction is independent of the control instruction output by the global energy efficiency optimization decision layer 130 and does not need to wait for upper-level decision, thus reducing fault response latency. Only the standardized time-series dataset that has been preprocessed and has been attached with a CRC32 data check code is uploaded to the global energy efficiency optimization decision layer 130. If the receiving end verifies and identifies a transmission error, the data segment is marked as abnormal and is not included in the subsequent energy efficiency calculation process to avoid deviations in the dynamic energy efficiency assessment model parameter iteration caused by transmission errors.

[0050] In this invention, the time series prediction unit generates a computing power operation prediction time series sequence corresponding to the heterogeneous computing power cluster within the next hour based on a standardized time series dataset. The time series prediction unit also generates a power distribution and cooling energy consumption prediction time series corresponding to the power supply and distribution equipment and cooling equipment within the next 24 hours based on the standardized time series dataset, externally acquired temperature information, and preset power grid time period identifiers. The computing power operation prediction time series sequence and the power distribution and cooling energy consumption prediction time series sequence are sent to the dynamic energy efficiency evaluation model for multi-device energy efficiency quantification calculation.

[0051] The time-series prediction unit sets different prediction windows for two types of prediction sequences: the computing power operation prediction time-series sequence adopts a short prediction window of one hour in the future, which is adapted to the characteristics of rapid fluctuations in the business load of heterogeneous computing power clusters and meets the real-time fine-grained control requirements of the computing power side; the power distribution and cooling energy consumption prediction time-series sequence adopts a long prediction window of twenty-four hours in the future, which is convenient for coordinating environmental conditions and power consumption periods to carry out medium- and long-term load optimization scheduling.

[0052] The external temperature information refers to the ambient temperature data accessed by the time-series prediction unit from outside the system. The external ambient temperature directly affects the heat exchange load of the refrigeration system; incorporating it into the input conditions can improve the accuracy of the power distribution refrigeration energy consumption prediction results.

[0053] The grid time period identifier is a pre-configured time tag used to distinguish between peak, off-peak, and valley electricity consumption periods. Different grid time periods correspond to different electricity consumption constraints and costs. Introducing this identifier allows energy consumption predictions to align with grid electricity consumption characteristics, supporting the formulation of subsequent energy efficiency optimization strategies.

[0054] The computing power operation prediction time series relies solely on the standardized time series dataset for extrapolation; the power distribution cooling energy consumption prediction time series, in addition to the standardized time series dataset, also incorporates external temperature information and power grid time period identifiers as auxiliary inputs. The two have different input data source compositions to adapt to the influencing factors characteristics of their respective prediction objects.

[0055] In this embodiment, the prediction duration is differentiated for computing power operation and power distribution cooling energy consumption to match the control rhythm of different devices; external temperature and grid time period identifiers are introduced into the power distribution cooling energy consumption prediction to expand the input dimensions and improve the reliability of energy consumption prediction; energy efficiency assessment is carried out in conjunction with short-term computing power prediction and long-term energy consumption prediction to achieve pre-performance efficiency optimization and control.

[0056] In this invention, the computing power scheduling unit of the global energy efficiency optimization decision layer 130 receives the multi-device energy efficiency quantification calculation results output by the dynamic energy efficiency evaluation model, and allocates computing power tasks to be allocated according to the calculation results. The computing power tasks are allocated to the heterogeneous computing power nodes with the best multi-device energy efficiency quantification calculation results. At the same time, for computing power nodes whose idle time exceeds a preset threshold, corresponding control instructions for hibernation and shutdown are generated. The control instructions are output from the global energy efficiency optimization decision layer 130 to the closed-loop execution layer 140.

[0057] The computing power scheduling unit uses the energy efficiency quantification calculation results of multiple devices as the allocation basis, prioritizing the distribution of computing power tasks to be allocated to heterogeneous computing power nodes with the best energy efficiency calculation results. This allocation method can minimize the power consumption and heat generated by computing power operation while fulfilling computing power service requirements, thereby optimizing data center energy efficiency from the perspective of computing power load allocation.

[0058] The computing power scheduling unit continuously monitors the idle time of each heterogeneous computing power node. When the continuous idle time of a node exceeds the preset time threshold, it determines that the node has no continuous computing power service demand and generates control instructions of the hibernation or shutdown type for the node to reduce ineffective energy consumption by shutting down idle computing power hardware.

[0059] The hibernation and shutdown control instructions generated by the computing power scheduling unit are uniformly summarized to the global energy efficiency optimization decision layer 130, and then uniformly output to the closed-loop execution layer 140. The closed-loop execution layer 140 completes the subsequent risk verification and equipment instruction issuance operations.

[0060] In this embodiment, computing power tasks are optimally allocated based on the energy efficiency quantification calculation results of multiple devices, which can achieve power consumption optimization during the computing power service scheduling phase; hibernation and shutdown control instructions are issued to long-term idle computing power nodes to eliminate the ineffective power consumption of idle hardware; control instructions are uniformly transmitted to the outside by the global energy efficiency optimization decision layer 130 to ensure that all energy efficiency optimization scheduling instructions are sent through a unified link, which facilitates the closed-loop execution layer 140 to uniformly complete the security risk verification.

[0061] In this invention, the closed-loop execution layer 140 has a built-in instruction risk verification module. Before issuing control instructions, the instruction risk verification module verifies the execution risk of the control instructions and eliminates abnormal instructions that may cause overheating of computing nodes or power overload. After the instruction is executed, energy efficiency change data and equipment operation status data are collected and sent back to the global energy efficiency optimization decision layer 130 to iteratively correct the weight parameters of the dynamic energy efficiency evaluation model.

[0062] The instruction risk verification module simulates the execution of the control instructions to be issued and deduces the operating conditions of each energy efficiency-related device after the instructions are issued. It focuses on determining whether there will be two types of safety scenarios: excessive temperature of computing nodes and power overload of power supply circuits. If the above risks are predicted, the control instruction is determined to be an abnormal instruction and is directly intercepted and not issued to the corresponding energy efficiency-related device.

[0063] After the instruction is executed, the closed-loop execution layer 140 collects two types of feedback data: one is energy efficiency change data, which reflects the changes in the overall energy efficiency indicators of the data center after the implementation of the control operation; the other is equipment operation status data, which records the actual operating conditions of each heterogeneous computing power cluster, power supply and distribution, and cooling equipment, such as actual temperature, power consumption, and load. The two types of data together constitute complete operation feedback data.

[0064] The closed-loop execution layer 140 transmits the collected operational feedback data back to the global energy efficiency optimization decision layer 130. The global energy efficiency optimization decision layer 130 uses real control implementation effect data to continuously update and adjust the weight parameters inside the dynamic energy efficiency assessment model, so that the model can fit the real operating characteristics of the equipment when conducting energy efficiency quantification calculations for multiple devices, and gradually improve the accuracy of energy efficiency assessment and scheduling optimization.

[0065] Therefore, by using the command risk verification module to intercept control commands with potential security risks in advance, the system balances the overall energy efficiency optimization goals with equipment operation safety, avoiding hardware failures caused by optimization scheduling operations. Based on the real feedback data after command execution, the dynamic energy efficiency evaluation model weights are iteratively corrected to form a continuously self-optimizing closed-loop control link. As the system operates, the energy efficiency calculation deviation is continuously reduced, and the optimization effect of computing power and cooling coordinated control is continuously improved.

[0066] In this invention, the dynamic energy efficiency optimization intelligent management and control system 100 for high-performance data centers also includes a digital twin simulation and deduction unit that is communicatively connected to the global energy efficiency optimization decision layer 130. The digital twin simulation and deduction unit has a built-in full-element digital twin model of the high-performance data center.

[0067] The digital twin simulation and inference unit and the global energy efficiency optimization decision layer 130 use bidirectional real-time message queue communication interaction. The data synchronization refresh cycle can be fixed at 10 seconds. Every time the global energy efficiency optimization decision layer 130 generates a set of control instructions to be verified, it simultaneously pushes the latest standardized time series dataset and the predicted time series sequence of power distribution and cooling energy consumption for the next 24 hours to the digital twin simulation and inference unit.

[0068] After the global energy efficiency optimization decision layer 130 generates the control command to be verified, it is first pushed to the digital twin simulation and deduction unit. The digital twin simulation and deduction unit performs a pre-simulation of the execution effect of the control command, evaluates the energy efficiency improvement and potential operational risks. When the evaluation result meets the preset threshold requirements, the global energy efficiency optimization decision layer 130 pushes the corresponding control command to the closed-loop execution layer 140.

[0069] Among them, the digital twin simulation and deduction unit establishes a two-way data communication link with the global energy efficiency optimization decision layer 130. It does not directly participate in the issuance of equipment commands, but only undertakes the role of simulation and pre-evaluation of the control scheme, as well as the pre-evaluation of the effects and risks, serving as a pre-verification link before the control commands are issued to the closed-loop execution layer 140.

[0070] The full-element digital twin model is a virtual simulation model that replicates all the hardware and operating logic of a high-performance data center. The model fully maps the hardware parameters, operating constraints, and energy consumption characteristics of all energy-related equipment, such as heterogeneous computing power clusters, power supply and distribution equipment, cooling equipment, and environmental sensing terminals. It can virtually reproduce the operating conditions of the entire data center after any control command is implemented.

[0071] The full-element digital twin model continuously receives standardized time-series datasets synchronously pushed by the global energy efficiency optimization decision layer 130, and synchronizes the current real operating conditions of the computer room in real time to ensure that the initial state of the simulation is completely consistent with the state of the equipment on site. The simulation pre-run duration is consistent with the duration of the power distribution and cooling energy consumption prediction time series. The duration of a single complete pre-run can be 24 hours, and the simulation energy efficiency indicators and equipment operating parameters for the corresponding time period are output hourly.

[0072] After the simulation unit completes the pre-simulation, it simultaneously outputs two types of evaluation results: one is the energy efficiency improvement, which compares the comprehensive energy efficiency index of the data center before and after the simulation and quantitatively calculates the energy-saving optimization space that the set of control commands can achieve; the other is the potential operational risks, which virtually predicts whether there will be hidden dangers of equipment exceeding limits such as overheating of computing nodes, power overload, and insufficient cooling capacity after the commands are executed.

[0073] The system is pre-configured with two types of quantitative evaluation threshold standards: the energy efficiency improvement threshold is that a single round of regulation reduces the total energy consumption by at least 2%, and the equipment risk threshold is that no equipment touches the rated operating limit during the entire simulation.

[0074] After the digital twin simulation and deduction unit completes the evaluation, it will send the evaluation results back to the global energy efficiency optimization decision layer 130. Only when both the energy efficiency improvement and operational risk evaluation indicators meet the preset threshold requirements will the global energy efficiency optimization decision layer 130 push the control command to the closed-loop execution layer 140. If the evaluation results do not meet the standards, the control command will be abandoned and the computing power scheduling unit and cooling capacity adjustment unit will regenerate the optimization scheme.

[0075] In this embodiment, a digital twin simulation pre-run is added before the control command is issued to the equipment. This allows for the early verification of optimization schemes in a virtual environment. On the one hand, it filters out inefficient control schemes that do not meet the optimization benefits in advance, thereby improving the overall energy-saving effect. On the other hand, it identifies potential equipment operation risks in advance in the virtual environment. The double-layered closed-loop execution layer 140 command risk verification significantly reduces the probability of actual data center hardware failures. The simulation is carried out based on a full-element digital twin model that completely replicates the characteristics of the data center hardware and a standardized time-series dataset that is synchronized with the field in real time with a 10-second cycle. This ensures that the pre-run conditions are highly consistent with the actual data center operation status, and the simulation evaluation results are true and reliable.

[0076] In this invention, the dynamic energy efficiency optimization intelligent management and control system 100 for high-performance data centers also includes an energy efficiency visualization unit that is communicatively connected to the global energy efficiency optimization decision layer 130. The energy efficiency visualization unit is connected to the data center operation and management platform, receives various energy efficiency statistics output by the global energy efficiency optimization decision layer 130, and displays the real-time PUE, computing power energy efficiency ratio, cumulative energy saving, and control command execution records of the entire data center in real time.

[0077] The energy efficiency visualization unit establishes a stable communication link with the global energy efficiency optimization decision-making layer 130. It is responsible for centrally receiving and parsing various energy efficiency statistics, transforming abstract values ​​into a visual display interface, allowing operations and maintenance personnel to intuitively grasp the overall energy efficiency and operational status of the data center. The energy efficiency visualization unit directly connects to the data center operation and management platform. On the one hand, it can synchronously push the visualization interface and energy efficiency statistics to the operation and management platform, enabling operations and maintenance personnel to view them on a unified platform. On the other hand, it can synchronously read the data center's basic configuration and operation and maintenance log information within the platform, assisting in improving the visualization content.

[0078] Real-time Power Usage Effectiveness (PUE) is the real-time ratio of total energy consumption to computing power equipment energy consumption in a data center. It is a core indicator for evaluating the basic energy efficiency level of a data center. The visualization interface refreshes this value in real time, intuitively reflecting the overall energy consumption of the data center. Computing power efficiency ratio (CPU) represents the scale of computing power that can be generated per unit of electricity consumption. It specifically measures the energy utilization efficiency of heterogeneous computing power clusters and is used to intuitively compare the energy-saving levels of different computing power nodes and different task allocation schemes. Cumulative energy savings represent the total amount of electricity saved by the system after each execution of energy efficiency optimization and control commands, compared to the uncontrolled baseline operating condition. It is used to quantitatively demonstrate the long-term energy-saving optimization results of the entire management and control system. The control command execution record fully stores the information of each control command issued from the global energy efficiency optimization decision layer 130 to the closed-loop execution layer 140, including the command generation time, control object, optimization target, simulation pre-run results, actual execution status, and energy efficiency changes after execution. This facilitates maintenance personnel to trace and review each energy efficiency control operation.

[0079] By setting up an energy efficiency visualization unit to centrally display multi-dimensional core energy efficiency indicators, the complex data center energy consumption and computing power operation data are presented intuitively, reducing the difficulty of data analysis for operation and maintenance personnel; the complete control command execution record is recorded simultaneously, which makes it convenient for operation and maintenance personnel to review and optimize strategies and locate the causes of energy efficiency fluctuations; at the same time, it connects with the data center operation and management platform to realize the integration of energy efficiency data with the data center operation and maintenance system, and improves the operation and maintenance adaptability of the entire control system.

[0080] Example 1

[0081] This embodiment applies to a dedicated high-performance computing data center for AI large-scale model training operated by a telecom operator. This data center deploys 1200 A800 GPU heterogeneous computing nodes, equipped with 12 in-row precision air conditioners and 2 sets of 10MVA UPS power distribution equipment, and continuously undertakes AI large-scale model pre-training and fine-tuning tasks. The existing technology manages the system with a fixed supply air temperature of 22℃ and random allocation of computing tasks. Real-time PUE remains around 1.42 year-round, with idle energy consumption accounting for 12% of total energy consumption, resulting in low overall energy efficiency.

[0082] The data acquisition layer 110 of this system deploys 120 industrial acquisition gateways, which connect to the out-of-band management cards of all GPU nodes, precision air conditioning controllers, UPS monitoring systems, and hot and cold aisle environmental monitoring nodes to collect all raw operating data corresponding to various energy efficiency-related devices. Each rack is equipped with one edge preprocessing gateway, which has a built-in edge preprocessing layer 120 to complete local anomaly removal, time-series registration, and normalization data preprocessing operations. At the same time, it completes minor fault identification and temporary safety control command issuance and handling based on pre-stored device judgment thresholds.

[0083] The system's core global energy efficiency optimization decision layer 130 incorporates a time-series prediction unit, employing a dual spatiotemporal granularity prediction logic: it uses a 5-minute base granularity to generate a predicted time-series sequence of computing power operation for the next hour, and uses a 1-hour base granularity combined with external temperature and power grid time period identifiers to generate a predicted time-series sequence of power distribution and cooling energy consumption for the next 24 hours. The internal computing power scheduling unit of the decision layer receives the multi-device energy efficiency quantification calculation results output by the dynamic energy efficiency evaluation model, prioritizing the allocation of training tasks to GPU heterogeneous computing power nodes with the top 40% energy efficiency ratio, and targeting nodes with continuous idle time exceeding [a certain threshold]. The computing nodes with a 15-minute threshold generate control commands of the hibernation or shutdown type; the decision-making layer supports the infrastructure linkage optimization logic, adjusts the cooling distribution of each area based on the thermal cloud map data, raises the air supply temperature of the computer room to 24℃, and increases the cooling supply only in areas with concentrated heat load; the control commands to be verified generated by the global energy efficiency optimization decision layer 130 need to be pushed to the digital twin simulation and deduction unit to complete the simulation pre-run, evaluate the energy efficiency improvement and potential operational risks, and after the simulation evaluation results meet the preset threshold requirements, the control commands are then output to the closed-loop execution layer 140.

[0084] Under the existing random scheduling mode, uneven task allocation leads to some GPU heterogeneous computing nodes operating at low load for a long time, resulting in a high proportion of idle energy consumption. The cooling control method with fixed air supply temperature cannot match the real-time computing power heat load, resulting in a large amount of wasted cooling capacity. After applying the intelligent management and control system for dynamic energy efficiency optimization of large computing power data centers proposed in this application, the real-time PUE of the data center dropped to below 1.26, the proportion of idle energy consumption dropped to less than 3%, and the overall energy efficiency was improved by more than 11%.

[0085] Example 2

[0086] This embodiment is applied to a multi-service hybrid deployment public cloud high-computing-power data center of a cloud vendor. The data center deploys 4,000 heterogeneous computing nodes, including three types of hardware devices: CPU, GPU, and FPGA, and supports multiple services such as e-commerce, video transcoding, AI inference, and cloud server leasing. The peak-valley load difference of the data center exceeds 60%. The existing technical management mode is to keep all devices fully operational during peak hours and maintain 80% of devices online during off-peak hours. The idle energy consumption of the devices is seriously wasted. Moreover, the energy efficiency requirements of different services for computing power and cooling vary greatly. The original management mode cannot achieve differentiated energy efficiency optimization and control.

[0087] This system incorporates a dynamic energy efficiency assessment model, which has been pre-trained through horizontal federated learning and fine-tuned specifically for the mixed business load scenario in this embodiment. The global energy efficiency optimization decision layer's internal time-series prediction unit performs full-time load simulation, generating two types of predicted time-series sequences corresponding to the peak and valley changes in business over the next 24 hours. The computing power scheduling unit allocates computing resources differently based on the time-series prediction results: during peak electricity consumption periods, dormant computing nodes are woken up one hour in advance to reserve business carrying resources; during off-peak electricity consumption periods, low-priority businesses are migrated to high-energy-efficiency heterogeneous computing with the best energy efficiency quantification results from multiple devices. The system generates control commands to shut down 30% of redundant and idle computing power nodes. Infrastructure linkage optimization logic dynamically adjusts cooling supply based on peak and off-peak loads, increasing cooling output during peak hours and decreasing it during off-peak hours to match real-time heat load demands. The system is equipped with an energy efficiency visualization unit, which establishes a communication connection with the global energy efficiency optimization decision layer 130 and connects to existing cloud vendor operation and management platforms to provide operation and maintenance personnel with real-time energy efficiency statistics for each business line, including real-time PUE, computing power efficiency ratio, cumulative energy savings, and control command execution records.

[0088] Under the existing technology management model, the drastic fluctuations in peak and valley loads make it impossible to coordinate the control of computing power and cooling equipment. A large number of computing power devices operate idle for a long time, and the cooling capacity supply does not match the real-time computing power heat load. These two factors result in a large amount of ineffective energy consumption and waste. After applying the intelligent management and control system for dynamic energy efficiency optimization of large computing power data centers in this application, the overall energy consumption of the data center during the off-peak hours has decreased by more than 22%, and the real-time PUE has been stably maintained below 1.28 throughout the day. It can perfectly adapt to the differentiated energy efficiency optimization and control needs of multi-service mixed deployment scenarios.

[0089] The core optimization objective of the dynamic energy efficiency assessment model of this system is to minimize the total energy consumption of the data center. The optimization objective function is as follows: ; ; ; ; ; in, Total energy consumption of the data center, in kWh; This represents the total energy consumption of the heterogeneous computing cluster, in kWh. Total energy consumption of power supply and distribution equipment, in kWh; Total energy consumption of refrigeration equipment, in kWh; Let be the operating core temperature of the i-th computing node. The maximum allowable operating temperature for computing nodes. It is the set of all computing power nodes; Let j be the load factor of the j-th distribution circuit. This represents the maximum allowable load rate of the power distribution circuit. It is the collection of all power distribution circuits; For the first Service level achievement rate for each business task The minimum service level achievement rate required for the task. This is the set of all tasks to be executed.

[0090] The objective function is embedded in the multi-device energy efficiency quantification calculation process of the global energy efficiency optimization decision layer 130, and the standardized time series dataset, computing power operation prediction time series, and power distribution cooling energy consumption prediction time series are used as model input parameters to carry out the solution calculation.

[0091] The model employs a constrained gradient descent optimization algorithm to solve for the optimal solution of the objective function. The iteration step size is set to a fixed 0.01, and the convergence criterion is the total energy consumption after three consecutive iterations. The iteration stops when the change is less than 0.005 kWh.

[0092] The constraints correspond to three types of control boundaries: the temperature safety limit of computing nodes, the load safety limit of power distribution circuits, and the minimum requirements for service quality. After solving, the optimal energy efficiency quantitative calculation results corresponding to each group of computing power allocation and cooling capacity allocation are output.

[0093] The dynamic energy efficiency assessment model weight parameter iteration correction is triggered when the closed-loop execution layer 140 completes a full round of control command execution and sends back all equipment operation feedback data. Each iteration retrains and updates the model's internal weights with the latest collected energy efficiency change data and equipment operation status data.

[0094] The energy efficiency quantification results obtained from the solution will serve as the calculation basis for the computing power scheduling unit and cooling capacity adjustment unit to generate control commands. Under the premise of ensuring the safe operation of energy efficiency-related equipment and the quality of business services, the objective function aims to minimize the energy consumption of the entire system, adapting to the needs of large computing power data centers to carry out global energy efficiency optimization based on this control system. At the same time, the optimization scheme output by the solution also needs to be pushed to the digital twin simulation and deduction unit to complete the execution effect pre-verification. Only after the verification meets the standards can the formal control command be generated and sent to the closed-loop execution layer 140.

[0095] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A dynamic energy efficiency optimization intelligent management and control system for high-computing-power data centers, characterized in that, It includes a data acquisition layer, an edge preprocessing layer, a global energy efficiency optimization decision layer, and a closed-loop execution layer; The data acquisition layer interfaces with energy efficiency-related devices to collect raw operational data from these devices. These energy efficiency-related devices include heterogeneous computing clusters, power supply and distribution equipment, refrigeration equipment, and environmental sensing terminals. The input of the edge preprocessing layer is connected to the output of the data acquisition layer. The edge preprocessing layer performs anomaly removal, time series registration and normalization on the received raw running data, filters and forms a standardized time series dataset and transmits it to the global energy efficiency optimization decision layer. The global energy efficiency optimization decision layer includes a time-series prediction unit, a computing power scheduling unit, a cooling capacity adjustment unit, and a dynamic energy efficiency evaluation model. The time-series prediction unit receives the standardized time-series dataset and generates a computing power operation prediction time-series sequence and a power distribution cooling energy consumption prediction time-series sequence, respectively. The dynamic energy efficiency evaluation model completes the energy efficiency quantification calculation of multiple devices based on the standardized time-series dataset and the two types of prediction time-series sequences. The computing power scheduling unit and the cooling capacity adjustment unit output control commands according to the calculation results. The input of the closed-loop execution layer is connected to the output of the global energy efficiency optimization decision layer. The closed-loop execution layer receives the control command and performs risk verification. It sends the verified command to the corresponding energy efficiency related device and collects the device operation feedback data to the global energy efficiency optimization decision layer to update the parameters of the dynamic energy efficiency assessment model.

2. The intelligent control system for dynamic energy efficiency optimization of high-computing-power data centers according to claim 1, characterized in that, The raw operational data collected by the data acquisition layer includes: real-time load, operating temperature, real-time power consumption, and task queue length of each node in the heterogeneous computing power cluster; input power, output power, and real-time PUE values ​​of power supply and distribution equipment; supply air temperature, return air temperature, cooling capacity allocation ratio, and operating power of precision air conditioners of cooling equipment; and environmental temperature and humidity parameters of each partition of the data center.

3. The intelligent control system for dynamic energy efficiency optimization of high-computing-power data centers according to claim 1, characterized in that, The edge preprocessing layer stores device judgment thresholds. It is also used to compare the data after anomaly removal, time-series registration, and normalization with the device judgment thresholds to identify overheating and power supply abnormality faults of computing nodes and generate fault alarm signals simultaneously. It also sends temporary safety control commands to the energy efficiency-related devices corresponding to the faults. At the same time, it filters out data from the processed data that has no missing content, continuous timestamps, and whose values ​​do not exceed the rated operating range of the devices and have no distortion jumps, forming a standardized time-series dataset that is sent to the global energy efficiency optimization decision layer.

4. The intelligent control system for dynamic energy efficiency optimization of high-computing-power data centers according to claim 1, characterized in that, The time-series prediction unit generates a computing power operation prediction time-series sequence corresponding to the heterogeneous computing power cluster within the next hour based on a standardized time-series dataset. The time-series prediction unit also generates a power distribution and cooling energy consumption prediction time-series sequence corresponding to the power supply and distribution equipment and cooling equipment within the next 24 hours based on the standardized time-series dataset, externally acquired temperature information, and preset power grid time period identifiers. The computing power operation prediction time-series sequence and the power distribution and cooling energy consumption prediction time-series sequence are sent to the dynamic energy efficiency assessment model for multi-device energy efficiency quantification calculation.

5. The intelligent control system for dynamic energy efficiency optimization of high-computing-power data centers according to claim 4, characterized in that, The computing power scheduling unit of the global energy efficiency optimization decision layer receives the multi-device energy efficiency quantification calculation results output by the dynamic energy efficiency evaluation model, and allocates computing power tasks to be allocated according to the calculation results. The computing power tasks are allocated to the heterogeneous computing power nodes with the best multi-device energy efficiency quantification calculation results. At the same time, for computing power nodes whose idle time exceeds a preset threshold, corresponding control instructions for hibernation and shutdown are generated. The control instructions are output from the global energy efficiency optimization decision layer to the closed-loop execution layer.

6. The intelligent control system for dynamic energy efficiency optimization of high-computing-power data centers according to claim 4, characterized in that, The cooling capacity regulation unit has a built-in infrastructure linkage optimization module. Based on the energy consumption change trend and the temperature distribution data of the computing node, the infrastructure linkage optimization module dynamically adjusts the air supply speed, air supply angle, and cold aisle cooling capacity distribution parameters of the precision air conditioner, generates cooling control commands, and outputs them to the closed-loop execution layer.

7. The intelligent control system for dynamic energy efficiency optimization of high-computing-power data centers according to claim 1, characterized in that, The closed-loop execution layer has a built-in instruction risk verification module. Before issuing control instructions, the instruction risk verification module verifies the execution risk of the control instructions and eliminates abnormal instructions that may cause overheating of computing nodes or power overload. After the instruction is executed, energy efficiency change data and equipment operation status data are collected and sent back to the global energy efficiency optimization decision layer to iteratively correct the weight parameters of the dynamic energy efficiency evaluation model.

8. The intelligent control system for dynamic energy efficiency optimization of high-computing-power data centers according to claim 1, characterized in that, It also includes a digital twin simulation and deduction unit that communicates with the global energy efficiency optimization decision-making layer. The digital twin simulation and deduction unit has a built-in full-element digital twin model of a high-computing-power data center. After the global energy efficiency optimization decision-making layer generates a control command to be verified, it first pushes it to the digital twin simulation and deduction unit. The digital twin simulation and deduction unit then performs a pre-simulation of the execution effect of the control command, evaluates the energy efficiency improvement and potential operational risks. When the evaluation result meets the preset threshold requirements, the global energy efficiency optimization decision-making layer then pushes the corresponding control command to the closed-loop execution layer.

9. The intelligent control system for dynamic energy efficiency optimization of high-computing-power data centers according to claim 1, characterized in that, It also includes an energy efficiency visualization unit that communicates with the global energy efficiency optimization decision-making layer. The energy efficiency visualization unit connects to the data center operation and management platform, receives various energy efficiency statistics output by the global energy efficiency optimization decision-making layer, and displays the real-time PUE, computing power energy efficiency ratio, cumulative energy saving, and control command execution records of the entire data center in real time.

Images