Intelligent calculation center multi-dimensional electric quantity monitoring and prediction control method and system

Abstract

The invention provides a multi-dimensional electric quantity monitoring and prediction control method and system for an intelligent calculation center, belongs to the technical field of power dispatching processing, and is used for realizing prospective and refined shaping of an overall energy consumption curve of the intelligent calculation center and remarkably improving energy efficiency and power grid stability. According to the method, high-precision analysis of the computing power load is realized through a low-cost sensing means, bus electrical waveform analysis and on-demand sampling verification of the server hardware state are combined, a software-hardware collaborative bidirectional calibration closed loop is constructed, expensive sensors do not need to be deployed on each device, and the cost is reduced. The operation states of equipment groups driven by different types of calculation tasks are accurately separated and identified from macroscopic mixed electrical signals, the inherent contradiction between the cost and the sensing granularity in a traditional monitoring scheme is solved, a high-quality data basis strongly associated with services is provided for subsequent accurate prediction and control, and the real-time performance of the system is improved. And the total load prediction accuracy and the dynamic response capability of the intelligent calculation center are improved.

Description

Technical Field

[0001] This invention relates to the field of power dispatching and processing technology, and in particular to a method and system for multi-dimensional power monitoring and predictive control in an intelligent computing center. Background Technology

[0002] Intelligent computing centers (also known as smart computing centers) are a new type of data center designed and built to meet the high computing power demands of artificial intelligence, big data analytics, and other fields. Their core feature is the large-scale deployment of high-performance computing chips, represented by graphics processing units (GPUs), specifically designed to perform computationally intensive tasks such as model training and data inference. Due to the enormous and dynamic nature of these tasks, intelligent computing centers have become a new type of infrastructure with huge energy consumption and drastic load fluctuations. Their refined energy management is of great significance for reducing operating costs and ensuring grid stability.

[0003] Energy management for data centers typically employs separate technologies. At the monitoring level, most solutions acquire total power consumption data by installing smart meters on the main power inlet line. Some solutions add monitoring of critical rack branches, but still cannot directly link power consumption to specific computing operations. At the prediction level, the mainstream method is to make predictions based on data from rack sensors, which often requires waiting for sensor readings. The cooling system is usually handled by a separate control system. For example, the environmental control system starts and stops the cooling equipment based on feedback from temperature sensors, while the computing task scheduling system mainly allocates tasks based on task priority and computing resource usage. There is a lack of direct coordination and linkage between the two.

[0004] The aforementioned existing technical solutions have obvious technical defects. First, in terms of perception, they can only judge the energy consumption of the equipment, resulting in a complete disconnect between computing power services and power consumption data. Because they do not consider the computing tasks to be executed and ignore the physical time delay effect between equipment heat generation and the response of the cooling system, their prediction models have low accuracy when faced with drastic dynamic changes in the load of the intelligent computing center, making it difficult to support fine-grained real-time scheduling, and there is room for improvement. Summary of the Invention

[0005] This invention provides a method and system for multi-dimensional power monitoring and predictive control of intelligent computing centers, which enables forward-looking and refined shaping of the overall energy consumption curve of intelligent computing centers, significantly improving energy efficiency and grid stability.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] In a first aspect, embodiments of the present invention provide a method for multi-dimensional power monitoring and predictive control in a smart computing center, the method comprising:

[0008] Obtain the device group operating status sequence from the current waveform data based on the power supply bus of the intelligent computing center;

[0009] Obtain the queue of computing power tasks to be executed and the parameters of computing power tasks, map the parameters of computing power tasks to the power consumption and heat of the device group, and generate a heat source timing input driven by the queue of computing power tasks to be executed based on the queue of computing power tasks to be executed and the power consumption and heat of the device group.

[0010] Based on the thermal capacity parameters of the set of device nodes divided by the physical space of the intelligent computing center, a thermal resistance network for heat transfer is established between the device nodes. The temperature state vector of the entire node is calculated based on the thermal capacity parameters and the thermal resistance network. The thermal capacity parameters and the thermal resistance network are pre-calibrated by combining computational fluid dynamics simulation and system identification.

[0011] The full node temperature state vector is input into the pre-trained heat flow network dynamic model, and the heat flow network dynamic state for future time periods is output.

[0012] Based on the dynamic state of the heat flow network and the energy efficiency performance coefficient of the server heat dissipation system of the intelligent computing center, a power prediction sequence for the heat dissipation system in future time periods is generated.

[0013] Based on the operating status sequence of the high-confidence equipment group and the power prediction sequence of the heat dissipation system for future periods, a total load prediction curve is constructed;

[0014] The total load forecast curve is input into a multi-objective optimization game model for solution, generating a multi-resource collaborative scheduling scheme that includes timestamps and control parameters;

[0015] According to the multi-resource collaborative scheduling scheme, the computing power task timing of the intelligent computing center, the air conditioning setting strategy of the intelligent computing center, and the energy storage system action of the intelligent computing center are generated and controlled in a coordinated manner.

[0016] Optionally, the acquisition of a high-reliability device group operating status sequence based on current waveform data from the intelligent computing center power supply bus includes:

[0017] Acquire current waveform data of the power supply bus of the intelligent computing center and power consumption data of multiple internal devices;

[0018] The current waveform data is subjected to load characteristic decomposition and source tracing processing to generate a preliminary load group characteristic waveform;

[0019] Acquire a database of device feature waveforms containing electrical characteristics of devices with different computing power under typical operating conditions, and extract a reference feature waveform from it;

[0020] The current waveform data is convolved with the reference feature waveform to generate the initial load contribution of all device types.

[0021] The current waveform data is decoupled using a blind source separation algorithm to generate anonymous independent source signal waveforms, wherein the blind source separation algorithm is a fast independent component analysis algorithm;

[0022] The correlation analysis is performed between the independent source signal waveform and the initial load contribution. Based on the analysis results, the preliminary load group characteristic waveform is determined.

[0023] The preliminary load group characteristic waveform is calibrated by combining the sampled power consumption data to generate a high-confidence device group operating status sequence.

[0024] Optionally, the step of calibrating the preliminary load group characteristic waveform using the sampled power consumption data to generate a high-confidence device group operating state sequence includes:

[0025] Monitor the initial load group characteristic waveform, and when the load change rate indicated by it exceeds the change threshold, generate and send a sampling command;

[0026] According to the sampling instruction, the real-time power consumption sampling data of the sample server is obtained from the hardware acquisition terminal of the server.

[0027] The real-time power consumption sampling data is compared with the power consumption data inferred from the preliminary load group characteristic waveform at the same time, and a dimensionless waveform calibration coefficient is calculated and generated.

[0028] The waveform calibration coefficients are applied to correct the initial load group characteristic waveforms to generate the high-reliability device group operating status sequence.

[0029] Optionally, the step of acquiring the queue of computing power tasks to be executed and generating a heat source timing input driven by the queue of computing power tasks to be executed, based on the power consumption heat of the device group that maps the computing power tasks to heat generation, includes:

[0030] Analyze the queue of computing power tasks to be executed and extract the task type identifier and pre-computation power scale;

[0031] Using the task type identifier and the pre-computation capacity, query the power consumption and heat of the device group to obtain the basic thermal power value of each task;

[0032] By combining the planned execution time window of each task with the server resource allocation strategy, the basic thermal power values ​​are allocated and superimposed in the time dimension to generate the heat source time sequence input.

[0033] Specifically: The benchmark library is pre-calibrated through experiments, and for each typical computing task type (such as AI model training, data inference, large file storage, etc.), its energy consumption characteristic curves under different computing power utilization rates are established. This curve usually includes two parts: basic static power consumption (heat generation when the device is powered on but has no task) and dynamic incremental power consumption (heat generation that increases non-linearly with computing power utilization rate).

[0034] Dynamic calculation involves first extracting the pre-computational power scale of the task to be executed during actual prediction and normalizing it to the processing utilization rate of the device (e.g., GPU utilization rate).

[0035] Lookup table and interpolation: The corresponding characteristic curve is indexed according to the task type, and the utilization rate is used to perform lookup table or interpolation calculation on the curve to obtain the real-time heat generation power of the task under a specific load pressure, reflecting the nonlinear surge of heat power caused by the increase of chip leakage current under high load.

[0036] Optionally, the step of loading the heat source time-series input into the heat flux network dynamic model for simulation and generation of the heat flux network dynamic state for future time periods includes:

[0037] The heat source timing input is loaded as heat flow into the corresponding heat source node in the heat flow network dynamic model;

[0038] The dynamic model of the heat flow network is calculated by solving the differential equations describing the heat balance between each node, with an iterative time step.

[0039] The calculation is repeated throughout the prediction period to obtain the predicted trajectory of future temperature changes at each key node. These predicted trajectories together constitute the dynamic state of the heat flow network.

[0040] Optionally, the step of calculating and generating a power prediction sequence for the cooling system for future periods based on the dynamic state of the heat flow network and in conjunction with performance coefficients characterizing the energy efficiency of the server cooling system includes:

[0041] Obtain the set threshold for the computer room temperature;

[0042] The predicted temperature trajectory in the dynamic state of the heat flow network is compared with the set threshold temperature of the computer room to calculate the heat required to be removed in order to keep the temperature below the threshold and generate a time series of temperature regulation requirements.

[0043] Divide the temperature regulation demand time series by the performance coefficient characterizing the energy efficiency of the server heat dissipation system to calculate and generate the power prediction sequence of the heat dissipation system, wherein the performance coefficient is the cooling energy efficiency ratio determined according to the air conditioning operating conditions.

[0044] Optionally, the step of inputting the total load forecast curve into a multi-objective optimization game model for balancing multi-dimensional operational objectives to generate a multi-resource collaborative scheduling scheme including timestamps and control parameters includes:

[0045] The power grid dispatch demand signal, the computing power task priority list, and the energy storage system status parameters are obtained and used as constraints for the multi-objective optimization game model.

[0046] Construct a weighted comprehensive objective function that includes grid interaction objectives, computing power task completion objectives, air conditioning energy efficiency objectives, and energy storage economic objectives;

[0047] The comprehensive objective function is solved under the constraints using an intelligent optimization algorithm to calculate the optimal solution for optimizing the computing task start time, air conditioning setpoint temperature, and energy storage system charging and discharging power. The optimal solution constitutes the multi-resource collaborative scheduling scheme, wherein the intelligent optimization algorithm is a genetic algorithm.

[0048] Optionally, parsing and executing the multi-resource collaborative scheduling scheme includes:

[0049] At the first point in time, a pre-scheduling instruction for computing power tasks, parsed from the multi-resource collaborative scheduling scheme, is sent to the computing task management platform;

[0050] Within a time window following the first time point, send the energy storage system pre-control command parsed from the multi-resource collaborative scheduling scheme to the energy storage management system;

[0051] At a second time point later than the first time point, the air conditioning pre-adjustment command parsed from the multi-resource collaborative scheduling scheme is sent to the air conditioning control system, thereby realizing the coordination and timing linkage of the three resources.

[0052] Optionally, after the execution of the linkage control, the method further includes:

[0053] Monitor and collect actual total load data and actual computer room ambient temperature data after the execution of linkage control;

[0054] The actual total load data is compared with the total load prediction curve, and the actual computer room ambient temperature data is compared with the dynamic status of the heat flow network to generate model error feedback data.

[0055] Using the model error feedback data, the power consumption and heat generation of the device group, as well as the thermal resistance and heat capacity parameters in the dynamic model of the heat flow network, are adaptively updated; and / or,

[0056] The dynamic model of the heat flow network is trained by collecting historical operating data to construct a loss function and using gradient descent to optimize the heat capacity and thermal resistance parameters.

[0057] Secondly, embodiments of the present invention provide a multi-dimensional power monitoring and prediction control system for intelligent computing centers, wherein the system is configured as follows:

[0058] The acquisition module is used to acquire the device group operating status sequence from the current waveform data based on the power supply bus of the intelligent computing center;

[0059] The generation module is used to obtain the queue of computing power tasks to be executed and the computing power task parameters; map the computing power task parameters to the power consumption and heat of the device group; and generate a heat source timing input driven by the queue of computing power tasks to be executed based on the queue of computing power tasks to be executed and the power consumption and heat of the device group.

[0060] The calculation module is used to establish a thermal resistance network for heat transfer between device nodes based on the thermal capacity parameters of the device node set divided by the physical space of the intelligent computing center; and to calculate the temperature state vector of all nodes based on the thermal capacity parameters and the thermal resistance network; wherein the thermal capacity parameters and the thermal resistance network are pre-calibrated by a combination of computational fluid dynamics simulation and system identification.

[0061] The future time period heat flow network dynamic state generation module is used to input the temperature state vector of all nodes into the pre-trained heat flow network dynamic model and output the future time period heat flow network dynamic state.

[0062] The power prediction sequence generation module is used to generate a power prediction sequence for the cooling system in future time periods based on the dynamic state of the heat flow network and the energy efficiency performance coefficient of the server cooling system of the intelligent computing center.

[0063] The prediction curve construction module is used to construct a total load prediction curve based on the operating status sequence of the high-confidence equipment group and the power prediction sequence of the heat dissipation system in the future period.

[0064] The scheduling scheme generation module is used to input the total load prediction curve into a multi-objective optimization game model for solving, and generate a multi-resource collaborative scheduling scheme that includes timestamps and control parameters.

[0065] The scheduling module is used to generate the computing power task timing of the intelligent computing center, the air conditioning setting strategy of the intelligent computing center, and to perform linkage control on the actions of the equipment in the intelligent computing center in accordance with the multi-resource collaborative scheduling scheme.

[0066] Thirdly, embodiments of the present invention provide an electronic device, including: a processor and a memory; the memory is used to store a computer program, and when the processor executes the computer program, the electronic device performs the intelligent computing center multi-dimensional power monitoring and predictive control method described in the first aspect.

[0067] In one possible design, the electronic device described in the third aspect may further include a transceiver. This transceiver may be a transceiver circuit or an interface circuit. The transceiver can be used for communication between the electronic device described in the third aspect and other electronic devices.

[0068] In the embodiments of the present invention, the electronic device described in the third aspect may be a terminal, or a chip (system) or other component or assembly disposed in the terminal, or a system containing the terminal.

[0069] Fourthly, embodiments of the present invention provide a computer-readable storage medium, including: a computer program or instructions; when the computer program or instructions are run on a computer, the computer causes the computer to execute the intelligent computing center multi-dimensional power monitoring and predictive control method described in the first aspect.

[0070] In summary, the above methods and systems have the following technical effects:

[0071] This invention achieves high-precision analysis of computing load through low-cost sensing methods. It combines bus electrical waveform analysis with on-demand sampling verification of server hardware status to construct a two-way calibration closed loop that integrates software and hardware. This eliminates the need to deploy expensive sensors on every device, accurately separating and identifying the operating status of device groups driven by different types of computing tasks from macroscopic mixed electrical signals. It resolves the inherent contradiction between cost and sensing granularity in traditional monitoring schemes, providing a high-quality, business-relevant data foundation for subsequent accurate prediction and control. This improves the accuracy and dynamic response capability of total load prediction for intelligent computing centers. It abandons the traditional prediction paradigm that relies solely on historical power data, creatively establishing a physical causal prediction chain from computing power operations to power consumption. By using the queue of computing tasks to be executed as the initial driving source for prediction, and combining it with a physical model considering the thermal inertia of the data center, it deduces the dynamic changes in the cooling system load caused by the heat generated by IT equipment, improving the prediction accuracy for complex fluctuations and time-delay effects caused by sudden tasks. Attached Figure Description

[0072] Figure 1 A flowchart illustrating the multi-dimensional power monitoring and predictive control method for intelligent computing centers provided in this embodiment of the invention;

[0073] Figure 2 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0074] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0075] In this embodiment of the invention, "instruction" can include direct and indirect instructions, as well as explicit and implicit instructions. The information indicated by a certain piece of information is called the information to be instructed. In specific implementation, there are many ways to instruct the information to be instructed, such as, but not limited to, directly instructing the information to be instructed, such as the information to be instructed itself or its index. It can also indirectly instruct the information to be instructed by instructing other information, where there is a correlation between the other information and the information to be instructed. It can also instruct only a part of the information to be instructed, while the other parts are known or pre-agreed upon. For example, the instruction of specific information can be achieved by using a pre-agreed (e.g., protocol-defined) arrangement of various pieces of information, thereby reducing instruction overhead to some extent. Simultaneously, common parts of various pieces of information can be identified and uniformly indicated to reduce the instruction overhead caused by individually indicating the same information.

[0076] Furthermore, the specific indication method can also be any existing indication method, such as, but not limited to, the above-mentioned indication methods and their various combinations. Specific details of various indication methods can be found in existing technologies, and will not be elaborated upon here. As described above, for example, when multiple pieces of information of the same type need to be indicated, the indication methods for different pieces of information may differ. In specific implementation, the required indication method can be selected according to specific needs. This embodiment of the invention does not limit the selected indication method; therefore, the indication methods involved in this embodiment of the invention should be understood to cover various methods that enable the party to be indicated to obtain the information to be indicated.

[0077] It should be understood that the information to be indicated can be sent as a whole or divided into multiple sub-information messages sent separately, and the sending period and / or timing of these sub-information messages can be the same or different. The specific sending method is not limited in this embodiment of the invention. The sending period and / or timing of these sub-information messages can be predefined, for example, according to a protocol, or configured by the sending device by sending configuration information to the receiving device.

[0078] "Predefined" or "pre-configured" can be achieved by pre-saving corresponding codes, tables, or other means that can be used to indicate relevant information in the device. This embodiment of the invention does not limit the specific implementation method. "Saving" can refer to saving in one or more memories. These memories can be separate installations or integrated into the encoder, decoder, processor, or electronic device. Alternatively, some memories can be separately installed, while others are integrated into the decoder, processor, or electronic device. The type of memory can be any form of storage medium, and this embodiment of the invention does not limit this.

[0079] In the embodiments of this invention, "protocol" may refer to a protocol family in the field of communication, a standard protocol with a similar protocol family frame structure, or a related protocol applied to a future intelligent computing center multi-dimensional power monitoring and predictive control method system. The embodiments of this invention do not specifically limit this.

[0080] In this embodiment of the invention, descriptions such as "when," "under the circumstances," "if," and "if" all refer to the device making corresponding processing under certain objective circumstances, and are not limited to a specific time. They do not require the device to make a judgment action during implementation, nor do they imply any other limitations.

[0081] In the description of the embodiments of the present invention, unless otherwise stated, " / " indicates that the objects before and after are in an "or" relationship. For example, A / B can represent A or B. "And / or" in the embodiments of the present invention is merely a description of the relationship between the related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone, where A and B can be singular or plural. Furthermore, in the description of the embodiments of the present invention, unless otherwise stated, "multiple" refers to two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple. Additionally, to facilitate a clear description of the technical solutions of the embodiments of the present invention, the terms "first" and "second" are used in the embodiments of the present invention to distinguish identical or similar items with essentially the same function and effect. Those skilled in the art will understand that the terms "first," "second," etc., do not limit the quantity or order of execution, and that "first," "second," etc., are not necessarily different. Furthermore, in the embodiments of this invention, words such as "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or description. Any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of this invention should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner for ease of understanding.

[0082] The network architecture and business scenarios described in the embodiments of this invention are for the purpose of more clearly illustrating the technical solutions of the embodiments of this invention, and do not constitute a limitation on the technical solutions provided by the embodiments of this invention. As those skilled in the art will know, with the evolution of network architecture and the emergence of new business scenarios, the technical solutions provided by the embodiments of this invention are also applicable to similar technical problems.

[0083] Figure 1 This is a flowchart illustrating the multi-dimensional power monitoring and predictive control method for intelligent computing centers provided in an embodiment of the present invention. This method can be applied to the aforementioned control terminal, and the specific process is as follows:

[0084] Obtain the device group operating status sequence from the current waveform data based on the power supply bus of the intelligent computing center;

[0085] Obtain the queue of computing power tasks to be executed and the parameters of computing power tasks, map the parameters of computing power tasks to the power consumption and heat of the device group, and generate a heat source timing input driven by the queue of computing power tasks to be executed based on the queue of computing power tasks to be executed and the power consumption and heat of the device group.

[0086] Based on the thermal capacity parameters of the set of device nodes divided by the physical space of the intelligent computing center, a thermal resistance network for heat transfer is established between the device nodes. The temperature state vector of the entire node is calculated based on the thermal capacity parameters and the thermal resistance network. The thermal capacity parameters and the thermal resistance network are pre-calibrated by combining computational fluid dynamics simulation and system identification.

[0087] The full node temperature state vector is input into the pre-trained heat flow network dynamic model, and the heat flow network dynamic state for future time periods is output.

[0088] Based on the dynamic state of the heat flow network and the energy efficiency performance coefficient of the server heat dissipation system of the intelligent computing center, a power prediction sequence for the heat dissipation system in future time periods is generated.

[0089] Based on the operating status sequence of the high-confidence equipment group and the power prediction sequence of the heat dissipation system for future periods, a total load prediction curve is constructed;

[0090] The total load forecast curve is input into a multi-objective optimization game model for solution, generating a multi-resource collaborative scheduling scheme that includes timestamps and control parameters;

[0091] According to the multi-resource collaborative scheduling scheme, the computing power task timing of the intelligent computing center, the air conditioning setting strategy of the intelligent computing center, and the energy storage system action of the intelligent computing center are generated and controlled in a coordinated manner.

[0092] Optionally, obtaining a high-reliability sequence of device group operating status based on current waveform data from the intelligent computing center's power supply bus includes:

[0093] Acquire current waveform data of the power supply bus of the intelligent computing center and power consumption data of multiple internal devices;

[0094] The current waveform data is subjected to load characteristic decomposition and source tracing processing to generate a preliminary load group characteristic waveform;

[0095] Acquire a database of device feature waveforms containing electrical characteristics of devices with different computing power under typical operating conditions, and extract a reference feature waveform from it;

[0096] The current waveform data is convolved with the reference feature waveform to generate the initial load contribution of all device types.

[0097] The current waveform data is decoupled using a blind source separation algorithm to generate anonymous independent source signal waveforms, wherein the blind source separation algorithm is a fast independent component analysis algorithm;

[0098] The correlation analysis is performed between the independent source signal waveform and the initial load contribution. Based on the analysis results, the preliminary load group characteristic waveform is determined.

[0099] The preliminary load group characteristic waveform is calibrated by combining the sampled power consumption data to generate a high-confidence device group operating status sequence.

[0100] The blind source separation algorithm mentioned above is a fast independent component analysis algorithm;

[0101] The construction method of the "device feature waveform library" is as follows: In an electrically isolated test environment, a single typical computing device to be calibrated (such as a specific model of GPU server or AI accelerator card cluster) is connected to the power supply bus, and a data acquisition device is configured to acquire the current waveform of its bus at a high sampling rate (such as 100kHz). Subsequently, the device is controlled to operate under various typical steady-state and transient conditions (such as power-on, idle, full-load computing, and load step change), and continuous and stable current waveform segments under various conditions are recorded. Feature extraction is performed on these raw waveforms. The extracted features include, but are not limited to: the amplitude and phase of the fundamental wave and each harmonic, waveform distortion, and spectral energy distribution of a specific frequency band. The extracted feature vectors are associated and stored with the corresponding device model and operating condition labels, thus forming the device feature waveform library. This library serves as the benchmark for subsequent feature matching and tracing of mixed waveforms.

[0102] The specific method for identity identification is as follows: Calculate the Pearson correlation coefficient between each anonymous independent source signal waveform and the initial load contribution waveforms of various device types generated through convolutional matching. Assign the device type with the highest correlation coefficient (requiring a correlation coefficient threshold greater than 0.75) to the independent source signal to complete its identity identification. For example, if an independent source signal has a correlation coefficient of 0.88 with the initial load contribution waveform of the "A100 GPU server cluster," but only 0.12 with the waveform of the "CPU computing server," then the source signal is identified as the operating status signal of the "A100 GPU server cluster."

[0103] Specifically: Based on the principle of signal statistical independence, the decoupling of mixed signals is achieved without needing to know the specific waveform characteristics of each source signal in advance.

[0104] The specific processing logic is as follows: Data preprocessing involves removing the mean and whitening the multi-channel current data to eliminate the second-order correlation between signals, making the data more statistically compact.

[0105] Independence Maximization Search: The core assumption of the algorithm is that the source signals (power waveforms of different types of devices) are statistically independent and have the strongest non-Gaussianity. According to the central limit theorem, the Gaussianity of the mixed signal will be enhanced. By iteratively adjusting a demixing matrix, the mixed signal is projected into a new coordinate system.

[0106] Iterative convergence: In each iteration, the algorithm calculates a measure of the non-Gaussianity of the separated signals (such as kurtosis or negative entropy). The algorithm continuously fine-tunes the coefficients of the mixing matrix until the non-Gaussianity of each separated signal component reaches its maximum value. At this point, the signals are considered to have achieved statistical maximal independence, thus separating independent source signal waveforms representing the operating characteristics of different equipment groups.

[0107] Optionally, the step of calibrating the preliminary load group characteristic waveform using the sampled power consumption data to generate a high-confidence device group operating state sequence includes:

[0108] Monitor the initial load group characteristic waveform, and when the load change rate indicated by it exceeds the change threshold, generate and send a sampling command;

[0109] According to the sampling instruction, the real-time power consumption sampling data of the sample server is obtained from the hardware acquisition terminal of the server.

[0110] Among them, the hardware acquisition terminal is the corresponding baseboard management controller;

[0111] The real-time power consumption sampling data is compared with the power consumption data inferred from the preliminary load group characteristic waveform at the same time, and a dimensionless waveform calibration coefficient is calculated and generated.

[0112] The specific formula for calculating the dimensionless waveform calibration coefficient is: waveform calibration coefficient k = P_sample / P_inferred. Where P_sample is the average value of the real-time power consumption sampling data of the sample server obtained from the baseboard management controller (BMC); P_inferred is the total power consumption value inferred from the preliminary load group characteristic waveform within the same time window, corresponding to the same sample server device group. This coefficient k directly reflects the ratio between the waveform inferred power consumption and the actual sampled power consumption. If k significantly deviates from 1.0, it indicates a deviation in the preliminary waveform inference.

[0113] The waveform calibration coefficients are applied to correct the initial load group characteristic waveforms to generate the high-reliability device group operating status sequence.

[0114] Optionally, the step of acquiring the queue of computing power tasks to be executed and generating a heat source timing input driven by the queue of computing power tasks to be executed, based on the power consumption heat of the device group that maps the computing power tasks to heat generation, includes:

[0115] Analyze the queue of computing power tasks to be executed and extract the task type identifier and pre-computation power scale;

[0116] Using the task type identifier and the pre-computation capacity, query the power consumption and heat of the device group to obtain the basic thermal power value of each task;

[0117] By combining the planned execution time window of each task with the server resource allocation strategy, the basic thermal power values ​​are allocated and superimposed in the time dimension to generate the heat source time sequence input.

[0118] To map computing power task parameters to device cluster power consumption and heat, this invention pre-constructs a "task type-power consumption / heat mapping benchmark library." This benchmark library is established through experimental calibration, specifically as follows: On servers with typical configurations (such as servers equipped with NVIDIA A100 GPUs), different typical computing power tasks (including but not limited to AI model training, batch data inference, and high-performance computing) are deployed, and their computing power utilization is precisely controlled using performance monitoring tools (such as nvidia-smi) (e.g., GPU utilization from 10% to 100%, in 10% increments). At each stable utilization rate, the server's total power consumption data (obtainable through the Baseboard Management Controller, BMC) and the surface temperature of the processor chip are simultaneously collected. More than 95% of the total power consumption is considered as the portion ultimately converted into heat, thus obtaining the real-time heat power value for each task type at different computing power utilization rates, forming a mapping table. The table below shows a simplified mapping example:

[0119] Task type identifier Pre-computing power scale (GPU utilization) Basic thermal power value (KW) TRAIN_CNN 30% 0.75 TRAIN_CNN 70% 1.65 INFERENCE_LLM 20% 0.55 INFERENCE_LLM 50% 1.20 HPC_FLOPS 90% 2.10

[0120] In actual prediction, the system parses the type identifier and pre-calculated power scale of the task to be executed. By querying this benchmark library and calculating using linear interpolation, the accurate basic thermal power value can be obtained. This method directly links the abstract "computing power" with the physical "heat," forming an accurate input source for subsequent thermodynamic simulations.

[0121] Optionally, the step of loading the heat source time-series input into the heat flux network dynamic model for simulation and generation of the heat flux network dynamic state for future time periods includes:

[0122] The heat source timing input is loaded as heat flow into the corresponding heat source node in the heat flow network dynamic model;

[0123] The dynamic model of the heat flow network is calculated by solving the differential equations describing the heat balance between each node, with an iterative time step.

[0124] The calculation is repeated throughout the prediction period to obtain the predicted trajectory of future temperature changes at each key node. These predicted trajectories together constitute the dynamic state of the heat flow network.

[0125] Specifically, the dynamic heat flow network model is a thermal resistance-thermal capacity (RC) network model constructed based on the lumped parameter method. In this model, the physical space of the intelligent computing center is discretized into N temperature nodes (e.g., server internal heat source nodes, rack air nodes, cold aisle nodes, hot aisle nodes, and enclosure structure nodes). The differential equations describing the heat balance between each node specifically adopt the following form of first-order ordinary differential equations:

[0126]

[0127] Where: i represents the index of the currently calculated node, and j represents the index of the node adjacent to node i. i is the set of adjacent nodes of node i; Ti(t) represents the temperature of node i at time t (state variable, unit: °C); Ci represents the heat capacity of node i (model parameter, unit: J / K), which characterizes the thermal inertia of the region; Qi(t) represents the input heat power acting on node i at time t (input variable, unit: W), which is the "heat source time sequence input" generated above; Rij represents the heat transfer resistance between node i and node j (model parameter, unit: K / W), which characterizes the resistance to heat transfer, including conductive thermal resistance and convective thermal resistance;

[0128] The embodiment uses the fourth-order Runge-Kutta method for discretization and iterative calculation to determine the model with the optimal parameter matrix. The specific training process is as follows: collect the heat source input Q(t) and measured temperature T_{real}(t) from the historical operation of the intelligent computing center, construct the loss function L=\sum(T_{pred}(t)-T_{real}(t))^2, and use the least squares method or gradient descent method to solve in reverse the values ​​of heat capacity C_i and thermal resistance R_{ij} that minimize the loss function.

[0129] The dynamic model of the heat flow network is trained by collecting historical operating data to construct a loss function and using gradient descent to optimize the heat capacity and thermal resistance parameters.

[0130] The training data for the dynamic model of the heat flow network is "historical operating data," which is specifically composed and preprocessed as follows: "Historical heat source input Q(t)" is obtained by back-calculating from the computing power tasks executed in the same historical period, based on the aforementioned "task type-power consumption / heat mapping benchmark library"; "Measured temperature T_real(t)" is the historical reading of temperature sensors deployed at each key node (corresponding to the model node) in the computer room. Before training, the two types of time series data need to be time-aligned and outlier removed. The processed Q(t) is used as input to the heat flow network model to be trained for simulation, resulting in the predicted temperature sequence T_pred(t). The loss function L=Σ(T_pred(t)-T_real(t))^2 measures the deviation between the prediction and the actual value. The gradient descent method is used to iteratively adjust the thermal resistance R_ij and heat capacity C_i parameters in the model until the loss function converges to the minimum value. At this point, a pre-trained model that reflects the specific thermal dynamic characteristics of the intelligent computing center is obtained.

[0131] Optionally, the step of calculating and generating a power prediction sequence for the cooling system for future periods based on the dynamic state of the heat flow network and in conjunction with performance coefficients characterizing the energy efficiency of the server cooling system includes:

[0132] Obtain the set threshold for the computer room temperature;

[0133] The predicted temperature trajectory in the dynamic state of the heat flow network is compared with the set threshold temperature of the computer room to calculate the heat required to be removed in order to keep the temperature below the threshold and generate a time series of temperature regulation requirements.

[0134] Divide the temperature regulation demand time series by the performance coefficient characterizing the energy efficiency of the server heat dissipation system to calculate and generate the power prediction sequence of the heat dissipation system, wherein the performance coefficient is the cooling energy efficiency ratio determined according to the air conditioning operating conditions.

[0135] Optionally, the step of inputting the total load forecast curve into a multi-objective optimization game model for balancing multi-dimensional operational objectives to generate a multi-resource collaborative scheduling scheme including timestamps and control parameters includes:

[0136] The power grid dispatch demand signal, the computing power task priority list, and the energy storage system status parameters are obtained and used as constraints for the multi-objective optimization game model.

[0137] What needs to be solved is the control sequence within the future scheduling period T (e.g., the next 24 hours). A weighted comprehensive objective function is constructed, encompassing grid interaction objectives, computing power task completion objectives, air conditioning energy efficiency objectives, and energy storage economic objectives, defining the decision vector X:

[0138]

[0139] Jgrid The goal is to evaluate the alignment between the total load curve and the grid dispatch instructions, with a smaller deviation resulting in a higher score, as shown in the following formula:

[0140]

[0141] In the formula: Ptotal(t) is the total load forecast value of the intelligent computing center at time t, C(t) is the time-of-use electricity price at time t, Plimit(t) is the demand response signal of the maximum power limit curve issued by the power grid, and λ is the over-limit penalty coefficient.

[0142] J task To assess the on-time start rate of high-priority tasks as the target for computing power task completion, the lower the latency, the higher the score, as shown in the following formula:

[0143]

[0144] In the formula: ρi: the priority weight of the i-th task; the higher the priority, the greater the weight. start i: the actual startup time obtained from the optimization solution, t plan i: The originally planned start time of the task;

[0145] J ac To achieve the energy efficiency target for air conditioning, the total cooling energy consumption of the heat dissipation system is evaluated, and the electrical energy consumption of the heat dissipation system is minimized. The lower the energy consumption, the higher the score.

[0146]

[0147] In the formula: Q load For task-based scheduling t start The heat load predicted by the heat flow network model, COP is the air conditioning performance coefficient, which is the set temperature T. set and ambient temperature T env The function;

[0148] J storage To achieve the economic goals of energy storage, the cost of battery life loss is taken into account, and frequent ineffective charging and discharging is avoided. The profitability of the energy storage system in utilizing the peak-valley electricity price difference is evaluated. The higher the profitability, the higher the score.

[0149]

[0150] In the formula: C cycle Battery depreciation cost factor per unit power throughput, P ba t(t) represents the charging and discharging power of the energy storage system at time $t$. The goal is to find the optimal balance point (i.e., Pareto optimal solution) among four mutually constraining objectives: grid demand, computing power performance, air conditioning energy efficiency, and energy storage cost.

[0151] The scores of the above four dimensions are weighted and synthesized using preset priority weight coefficients to form a unique fitness index for measuring the quality of the scheduling scheme.

[0152] The optimization algorithm is executed, and an intelligent evolutionary algorithm (such as a genetic algorithm or particle swarm optimization algorithm) is used to solve the problem.

[0153] Initialization generates hundreds or thousands of possible scheduling schemes (including different combinations of task start time, air conditioning setpoint and energy storage charging and discharging power).

[0154] Simulation and scoring: Each scheme is substituted into the aforementioned heat flow network model for pre-simulation, and its fitness index is calculated based on the aforementioned comprehensive evaluation system.

[0155] Through survival of the fittest and evolution, the best solutions with the highest fitness index are retained, while inferior solutions are eliminated. The parameters of the best solutions are then cross-combined and slightly mutated to generate a new generation of solutions.

[0156] Convergence Output: After multiple rounds of iterative evolution, when the fitness index of the optimal solution no longer improves significantly, the solution is determined as the final multi-resource collaborative scheduling solution, wherein the intelligent optimization algorithm is a genetic algorithm.

[0157] The constraints include: power balance constraints, energy storage system constraints (including capacity, power and state continuity), data center thermal environment constraints, and task scheduling window constraints.

[0158] Optionally, parsing and executing the multi-resource collaborative scheduling scheme includes:

[0159] At the first point in time, a pre-scheduling instruction for computing power tasks, parsed from the multi-resource collaborative scheduling scheme, is sent to the computing task management platform;

[0160] Within a time window following the first time point, send the energy storage system pre-control command parsed from the multi-resource collaborative scheduling scheme to the energy storage management system;

[0161] At a second time point later than the first time point, the air conditioning pre-adjustment command parsed from the multi-resource collaborative scheduling scheme is sent to the air conditioning control system, thereby realizing the coordination and timing linkage of the three resources.

[0162] Optionally, after the execution of the linkage control, the method further includes:

[0163] Monitor and collect actual total load data and actual computer room ambient temperature data after the execution of linkage control;

[0164] The actual total load data is compared with the total load prediction curve, and the actual computer room ambient temperature data is compared with the dynamic status of the heat flow network to generate model error feedback data.

[0165] Using the model error feedback data, the power consumption and heat of the device group and the thermal resistance and thermal capacity parameters in the dynamic model of the heat flow network are adaptively updated.

[0166] In another embodiment, this embodiment also discloses a multi-dimensional power monitoring and predictive control system for intelligent computing centers, the system being configured as follows:

[0167] The acquisition module is used to acquire the device group operating status sequence from the current waveform data based on the power supply bus of the intelligent computing center;

[0168] The generation module is used to obtain the queue of computing power tasks to be executed and the computing power task parameters; map the computing power task parameters to the power consumption and heat of the device group; and generate a heat source timing input driven by the queue of computing power tasks to be executed based on the queue of computing power tasks to be executed and the power consumption and heat of the device group.

[0169] The calculation module is used to establish a thermal resistance network for heat transfer between device nodes based on the thermal capacity parameters of the device node set divided by the physical space of the intelligent computing center; and to calculate the temperature state vector of all nodes based on the thermal capacity parameters and the thermal resistance network; wherein the thermal capacity parameters and the thermal resistance network are pre-calibrated by a combination of computational fluid dynamics simulation and system identification.

[0170] The future time period heat flow network dynamic state generation module is used to input the temperature state vector of all nodes into the pre-trained heat flow network dynamic model and output the future time period heat flow network dynamic state.

[0171] The power prediction sequence generation module is used to generate a power prediction sequence for the cooling system in future time periods based on the dynamic state of the heat flow network and the energy efficiency performance coefficient of the server cooling system of the intelligent computing center.

[0172] The prediction curve construction module is used to construct a total load prediction curve based on the operating status sequence of the high-confidence equipment group and the power prediction sequence of the heat dissipation system in the future period.

[0173] The scheduling scheme generation module is used to input the total load prediction curve into a multi-objective optimization game model for solving, and generate a multi-resource collaborative scheduling scheme that includes timestamps and control parameters.

[0174] The scheduling module is used to generate the computing power task timing of the intelligent computing center, the air conditioning setting strategy of the intelligent computing center, and to perform linkage control on the actions of the equipment in the intelligent computing center in accordance with the multi-resource collaborative scheduling scheme.

[0175] During the prediction phase, the coupled prediction module receives the aforementioned sequence of operating states and simultaneously acquires the queue of computing power tasks to be executed from external input. This module does not rely solely on historical power data but rather deduces a total load prediction curve that includes IT equipment load and cooling system load based on physical causal chains.

[0176] During the control phase, the collaborative scheduling control module receives the total load forecast curve, combines the constraints of the power grid, energy storage and data center environment, generates a multi-resource collaborative scheduling scheme through a multi-objective optimization game model, and parses it into specific control commands to be sent to the physical layer.

[0177] The feedback loop collects actual load and temperature data after the control is executed, generates model error feedback data, and sends it back to the coupled prediction module to correct physical parameters such as thermal resistance and heat capacity.

[0178] This invention provides a schematic diagram of the structure of an electronic device. Exemplarily, the electronic device may be a network device, or a chip (system) or other component or assembly that can be disposed in a network device. The electronic device may include a processor. Optionally, the electronic device may further include a memory and / or a transceiver. The processor is coupled to the memory and transceiver, for example, by means of a communication bus connection.

[0179] The following is a detailed introduction to the various components of the electronic device:

[0180] In this context, the processor is the control center of the electronic device. It can be a single processor or a collective term for multiple processing elements. For example, a processor can be one or more central processing units (CPUs), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention, such as one or more digital signal processors (DSPs), or one or more field-programmable gate arrays (FPGAs).

[0181] Optionally, the processor can perform various functions of the electronic device by running or executing software programs stored in memory and calling data stored in memory, such as executing the above-mentioned power system computing power and power collaborative scheduling method based on regional intelligent computing centers.

[0182] In a specific implementation, as one example, the processor may include one or more CPUs, such as CPU0 and CPU1.

[0183] In a specific implementation, as one example, the electronic device may also include multiple processors. Each of these processors may be a single-core processor (single-CPU) or a multi-core processor (multi-CPU). Here, a processor may refer to one or more devices, circuits, and / or processing cores used to process data (e.g., computer program instructions).

[0184] The memory is used to store the software program that executes the solution of the present invention, and the execution is controlled by the processor. The specific implementation method can be referred to the above method embodiment, and will not be repeated here.

[0185] Optionally, the memory can be read-only memory (ROM) or other types of static storage devices capable of storing static information and instructions, random access memory (RAM) or other types of dynamic storage devices capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. The memory can be integrated with the processor or exist independently and coupled to the processor through an interface circuit of an electronic device; the embodiments of the present invention do not specifically limit this.

[0186] A transceiver is used for communication with other electronic devices. For example, if the electronic device is a terminal, the transceiver can be used to communicate with a network device or with another terminal device. Similarly, if the electronic device is a network device, the transceiver can be used to communicate with a terminal or with another network device.

[0187] Optionally, the transceiver may include a receiver and a transmitter. The receiver is used to implement the receiving function, and the transmitter is used to implement the sending function.

[0188] Optionally, the transceiver can be integrated with the processor or exist independently and coupled to the processor through the interface circuit of the electronic device. This embodiment of the invention does not specifically limit this.

[0189] It is understood that the structure of an electronic device does not constitute a limitation on the electronic device. An actual electronic device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0190] Furthermore, the technical effects of the electronic devices can be referenced from the technical effects of the power system computing power and power collaborative scheduling method based on regional intelligent computing centers described in the above method embodiments, and will not be repeated here.

[0191] It should be understood that the processor in the embodiments of the present invention can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0192] It should also be understood that the memory in the embodiments of the present invention can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDRSDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DRRAM).

[0193] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of 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. A semiconductor medium can be a solid-state drive.

[0194] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0195] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0196] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0197] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0198] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0199] In the embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0200] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0201] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0202] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0203] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for multi-dimensional power monitoring and predictive control in an intelligent computing center, characterized in that, The method includes: Obtain the device group operating status sequence from the current waveform data based on the power supply bus of the intelligent computing center; Obtain the queue of computing power tasks to be executed and the parameters of computing power tasks, map the parameters of computing power tasks to the power consumption and heat of the device group, and generate a heat source timing input driven by the queue of computing power tasks to be executed based on the queue of computing power tasks to be executed and the power consumption and heat of the device group. Based on the thermal capacity parameters of the set of device nodes divided by the physical space of the intelligent computing center, a thermal resistance network for heat transfer is established between the device nodes. The temperature state vector of the entire node is calculated based on the thermal capacity parameters and the thermal resistance network. The thermal capacity parameters and the thermal resistance network are pre-calibrated by combining computational fluid dynamics simulation and system identification. The full node temperature state vector is input into the pre-trained heat flow network dynamic model, and the heat flow network dynamic state for future time periods is output. Based on the dynamic state of the heat flow network and the energy efficiency performance coefficient of the server heat dissipation system of the intelligent computing center, a power prediction sequence for the heat dissipation system in future time periods is generated. Based on the operating status sequence of the high-confidence equipment group and the power prediction sequence of the heat dissipation system for future periods, a total load prediction curve is constructed; The total load forecast curve is input into a multi-objective optimization game model for solution, generating a multi-resource collaborative scheduling scheme that includes timestamps and control parameters; According to the multi-resource collaborative scheduling scheme, the computing power task timing of the intelligent computing center, the air conditioning setting strategy of the intelligent computing center, and the energy storage system action of the intelligent computing center are generated and controlled in a coordinated manner.

2. The method according to claim 1, characterized in that, The acquisition of a high-reliability device group operating status sequence based on current waveform data from the intelligent computing center power supply bus includes: Acquire current waveform data of the power supply bus of the intelligent computing center and power consumption data of multiple internal devices; The current waveform data is subjected to load characteristic decomposition and source tracing processing to generate a preliminary load group characteristic waveform; Acquire a database of device feature waveforms containing electrical characteristics of devices with different computing power under typical operating conditions, and extract a reference feature waveform from it; The current waveform data is convolved with the reference feature waveform to generate the initial load contribution of all device types. The current waveform data is decoupled using a blind source separation algorithm to generate anonymous independent source signal waveforms, wherein the blind source separation algorithm is a fast independent component analysis algorithm; The correlation analysis is performed between the independent source signal waveform and the initial load contribution. Based on the analysis results, the preliminary load group characteristic waveform is determined. The preliminary load group characteristic waveform is calibrated by combining the sampled power consumption data to generate a high-confidence device group operating status sequence.

3. The method according to claim 2, characterized in that, The step of calibrating the preliminary load group characteristic waveform based on the sampled power consumption data to generate a high-confidence device group operating status sequence includes: Monitor the initial load group characteristic waveform, and when the load change rate indicated by it exceeds the change threshold, generate and send a sampling command; According to the sampling instruction, the real-time power consumption sampling data of the sample server is obtained from the hardware acquisition terminal of the server. The real-time power consumption sampling data is compared with the power consumption data inferred from the preliminary load group characteristic waveform at the same time, and a dimensionless waveform calibration coefficient is calculated and generated. The waveform calibration coefficients are applied to correct the initial load group characteristic waveforms to generate the high-reliability device group operating status sequence.

4. The method according to claim 1, characterized in that, The step of acquiring the queue of computing power tasks to be executed and generating a heat source timing input driven by the queue of computing power tasks to be executed, based on mapping the computing power tasks to the power consumption heat of the device group that generates heat, includes: Analyze the queue of computing power tasks to be executed and extract the task type identifier and pre-computation power scale; Using the task type identifier and the pre-computation capacity, query the power consumption and heat of the device group to obtain the basic thermal power value of each task; By combining the planned execution time window of each task with the server resource allocation strategy, the basic thermal power values ​​are allocated and superimposed in the time dimension to generate the heat source time sequence input.

5. The method according to claim 1, characterized in that, The step of loading the heat source time-series input into the heat flow network dynamic model for simulation and generation of the heat flow network dynamic state for future time periods includes: The heat source timing input is loaded as heat flow into the corresponding heat source node in the heat flow network dynamic model; The dynamic model of the heat flow network is calculated by solving the differential equations describing the heat balance between each node, with an iterative time step. The calculation is repeated throughout the prediction period to obtain the predicted trajectory of future temperature changes at each key node. These predicted trajectories together constitute the dynamic state of the heat flow network.

6. The method according to claim 5, characterized in that, The step of calculating and generating a power prediction sequence for the cooling system in future time periods based on the dynamic state of the heat flow network and in conjunction with the performance coefficient characterizing the energy efficiency of the server cooling system includes: Obtain the set threshold for the computer room temperature; The predicted temperature trajectory in the dynamic state of the heat flow network is compared with the set threshold temperature of the computer room to calculate the heat required to be removed in order to keep the temperature below the threshold and generate a time series of temperature regulation requirements. Divide the temperature regulation demand time series by the performance coefficient characterizing the energy efficiency of the server heat dissipation system to calculate and generate the power prediction sequence of the heat dissipation system, wherein the performance coefficient is the cooling energy efficiency ratio determined according to the air conditioning operating conditions.

7. The method according to claim 1, characterized in that, The step of inputting the total load forecast curve into a multi-objective optimization game model for balancing multi-dimensional operational objectives for solution, generating a multi-resource collaborative scheduling scheme including timestamps and control parameters, includes: The power grid dispatch demand signal, the computing power task priority list, and the energy storage system status parameters are obtained and used as constraints for the multi-objective optimization game model. Construct a weighted comprehensive objective function that includes grid interaction objectives, computing power task completion objectives, air conditioning energy efficiency objectives, and energy storage economic objectives; The comprehensive objective function is solved under the constraints using an intelligent optimization algorithm to calculate the optimal solution for optimizing the computing task start time, the air conditioning setpoint temperature, and the energy storage system charging and discharging power. The optimal solution constitutes the multi-resource collaborative scheduling scheme, wherein the intelligent optimization algorithm is a genetic algorithm.

8. The method according to claim 7, characterized in that, The process of parsing and executing the multi-resource collaborative scheduling scheme includes: At the first point in time, a pre-scheduling instruction for computing power tasks, parsed from the multi-resource collaborative scheduling scheme, is sent to the computing task management platform; Within a time window following the first time point, send the energy storage system pre-control command parsed from the multi-resource collaborative scheduling scheme to the energy storage management system; At a second time point later than the first time point, the air conditioning pre-adjustment command parsed from the multi-resource collaborative scheduling scheme is sent to the air conditioning control system, thereby realizing the coordination and timing linkage of the three resources.

9. The method according to claim 1, characterized in that, After the execution of the linkage control, the method further includes: Monitor and collect actual total load data and actual computer room ambient temperature data after the execution of linkage control; The actual total load data is compared with the total load prediction curve, and the actual computer room ambient temperature data is compared with the dynamic status of the heat flow network to generate model error feedback data. Using the model error feedback data, the power consumption and heat generation of the device group, as well as the thermal resistance and heat capacity parameters in the dynamic model of the heat flow network, are adaptively updated; and / or, The dynamic model of the heat flow network is trained by collecting historical operating data to construct a loss function and using gradient descent to optimize the heat capacity and thermal resistance parameters.

10. A multi-dimensional power monitoring and predictive control system for intelligent computing centers, applied to the multi-dimensional power monitoring and predictive control method for intelligent computing centers as described in any one of claims 1-9, characterized in that, The system is configured as follows: The acquisition module is used to acquire the device group operating status sequence from the current waveform data based on the power supply bus of the intelligent computing center; The generation module is used to obtain the queue of computing power tasks to be executed and the parameters of the computing power tasks; The computing power task parameters are mapped to the power consumption and heat of the device cluster. Based on the queue of computing power tasks to be executed and the power consumption and heat of the device group, a heat source timing input driven by the queue of computing power tasks to be executed is generated. The calculation module is used to establish a thermal resistance network for heat transfer between device nodes based on the thermal capacity parameters of the device node set divided by the physical space of the intelligent computing center; and to calculate the temperature state vector of all nodes based on the thermal capacity parameters and the thermal resistance network; wherein the thermal capacity parameters and the thermal resistance network are pre-calibrated by a combination of computational fluid dynamics simulation and system identification. The future time period heat flow network dynamic state generation module is used to input the temperature state vector of all nodes into the pre-trained heat flow network dynamic model and output the future time period heat flow network dynamic state. The power prediction sequence generation module is used to generate a power prediction sequence for the cooling system in future time periods based on the dynamic state of the heat flow network and the energy efficiency performance coefficient of the server cooling system of the intelligent computing center. The prediction curve construction module is used to construct a total load prediction curve based on the operating status sequence of the high-confidence equipment group and the power prediction sequence of the heat dissipation system in the future period. The scheduling scheme generation module is used to input the total load prediction curve into a multi-objective optimization game model for solving, and generate a multi-resource collaborative scheduling scheme that includes timestamps and control parameters. The scheduling module is used to generate the computing power task timing of the intelligent computing center, the air conditioning setting strategy of the intelligent computing center, and to perform linkage control on the actions of the equipment in the intelligent computing center in accordance with the multi-resource collaborative scheduling scheme.

11. A computer-readable storage medium, comprising: A computer program or instruction; when the computer program or instruction is run on a computer, it causes the computer to perform the intelligent computing center multi-dimensional power monitoring and predictive control method as described in any one of claims 1-9.

Images