Computer power consumption dynamic allocation method and system

Abstract

The present invention relates to the field of computer technology and discloses a method and system for dynamically allocating computer power consumption. The method comprises the following steps: acquiring data from sensors at various levels, preprocessing the data using a wavelet noise reduction algorithm, and obtaining preprocessed data; constructing a dynamic decision model, inputting the preprocessed data into the dynamic decision model, and outputting a load heat map for future time slices; collecting energy consumption data under different hardware configurations to obtain an energy consumption knowledge graph, performing federated learning using the DP-FedAvg algorithm, and generating a dynamic power supply curve by combining the load heat map, current hardware status, software operation trends, and historical energy consumption data; and employing quantized control technology to fine-tune power output according to the power supply curve and dynamically allocate power to various hardware components. The method significantly improves power utilization efficiency and reduces overall computer energy consumption.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a method and system for dynamically allocating computer power consumption. Background Art

[0002] As computer performance continues to improve, the power consumption requirements of its internal hardware components, such as CPU, GPU, hard disk, and memory, are becoming increasingly diverse. In traditional computer power supply modes, the power supply is usually statically configured according to the peak power consumption of the hardware components. As a result, during the actual operation of the computer, the hardware components are not in a fully loaded state most of the time, and a large amount of power is wasted and idle. At the same time, the cooling system also needs to be designed according to the peak power consumption, which increases costs and energy consumption. In addition, when multiple hardware components require high power operation at the same time, if the power supply cannot flexibly allocate power, insufficient power may occur, affecting the stability of the computer system and reducing the user experience. Summary of the Invention

[0003] The purpose of the present invention is to solve the above problems and to design a method and system for dynamically allocating computer power consumption.

[0004] A first aspect of the present invention provides a method for dynamically allocating computer power consumption, the method comprising the following steps:

[0005] Acquire data from sensors at all levels, pre-process the data using a wavelet noise reduction algorithm, and obtain pre-processed data;

[0006] Using the lightweight Transformer-XL as the basic architecture, combined with neural architecture search, it generates the optimal subnetwork structure in real time, performs hyperparameter optimization, builds a dynamic decision model, inputs preprocessed data, and predicts future load heat maps.

[0007] Integrate energy consumption data under different hardware configurations, build an energy consumption knowledge graph, use the DP-FedAvg algorithm for federated learning, and combine load heat maps, current hardware status, software operation trends, and historical energy consumption data to generate a dynamic power supply curve;

[0008] Using quantum control technology, it fine-tunes power output according to the power supply curve and dynamically distributes power to each hardware component.

[0009] Optionally, in a first implementation of the first aspect of the present invention, acquiring data from sensors at each level and preprocessing the data using a wavelet noise reduction algorithm to obtain preprocessed data includes:

[0010] Obtain sensor data, initialize the algorithm, and set the number of wavelet decomposition layers to 4-6;

[0011] Extract data segments in sequence for decomposition, the length of which is determined by the number of decomposition levels and sampling frequency;

[0012] Perform wavelet decomposition on the data segment using high-pass and low-pass filters to obtain detail and approximate components;

[0013] Perform threshold processing on high-frequency detail components to remove noise;

[0014] The processed components and low-frequency approximate components are used to restore the data through the wavelet reconstruction algorithm to obtain the preprocessed results.

[0015] Optionally, in a second implementation of the first aspect of the present invention, the lightweight Transformer-XL is used as the basic architecture, combined with neural architecture search, to generate the optimal subnet structure in real time, perform hyperparameter optimization, and build a dynamic decision model, including:

[0016] Use the lightweight Transformer-XL architecture, and set the number of neurons and multi-head attention heads to 4-8;

[0017] Integrate the neural architecture search engine into the Transformer-XL model in parallel, collect hardware and load information, use reinforcement learning algorithms to search, and initialize parameters;

[0018] Divide the search space into multiple subspaces, search independently and in parallel, and randomly generate an initial subnetwork structure as a starting point;

[0019] Deploy the subnetwork structure to the hardware environment, train and evaluate with load data, feed the evaluation indicators into the reinforcement learning algorithm, calculate the reward value, update the policy network, optimize the subnetwork structure, and select the optimal one;

[0020] The multi-objective Bayesian optimization algorithm is used to optimize the model hyperparameters, the results are applied to the subnetwork structure, and the genetic algorithm is used to dynamically optimize the hyperparameters to form a dynamic decision-making model.

[0021] Optionally, in a third implementation of the first aspect of the present invention, the method of using a multi-objective Bayesian optimization algorithm to optimize model hyperparameters, applying the results to the subnet structure, and dynamically optimizing the hyperparameters using a genetic algorithm to form a dynamic decision model includes:

[0022] Randomly select a set of hyperparameters, calculate the optimization target value, and use it as the initial sample set to train the Gaussian process model and estimate the parameters;

[0023] Select the next set of hyperparameters by maximizing the acquisition function, find the maximum point using a global search algorithm, calculate the target value, and update the sample set;

[0024] Retrain the model with the updated sample set, update the parameters, determine whether the maximum number of iterations has been reached, stop if reached, and output the hyperparameter combination;

[0025] Encode the hyperparameter combination as a chromosome and randomly generate the initial population, where each chromosome represents a set of hyperparameters;

[0026] Calculate the chromosome fitness value, select some chromosomes as parents, randomly select parents for crossover operation, and generate offspring chromosomes with a crossover probability of 0.6-0.8;

[0027] Perform mutation operations on offspring chromosomes, introduce new genetic information, update the population, determine whether the maximum number of iterations has been reached, stop if reached, output the hyperparameter combination corresponding to the chromosome with the highest fitness, and build a dynamic decision model.

[0028] Optionally, in a fourth implementation of the first aspect of the present invention, integrating energy consumption data under different hardware configurations to construct an energy consumption knowledge graph includes:

[0029] Collect energy consumption data of hardware configuration, clean it to remove abnormalities and missing values, and extract energy consumption characteristics;

[0030] The energy consumption characteristics are converted into graph nodes through vector encoding, and a unique identifier is assigned to the energy consumption pattern of each hardware-software combination;

[0031] Analyze the energy consumption changes of hardware components during software operation, and determine the energy consumption changes during collaborative working mode and migration;

[0032] Calculate the correlation coefficient of energy consumption characteristics between nodes to represent the strength of association, and create edges to form an energy consumption knowledge graph.

[0033] Optionally, in a fifth implementation of the first aspect of the present invention, the DP-FedAvg algorithm is used for federated learning, and a dynamic power supply curve is generated by combining a load heat map, current hardware status, software operation trends, and historical energy consumption data, including:

[0034] Set the parameters for federated learning, including the privacy budget, determine the number of communication rounds, and initialize the global model on the server side;

[0035] During each round of training, the current global model is downloaded from the server. The local load heat map, current hardware status, software operation trends, and energy consumption history data are used as input to train the model and calculate the gradient of the model parameters.

[0036] Add Laplace-distributed noise to the calculated gradient and upload it to the server. The server then performs a weighted summation of the uploaded gradients based on the proportion of each client's data volume to the total data volume. Add Laplace-distributed noise to the summation to obtain the final aggregated gradient.

[0037] The server uses the aggregated gradient to update the global model, and then sends the updated global model to each client.

[0038] After T rounds of federated learning, the client obtains the final global model. The client then inputs the local real-time load heat map, current hardware status, software operation trends, and historical energy consumption data into the global model.

[0039] The global model is used to predict the energy consumption requirements of each hardware component in the future. Based on the predicted energy consumption requirements, a dynamic power supply curve is generated with time as the horizontal axis and power as the vertical axis.

[0040] Optionally, in a sixth implementation of the first aspect of the present invention, the method of using quantized control technology to fine-tune power output according to a power supply curve and dynamically allocate power to each hardware component includes:

[0041] Extract the target power supply power of hardware components in each time period and form a power sequence table;

[0042] Identify and list hardware components that require power regulation and control, and monitor their operating status in real time;

[0043] Initialize controllers related to pulse width modulation-based quantum control technology;

[0044] Calculate every 10 microseconds to determine the target power supply of each hardware component and the required voltage and current adjustments;

[0045] The adjustment amount is quantized and the controller output signal is adjusted.

[0046] A second aspect of the present invention provides a computer power consumption dynamic allocation system, the system comprising:

[0047] The data aggregation module is used to obtain data from sensors at all levels and pre-process the data using a wavelet noise reduction algorithm to obtain pre-processed data;

[0048] The model building module uses the lightweight Transformer-XL as the basic architecture, combines it with neural architecture search, generates the optimal subnetwork structure in real time, performs hyperparameter optimization, builds a dynamic decision model, inputs preprocessed data, and predicts future load heat maps;

[0049] The curve generation module is used to integrate energy consumption data under different hardware configurations, build an energy consumption knowledge graph, and use the DP-FedAvg algorithm for federated learning. It combines load heat maps, current hardware status, software operation trends, and historical energy consumption data to generate dynamic power supply curves.

[0050] The dynamic adjustment module uses quantum control technology to fine-tune power output according to the power supply curve and dynamically distribute power to various hardware components.

[0051] Optionally, in a first implementation of the second aspect of the present invention, the data aggregation module includes:

[0052] Initialization submodule, used to obtain sensor data, initialize the algorithm, and set the number of wavelet decomposition layers to 4-6;

[0053] The first decomposition submodule is used to extract data segments in sequence for decomposition, and the length is determined by the number of decomposition layers and the sampling frequency;

[0054] A second decomposition submodule is used to perform wavelet decomposition on the data segment using high-pass and low-pass filters to obtain detail and approximate components;

[0055] The noise removal submodule is used to perform threshold processing on high-frequency detail components to remove noise;

[0056] The wavelet reconstruction submodule is used to use the processed components and low-frequency approximate components to restore the data through the wavelet reconstruction algorithm to obtain the pre-processed results.

[0057] Optionally, in a second implementation of the second aspect of the present invention, the model building module includes:

[0058] Select submodules to use the lightweight Transformer-XL architecture, and set the number of neurons and multi-head attention heads to 4-8;

[0059] The search submodule is used to integrate the neural architecture search engine into the Transformer-XL model in parallel, collect hardware and load information, use the reinforcement learning algorithm to search, and initialize parameters;

[0060] The partitioning submodule is used to divide the search space into multiple subspaces, search independently and in parallel, and randomly generate an initial subnetwork structure as a starting point;

[0061] The update submodule is used to deploy the subnetwork structure to the hardware environment, train and evaluate it with load data, feed the evaluation indicators back to the reinforcement learning algorithm, calculate the reward value, update the policy network, optimize the subnetwork structure, and select the optimal one;

[0062] The application submodule is used to optimize the model hyperparameters using the multi-objective Bayesian optimization algorithm, apply the results to the subnetwork structure, and use the genetic algorithm to dynamically optimize the hyperparameters to form a dynamic decision model.

[0063] In the technical solution provided by the present invention, data from sensors at all levels is obtained, and the wavelet noise reduction algorithm is used to preprocess the data to obtain preprocessed data; a lightweight Transformer-XL is used as the basic architecture, combined with neural architecture search, to generate the optimal subnet structure in real time, perform hyperparameter optimization, build a dynamic decision model, input preprocessed data, and predict future load heat maps; energy consumption data under different hardware configurations are integrated to build an energy consumption knowledge graph, and the DP-FedAvg algorithm is used for federated learning. A dynamic power supply curve is generated by combining the load heat map, current hardware status, software operation trends, and historical energy consumption data; quantum control technology is used to fine-tune the power output according to the power supply curve and dynamically allocate power to each hardware component; the present invention dynamically allocates power consumption according to actual conditions, avoids power waste in traditional static power supply methods, significantly improves power utilization efficiency, reduces the overall energy consumption of the computer, effectively improves the operating stability of the computer system, reduces problems such as freezes and freezes caused by insufficient power supply, provides users with a smooth user experience, extends the service life of the computer's internal hardware, and reduces maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Various other advantages and benefits will become apparent to those skilled in the art by reading the following detailed description of the preferred embodiment.The accompanying drawings are only for the purpose of illustrating the preferred embodiment and are not to be considered as limiting the present invention.

[0065] Figure 1 A flowchart of a method for dynamically allocating computer power consumption provided by an embodiment of the present invention;

[0066] Figure 2 This is a structural diagram of a computer power consumption dynamic allocation system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0067] The terms "first," "second," "third," "fourth," and so forth (if any) in the description and claims of the present invention and in the accompanying drawings are used to distinguish similar items and are not necessarily used to describe a particular order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that shown or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, apparatus, product, or device that includes a series of steps or elements is not necessarily limited to those steps or elements expressly listed, but may include other steps or elements not expressly listed or inherent to such process, method, product, or device.

[0068] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 A flowchart of a method for dynamically allocating computer power consumption is provided in an embodiment of the present invention. The method specifically includes the following steps:

[0069] Step 101: Acquire data from sensors at each level, and pre-process the data using a wavelet noise reduction algorithm to obtain pre-processed data;

[0070] In this embodiment, data from sensors at each level are obtained. Before wavelet noise reduction is started, the algorithm is initialized based on the Daubechies9 wavelet basis, and the number of wavelet decomposition layers is determined to be 4-6 layers. The aggregated and sorted data are extracted and decomposed in sequence, wherein the length of each data segment is determined by the number of wavelet decomposition layers sampling frequency. The extracted data segments are subjected to wavelet decomposition using high-pass and low-pass filter coefficients, and the data are sequentially decomposed into detail components and approximate components of different frequency scales. The high-frequency detail components obtained by decomposition are thresholded, and a soft threshold function is used to set elements in the detail components whose absolute values are less than the threshold to zero, and retain elements greater than the threshold after subtracting the threshold, so as to remove noise. After noise removal is completed, the processed detail components and the unaffected low-frequency approximate components are used to reversely execute the data decomposition process according to the wavelet reconstruction algorithm, and the components are recombined to restore the processed data to obtain preprocessed data.

[0071] In this embodiment, at the hardware level, current / voltage ripple sensors are precisely installed in locations such as the motherboard power supply lines, CPU core power pins, and GPU power modules. These sensors capture subtle changes in current and voltage in real time at a high-frequency sampling rate of 1MHz, reflecting the real-time power consumption dynamics of the hardware. Simultaneously, thermal imaging cameras are placed in areas such as the chipset and heat sink surface to non-contactly capture heat distribution images several times per second, visually displaying the hardware temperature status. At the system level, a process-level energy consumption tracker is embedded in the operating system kernel. This closely tracks the instruction per cycle (IPC) and cache miss rate of each thread running in the system, recording data at specific intervals (e.g., 10 milliseconds) to accurately locate resource consumption hotspots. At the user level, an eye tracker is installed in a suitable location on the display screen frame to track the movement and lingering area of the user's gaze on the screen at a refresh rate of no less than 60Hz. Pressure sensors are placed inside or on the bottom of the keyboard and mouse where they connect to the computer to monitor the pressure applied and the rate of change when the user operates the peripherals in real time, fully capturing the user's interaction intent.

[0072] Step 102: Use the lightweight Transformer-XL as the basic architecture, combine it with neural architecture search, generate the optimal subnetwork structure in real time, perform hyperparameter optimization, build a dynamic decision model, input preprocessed data, and predict the future load heat map;

[0073] In this embodiment, a lightweight Transformer-XL is used as the basic architecture, and a neural architecture search engine is embedded in parallel. The optimal subnet structure is dynamically generated in real time according to the current hardware configuration and operating load. The multi-objective Bayesian optimization algorithm and genetic algorithm are used to optimize the model's hyperparameters, and a dynamic decision model is constructed. The preprocessed data is input into the dynamic decision model to output a load heat map for future time slices.

[0074] In this embodiment, a lightweight Transformer-XL is selected as the basic architecture, the number of neurons and attention mechanism of each layer are initialized, and the number of heads of multi-head attention is set to 4-8; the neural architecture search engine is embedded in the built Transformer-XL basic model in parallel, hardware information and load information are collected, and a reinforcement learning algorithm is used for search, and the parameters of the reinforcement learning algorithm are initialized; the search space is divided into multiple subspaces, and searches are performed in these subspaces in parallel. Each search task runs independently on different computing resources to generate an initial architecture. Each search task randomly generates a set of initial subnet structures from the search space as the starting point of the search; the generated The subnet structure is deployed to the current hardware environment and trained and evaluated using the current operating load data. The evaluation indicators of each subnet structure are calculated and fed back to the reinforcement learning algorithm. The reward value is calculated based on the evaluation indicators, the policy network is updated, and a new subnet structure is generated based on the adjusted policy network. Evaluation and optimization continue. When the maximum number of searches is reached, the subnet structure with the best evaluation indicator is selected from all generated subnet structures. At the initial stage of model hyperparameter optimization, a multi-objective Bayesian optimization algorithm is used, and the hyperparameter combination obtained by the multi-objective Bayesian optimization algorithm is applied to the optimal subnet structure. A genetic algorithm is used to dynamically optimize the hyperparameters to obtain a dynamic decision model.

[0075] In this embodiment, an initial set of hyperparameter combinations is randomly selected from the range of hyperparameter values. For each set of hyperparameter combinations, the values of each optimization objective are calculated. These hyperparameter combinations and their corresponding target values are used as an initial sample set. Based on the data of the initial sample set, a Gaussian process model is trained to estimate the model parameters. Within the range of hyperparameter values, the next hyperparameter combination to be sampled is selected by maximizing the acquisition function. A global search algorithm is used to find the maximum point of the acquisition function. The selected hyperparameter combination is applied to the model to calculate the corresponding optimization objective value. The new hyperparameter combination and its target value are added to the sample set. The Gaussian process model is retrained using the updated sample set, the model parameters are updated, and it is determined whether the maximum number of iterations has been reached. If so, the optimization process is stopped and the hyperparameter combination obtained by the multi-objective Bayesian optimization algorithm is output. The hyperparameter combination obtained by the multi-objective Bayesian optimization algorithm is encoded and converted into a chromosome form. The length of the chromosome is determined, and a set of initial chromosomes is randomly generated to form an initial population, where each chromosome represents a set of hyperparameter combinations. The fitness value of each chromosome is calculated. Based on the fitness value of the chromosome, a portion of the chromosomes is selected as parents, and two chromosomes are randomly selected from the parents as parents. According to the crossover probability, the chromosomes of the parents are crossovered to generate offspring chromosomes, where the crossover probability is 0.6-0.8; the genes in the offspring chromosomes are mutated to introduce new genetic information, and the offspring chromosomes are added to the population to replace the chromosomes with lower fitness. The population is updated and it is determined whether the maximum number of iterations has been reached. If so, the optimization process is stopped and the hyperparameter combination corresponding to the chromosome with the highest fitness is output. A dynamic decision model is constructed based on the hyperparameter combination output by the genetic algorithm.

[0076] In this embodiment, a neural architecture search engine (NAS) is embedded in parallel with the established Transformer-XL basic model. NAS operates based on the principles of reinforcement learning and defines its action space, that is, it clearly defines which structural adjustments can be made to the basic model, such as adding or deleting specific types of layers, such as convolutional layers, adding or removing fully connected layers, adjusting the inter-layer connection method, changing the neuron activation function, etc. At the same time, a reward mechanism is established, which is closely related to the current hardware configuration and operating load. When the subnet structure generated by NAS performs well in adapting hardware resources, such as reducing memory usage, fully utilizing GPU computing power, and matching the operating load, such as improving the response speed to high load fluctuations and accurately predicting load changes, a higher reward is given; the hardware configuration information of the current computer system is collected in real time, covering the specific hardware model and core parameters such as the CPU's main frequency and number of cores, the GPU's video memory capacity and number of processing cores, as well as the connection relationship and collaborative working mode between various hardware components; at the same time, the real-time operating load of the system is continuously monitored. Through the reserved interface within the established basic model framework, it receives multi-source data from system-level energy consumption trackers, hardware-level sensors, etc., and integrates key indicators reflecting the operating load, such as CPU utilization, GPU load rate, and memory read and write frequency, to provide a basis for the subsequent dynamic generation of subnet structures; based on the collected hardware configuration and operating load information, combined with the preset action space and reward mechanism, NAS begins to dynamically search and generate the optimal subnet structure adapted to the current scenario. In each exploration attempt, it randomly or strategically selects actions in the action space to modify the base model, constructing a candidate subnetwork structure. It then feeds preprocessed data into this candidate subnetwork, observing how closely the model's predictions match the actual load, and rewards the network accordingly. After multiple iterations of exploration, the subnetwork structure with the highest reward score is selected and determined to be the optimal subnetwork for the current hardware and workload.

[0077] Step 103: Integrate energy consumption data under different hardware configurations, build an energy consumption knowledge graph, use the DP-FedAvg algorithm for federated learning, and combine the load heat map, current hardware status, software operation trend, and energy consumption history data to generate a dynamic power supply curve;

[0078] In this embodiment, energy consumption data under different hardware configurations are collected, and the energy consumption patterns of various hardware combinations in different software environments are encoded as graph nodes. The nodes are connected by edges to represent the hardware collaborative working characteristics or the energy consumption association caused by software migration, and an energy consumption knowledge graph is obtained. The DP-FedAvg algorithm is used for federated learning, and a dynamic power supply curve is generated by combining the load heat map, current hardware status, software operation trend and historical energy consumption data.

[0079] In this embodiment, energy consumption data under different hardware configurations are collected, outliers and missing values are removed from the energy consumption data, and energy consumption features are extracted from the cleaned data; vector encoding is used to convert the extracted energy consumption features into a representation of graph nodes, the energy consumption pattern of each hardware-software combination is encoded as a graph node, and a unique identifier is assigned to each node; the energy consumption changes of different hardware components during the operation of the same software are analyzed, and the collaborative working mode between hardware and software and the energy consumption changes during migration are determined; the correlation strength is represented by calculating the correlation coefficient between the energy consumption features of two nodes, and edges are created between the corresponding graph nodes based on the correlation relationship. All generated graph nodes and edges are combined together to form an energy consumption knowledge graph.

[0080] In this embodiment, federated learning parameters are set, including a privacy budget and the number of communication rounds, and a global model is initialized on the server. During each training round, the current global model is downloaded from the server. The model is trained using the local load heat map, current hardware status, software operation trends, and historical energy consumption data as input, and the gradients of the model parameters are calculated. Laplace-distributed noise is added to the calculated gradients and uploaded to the server. The server then performs a weighted summation of the uploaded gradients based on the proportion of each client's data volume to the total data volume, and adds Laplace-distributed noise to the sum to obtain the final aggregated gradient. The server uses the aggregated gradient to update the global model, which is then distributed to each client. After T rounds of federated learning, the client obtains the final global model. The client then inputs the local real-time load heat map, current hardware status, software operation trends, and historical energy consumption data into the global model. The global model is used to predict the energy consumption requirements of each hardware component over a period of time. Based on the predicted energy consumption requirements, a dynamic power supply curve is generated, with time as the horizontal axis and power as the vertical axis.

[0081] In this embodiment, the model's built-in feature extraction module performs deep feature extraction on the classified and sorted input data. For hardware-level current and voltage data, frequency domain features are extracted using methods such as Fourier transforms to capture potential periodic changes in power consumption. For temperature data, time series analysis methods are combined to extract features such as the rate of temperature change and fluctuation amplitude to assist in determining the thermal stability of the hardware. At the system level, process information is encoded using techniques similar to natural language processing, converting metrics such as thread IPC and cache miss rates into semantic feature vectors that the model can understand for fusion analysis with hardware-level features. User-level eye movement and pressure sensor data are extracted using a specially designed attention mechanism module, focusing on the hardware load potential corresponding to the user's current focus area and the load change direction implied by the operating pressure trend. All extracted features are ultimately encoded and integrated into a feature vector of uniform dimension and input into the model's core processing layer.

[0082] In this embodiment, based on the load variation patterns learned from historical data and the real-time situation reflected by the current input data, the load of each hardware component in future time slices (e.g., 10 time slices divided into 100-millisecond intervals within the next second) is predicted. The prediction results are presented in numerical form, with each time slice corresponding to a set of load prediction values for hardware components. The predicted values are converted into a load heat map using a visualization module. Color mapping rules are set, for example, using dark red to represent high load and light blue to represent low load. Based on the predicted values, the load of each hardware component in different time slices is plotted on the heat map in light and dark colors, clearly and intuitively showing the system's load distribution and change trends in the short future.

[0083] In this embodiment, the energy consumption knowledge graph contains rich information such as the energy consumption patterns of different hardware configurations in various software environments, the hardware collaborative working characteristics, and the energy consumption association caused by software migration, which provides important data support for the federated learning of the DP-FedAvg algorithm. In the federated learning process, the algorithm can better understand the intrinsic structure of the data based on these relationships and patterns in the knowledge graph, so as to more accurately perform model training and parameter updates; through the energy consumption knowledge graph, it is possible to clearly understand the energy consumption association between different hardware combinations and software environments. When analyzing in combination with the current hardware status and software operation trends, the knowledge graph It can help determine whether the current hardware status conforms to previous energy consumption patterns and whether software operation trends will cause changes in energy consumption correlations between hardware. For example, based on the energy consumption correlations caused by software migration recorded in the knowledge graph, it can predict possible hardware energy consumption changes under the current software operation trends, providing more comprehensive information for generating dynamic power supply curves. The generation of dynamic power supply curves requires comprehensive consideration of multiple factors such as load heat maps, current hardware status, software operation trends, and historical energy consumption data. As part of this, the energy consumption knowledge graph can provide additional constraints and references for the generation of dynamic power supply curves from the perspective of energy consumption patterns and correlations. It helps determine the appropriate power distribution strategy for different hardware-software combinations, so that the generated dynamic power supply curve can more accurately meet the system's energy consumption requirements and achieve efficient and energy-saving power supply management.

[0084] Step 104: Using quantum control technology, fine-tune the power output according to the power supply curve and dynamically distribute power to each hardware component.

[0085] In this embodiment, the quantized control technology based on pulse width modulation adjusts the supply voltage and current in small discrete amounts according to the dynamic power supply curve, and dynamically adjusts the power output distribution to each hardware component.

[0086] In this embodiment, the target power supply value corresponding to each hardware component in each time period is extracted based on the dynamic power supply curve, and these target power supply values are organized into a power sequence table arranged in chronological order; all hardware components in the current computer system that require power supply regulation are identified, and a corresponding component list is established, and the current operating status information of each hardware component is collected in real time; according to the specifications of the computer power supply and the power supply demand characteristics of the hardware components, the controller involved in the pulse width modulation-based quantum control technology is initialized and set; the adjustment calculation process is started every 10 microseconds, and the target power supply power of each hardware component in the corresponding time period is found based on the power sequence table combined with the current time, and the voltage and current adjustment amounts required to achieve the target power supply power in the current state are determined based on the current operating status information; the calculated adjustment amounts are quantized to obtain discrete adjustment steps, and the output signal of the controller is adjusted based on the calculated discrete adjustment steps of voltage and current.

[0087] See also Figure 2 , a schematic diagram of the structure of a computer power consumption dynamic allocation system provided by an embodiment of the present invention, the system includes:

[0088] The data aggregation module is used to obtain data from sensors at all levels and pre-process the data using a wavelet noise reduction algorithm to obtain pre-processed data;

[0089] The model building module uses the lightweight Transformer-XL as the basic architecture, combines it with neural architecture search, generates the optimal subnetwork structure in real time, performs hyperparameter optimization, builds a dynamic decision model, inputs preprocessed data, and predicts future load heat maps;

[0090] The curve generation module is used to integrate energy consumption data under different hardware configurations, build an energy consumption knowledge graph, and use the DP-FedAvg algorithm for federated learning. It combines load heat maps, current hardware status, software operation trends, and historical energy consumption data to generate dynamic power supply curves.

[0091] The dynamic adjustment module uses quantum control technology to fine-tune power output according to the power supply curve and dynamically distribute power to various hardware components.

[0092] In this embodiment, the data aggregation module includes:

[0093] Initialization submodule, used to obtain sensor data, initialize the algorithm, and set the number of wavelet decomposition layers to 4-6;

[0094] The first decomposition submodule is used to extract data segments in sequence for decomposition, and the length is determined by the number of decomposition layers and the sampling frequency;

[0095] A second decomposition submodule is used to perform wavelet decomposition on the data segment using high-pass and low-pass filters to obtain detail and approximate components;

[0096] The noise removal submodule is used to perform threshold processing on high-frequency detail components to remove noise;

[0097] The wavelet reconstruction submodule is used to use the processed components and low-frequency approximate components to restore the data through the wavelet reconstruction algorithm to obtain the pre-processed results.

[0098] In this embodiment, the model building module includes:

[0099] Select submodules to use the lightweight Transformer-XL architecture, and set the number of neurons and multi-head attention heads to 4-8;

[0100] The search submodule is used to integrate the neural architecture search engine into the Transformer-XL model in parallel, collect hardware and load information, use the reinforcement learning algorithm to search, and initialize parameters;

[0101] The partitioning submodule is used to divide the search space into multiple subspaces, search independently and in parallel, and randomly generate an initial subnetwork structure as a starting point;

[0102] The update submodule is used to deploy the subnetwork structure to the hardware environment, train and evaluate it with load data, feed the evaluation indicators back to the reinforcement learning algorithm, calculate the reward value, update the policy network, optimize the subnetwork structure, and select the optimal one;

[0103] The application submodule is used to optimize the model hyperparameters using the multi-objective Bayesian optimization algorithm, apply the results to the subnetwork structure, and use the genetic algorithm to dynamically optimize the hyperparameters to form a dynamic decision model.

[0104] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for dynamically allocating computer power consumption, characterized in that: The method comprises the following steps: Acquire data from sensors at all levels, pre-process the data using a wavelet noise reduction algorithm, and obtain pre-processed data; Using the lightweight Transformer-XL as the basic architecture, combined with neural architecture search, it generates the optimal subnetwork structure in real time, performs hyperparameter optimization, builds a dynamic decision model, inputs preprocessed data, and predicts future load heat maps. Integrate energy consumption data under different hardware configurations, build an energy consumption knowledge graph, use the DP-FedAvg algorithm for federated learning, and combine load heat maps, current hardware status, software operation trends, and historical energy consumption data to generate a dynamic power supply curve; Using quantum control technology, it fine-tunes the power output according to the power supply curve and dynamically distributes power to each hardware component; The lightweight Transformer-XL is used as the basic architecture, combined with neural architecture search, to generate the optimal subnetwork structure in real time, perform hyperparameter optimization, and build a dynamic decision model, including: Use the lightweight Transformer-XL architecture, and set the number of neurons and multi-head attention heads to 4-8; Integrate the neural architecture search engine into the Transformer-XL model in parallel, collect hardware and load information, use reinforcement learning algorithms to search, and initialize parameters; Divide the search space into multiple subspaces, search independently and in parallel, and randomly generate an initial subnetwork structure as a starting point; Deploy the subnetwork structure to the hardware environment, train and evaluate with load data, feed the evaluation indicators into the reinforcement learning algorithm, calculate the reward value, update the policy network, optimize the subnetwork structure, and select the optimal one; Use the multi-objective Bayesian optimization algorithm to optimize the model hyperparameters, apply the results to the subnetwork structure, and use the genetic algorithm to dynamically optimize the hyperparameters to form a dynamic decision model; The DP-FedAvg algorithm is used for federated learning, combining load heat maps, current hardware status, software operation trends, and historical energy consumption data to generate a dynamic power supply curve, including: Set the parameters for federated learning, including the privacy budget, determine the number of communication rounds, and initialize the global model on the server side; During each round of training, the current global model is downloaded from the server. The local load heat map, current hardware status, software operation trends, and energy consumption history data are used as input to train the model and calculate the gradient of the model parameters. Add Laplace-distributed noise to the calculated gradient and upload it to the server. The server then performs a weighted summation of the uploaded gradients based on the proportion of each client's data volume to the total data volume. Add Laplace-distributed noise to the summation to obtain the final aggregated gradient. The server uses the aggregated gradient to update the global model, and then sends the updated global model to each client. After T rounds of federated learning, the client obtains the final global model. The client then inputs the local real-time load heat map, current hardware status, software operation trends, and historical energy consumption data into the global model. The global model is used to predict the energy consumption requirements of each hardware component in the future. Based on the predicted energy consumption requirements, a dynamic power supply curve is generated with time as the horizontal axis and power as the vertical axis.

2. The computer power consumption dynamic allocation method according to claim 1, wherein: The data from sensors at each level are obtained and preprocessed using a wavelet noise reduction algorithm to obtain preprocessed data, including: Obtain sensor data, initialize the algorithm, and set the number of wavelet decomposition layers to 4-6; Extract data segments in sequence for decomposition, the length of which is determined by the number of decomposition levels and sampling frequency; Perform wavelet decomposition on the data segment using high-pass and low-pass filters to obtain detail and approximate components; Perform threshold processing on high-frequency detail components to remove noise; The processed components and low-frequency approximate components are used to restore the data through the wavelet reconstruction algorithm to obtain the preprocessed results.

3. The computer power consumption dynamic allocation method according to claim 1, wherein: The method uses a multi-objective Bayesian optimization algorithm to optimize model hyperparameters, applies the results to the subnet structure, and uses a genetic algorithm to dynamically optimize hyperparameters to form a dynamic decision model, including: Randomly select a set of hyperparameters, calculate the optimization target value, and use it as the initial sample set to train the Gaussian process model and estimate the parameters; Select the next set of hyperparameters by maximizing the acquisition function, find the maximum point using a global search algorithm, calculate the target value, and update the sample set; Retrain the model with the updated sample set, update the parameters, determine whether the maximum number of iterations has been reached, stop if reached, and output the hyperparameter combination; Encode the hyperparameter combination as a chromosome and randomly generate the initial population, where each chromosome represents a set of hyperparameters; Calculate the chromosome fitness value, select some chromosomes as parents, randomly select parents for crossover operation, and generate offspring chromosomes with a crossover probability of 0.6-0.8; Perform mutation operations on offspring chromosomes, introduce new genetic information, update the population, determine whether the maximum number of iterations has been reached, stop if reached, output the hyperparameter combination corresponding to the chromosome with the highest fitness, and build a dynamic decision model.

4. The computer power consumption dynamic allocation method according to claim 1, wherein: The energy consumption data under different hardware configurations are integrated to construct an energy consumption knowledge graph, including: Collect energy consumption data of hardware configuration, clean it to remove abnormalities and missing values, and extract energy consumption characteristics; The energy consumption characteristics are converted into graph nodes through vector encoding, and a unique identifier is assigned to the energy consumption pattern of each hardware-software combination; Analyze the energy consumption changes of hardware components during software operation, and determine the energy consumption changes during collaborative working mode and migration; Calculate the correlation coefficient of energy consumption characteristics between nodes to represent the strength of association, and create edges to form an energy consumption knowledge graph.

5. The computer power consumption dynamic allocation method according to claim 1, wherein: The quantum control technology used here fine-tunes the power output according to the power curve and dynamically distributes power to various hardware components, including: Extract the target power supply power of hardware components in each time period and form a power sequence table; Identify and list hardware components that require power regulation and control, and monitor their operating status in real time; Initialize controllers related to pulse width modulation-based quantum control technology; Calculate every 10 microseconds to determine the target power supply of each hardware component and the required voltage and current adjustments; The adjustment amount is quantized and the controller output signal is adjusted.

6. A computer power consumption dynamic allocation system, characterized in that: The system includes: The data aggregation module is used to obtain data from sensors at all levels and pre-process the data using a wavelet noise reduction algorithm to obtain pre-processed data; The model building module uses the lightweight Transformer-XL as the basic architecture, combines it with neural architecture search, generates the optimal subnetwork structure in real time, performs hyperparameter optimization, builds a dynamic decision model, inputs preprocessed data, and predicts future load heat maps; The curve generation module is used to integrate energy consumption data under different hardware configurations, build an energy consumption knowledge graph, and use the DP-FedAvg algorithm for federated learning. It combines the load heat map, current hardware status, software operation trend, and energy consumption history data to generate a dynamic power supply curve: set the parameters of federated learning, including privacy budget, determine the number of communication rounds, initialize the global model on the server, download the current global model from the server in each round of training, train the model with the local load heat map, current hardware status, software operation trend, and energy consumption history data as input, calculate the gradient of the model parameters, add noise that conforms to the Laplace distribution to the calculated gradient, and upload the gradient with added noise to The server performs a weighted summation of the uploaded gradients based on the proportion of each client's data volume to the total data volume. It then adds noise that conforms to the Laplace distribution to the summation result to obtain the final aggregated gradient. The server uses the aggregated gradient to update the global model and sends the updated global model to each client. After T rounds of federated learning, the client obtains the final global model. The client inputs the local real-time load heat map, current hardware status, software operation trend, and energy consumption history data into the global model. The global model is used to predict the energy consumption requirements of each hardware component in the future. Based on the predicted energy consumption requirements, a dynamic power supply curve is generated with time as the horizontal axis and power as the vertical axis. Dynamic adjustment module, which uses quantum control technology to fine-tune power output according to the power supply curve and dynamically distribute power to various hardware components; The model building module includes: Select submodules to use the lightweight Transformer-XL architecture, and set the number of neurons and multi-head attention heads to 4-8; The search submodule is used to integrate the neural architecture search engine into the Transformer-XL model in parallel, collect hardware and load information, use the reinforcement learning algorithm to search, and initialize parameters; The partitioning submodule is used to divide the search space into multiple subspaces, search independently and in parallel, and randomly generate an initial subnetwork structure as a starting point; The update submodule is used to deploy the subnetwork structure to the hardware environment, train and evaluate it with load data, feed the evaluation indicators back to the reinforcement learning algorithm, calculate the reward value, update the policy network, optimize the subnetwork structure, and select the optimal one; The application submodule is used to optimize the model hyperparameters using the multi-objective Bayesian optimization algorithm, apply the results to the subnetwork structure, and use the genetic algorithm to dynamically optimize the hyperparameters to form a dynamic decision model.

7. The computer power consumption dynamic allocation system according to claim 6, wherein: The data aggregation module includes: Initialization submodule, used to obtain sensor data, initialize the algorithm, and set the number of wavelet decomposition layers to 4-6; The first decomposition submodule is used to extract data segments in sequence for decomposition, and the length is determined by the number of decomposition layers and the sampling frequency; A second decomposition submodule is used to perform wavelet decomposition on the data segment using high-pass and low-pass filters to obtain detail and approximate components; The noise removal submodule is used to perform threshold processing on high-frequency detail components to remove noise; The wavelet reconstruction submodule is used to use the processed components and low-frequency approximate components to restore the data through the wavelet reconstruction algorithm to obtain the pre-processed results.

Images