AI prediction-based pc power supply load adjustment method, device and medium

Abstract

The application provides a PC power supply load adjustment method and device based on AI prediction and a medium. By continuously collecting power supply signals at a PC power supply output end, processor running state signals and graphic component working state signals, load time series data is generated. After pattern recognition, power consumption signatures corresponding to different computing tasks are separated to obtain a power consumption signature set. The load prediction model is input into a pre-trained load prediction model. Through time series analysis and processing, the load intensity change profile in the future power supply cycle is obtained. The load intensity change profile is fed back and calibrated with the current double-path power supply state information of the PC power supply to generate a power supply power distribution strategy. The working points of the primary conversion path and the secondary conversion path in the PC power supply are adjusted according to the power supply power distribution strategy, so that the power supply output power and the load change are dynamically matched. The application reduces unnecessary power loss, thereby ensuring stable operation of the PC system while realizing fine and dynamic adjustment of the power supply load.

Description

Technical Field

[0001] This application relates to the field of data processing, and in particular to a method, device and medium for PC power load regulation based on AI prediction. Background Technology

[0002] With the development of computer technology, adjusting power supply output parameters can adapt power supply capacity to system load requirements. The core objective is to achieve dynamic power output management while ensuring stable hardware operation. Currently, common PC power supply load regulation technologies typically employ real-time feedback-based mechanisms. These mechanisms collect electrical signals such as voltage and current at the power output terminal and combine them with hardware protection circuits or simple control algorithms to dynamically adjust output power, thus addressing dynamic load changes to some extent. However, this type of regulation relies primarily on passive responses to the current electrical signals, resulting in regulation lag. When the load fluctuates rapidly, instantaneous power mismatch can easily occur. Furthermore, because it fails to fully consider the inherent differences in power consumption requirements across different computing tasks, the regulation accuracy is limited, making it difficult to optimize power efficiency in complex computing scenarios. Summary of the Invention

[0003] This invention provides a PC power load regulation method, device, and medium based on AI prediction.

[0004] In a first aspect, embodiments of the present invention provide a PC power supply load adjustment method based on AI prediction. The method includes: continuously collecting power supply signals, processor operating status signals, and graphics component operating status signals from the PC power supply output terminal; associating and integrating the collected signals according to time series to generate load time series data, wherein the load time series data includes signal acquisition time stamps and time series change association information of each signal; performing pattern recognition on the load time series data to separate power consumption signatures corresponding to different computing tasks, thereby obtaining a power consumption signature set associated with the load tasks; inputting the power consumption signature set associated with the load tasks into a pre-trained load prediction model, and obtaining a load intensity change profile within future power supply cycles through time series analysis processing, wherein the load intensity change profile includes time series regularity information of load changes; performing feedback calibration between the load intensity change profile and the current dual-path power supply status information of the PC power supply to generate a power allocation strategy, wherein the power allocation strategy adapts to the time series regularity of the load intensity change profile; and adjusting the operating points of the primary conversion path and the secondary conversion path inside the PC power supply according to the power allocation strategy to ensure dynamic matching between the power supply output power and load changes.

[0005] Secondly, embodiments of the present invention provide a computer device, including: a memory for storing computer-executable instructions or computer programs; and a processor for executing the computer-executable instructions or computer programs stored in the memory to implement the above-described AI-predictive PC power load regulation method.

[0006] Thirdly, embodiments of this application provide a computer-readable storage medium storing a computer program, which is loaded and executed by a processor to implement the AI-predictive PC power load regulation method described above.

[0007] This invention continuously collects power supply signals, processor operating status signals, and graphics component operating status signals from the PC power supply output terminal and integrates them according to time series to generate load time-series data. This comprehensively captures the multi-dimensional dynamic characteristics of the power load, providing a more complete data foundation for subsequent load analysis. By performing pattern recognition on the load time-series data to separate the power consumption signatures corresponding to different computing tasks, it achieves precise binding between computing tasks and power consumption characteristics, enabling load identification to delve from surface signal statistics to the essential characteristics of the tasks. Inputting the set of power consumption signatures associated with the load tasks into the load prediction model yields a load intensity change profile containing time-series pattern information, effectively capturing the dynamic change pattern of the load over time and providing a forward-looking basis for power supply adjustment. Through feedback calibration of the load intensity change profile and dual-path power supply status information, a power allocation strategy adapted to the time-series pattern is generated, achieving dynamic matching between load demand and power supply capacity, avoiding resource waste or insufficient power supply caused by static allocation. Adjusting the operating points of the primary and secondary conversion paths according to the power allocation strategy ensures that the power hardware parameters adapt to load changes in real time, effectively improving power energy conversion efficiency and reducing unnecessary power loss. Thus, while ensuring the stable operation of the PC system, it achieves refined and dynamic adjustment of the power load. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the structure of the computer device provided in the embodiments of this application;

[0009] Figure 2 This is a flowchart illustrating the PC power load regulation method based on AI prediction provided in an embodiment of this application. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0011] The following describes a computer device that implements the AI-predictive PC power load regulation method provided in the embodiments of this application. See also Figure 1 This is a schematic diagram of the structure of the computer device provided in the embodiments of this application. Figure 1 The computer device shown includes at least one processor 210, memory 250, at least one network interface 220, and external interface 230. The various components in the computer device 200 are coupled together via a bus system 240. It is understood that the bus system 240 is used to implement communication between these components. In addition to a data bus, the bus system 240 also includes a power bus, a control bus, and a status signal bus. However, for clarity, ... Figure 1 The general labeled all buses as Bus System 240.

[0012] Processor 210 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0013] External interface 230 may include, for example, one or more speakers and / or one or more visual displays. External interface 230 may also include one or more input devices 432, such as a keyboard, mouse, microphone, touchscreen display, camera, etc. Memory 250 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state memory, hard disk drives, optical disk drives, etc. Memory 250 may optionally include one or more storage devices physically located remote from processor 210.

[0014] The memory 250 may include volatile memory or non-volatile memory, or both. The non-volatile memory may be read-only memory (ROM), and the volatile memory may be random access memory (RAM). The memory 250 described in this application embodiment is intended to include any suitable type of memory.

[0015] In some embodiments, memory 250 is capable of storing data to support various operations, examples of which include programs, modules, and data structures or subsets or supersets thereof, as illustrated below.

[0016] Operating system 251 includes system programs for handling various basic system services and performing hardware-related tasks, such as framework layer, core library layer, driver layer, etc., for implementing various basic business functions and handling hardware-based tasks; network communication module 252 is used to reach external devices via one or more (wired or wireless) network interfaces 220, exemplary network interfaces 220 include: Bluetooth, WiFi, and Universal Serial Bus (USB), etc.; presentation module 253 is used to enable the display of information (e.g., external interfaces for operating peripheral devices and displaying content and information) via one or more output devices 231 associated with external interface 230 (e.g., display screen, speaker, etc.); input processing module 254 is used to detect and translate one or more user inputs or interactions from one or more input devices 232.

[0017] See Figure 2 This is a flowchart illustrating the PC power load regulation method based on AI prediction provided in this application embodiment. Next, we will combine... Figure 2 The steps shown are explained.

[0018] Step S100: Continuously collect the power supply signal, processor running status signal and graphics component working status signal from the PC power output terminal. The collected signals are correlated and integrated according to the time sequence to generate load timing data. The load timing data includes signal acquisition time markers and timing change correlation information of each signal.

[0019] Power supply signals reflect the power supply status at the PC power supply output terminal, indicating the power output, voltage, and current. Processor operating status signals present the real-time operating status of the processor, including metrics such as processor utilization, core frequency, and temperature. Graphics component operating status signals describe the operation of the graphics component, such as graphics processor utilization, video memory usage, and frame rate.

[0020] For power supply signals, current sensors are used to measure current values, and voltage sensors are used to acquire voltage values. These sensors capture changes in current and voltage in real time and convert them into analyzable electrical signals. Processor operating status signals can be acquired through the processor's built-in monitoring module. During the acquisition process, signals are recorded in chronological order, forming independent time series. Each data point in the time series corresponds to a specific acquisition time. These acquired signals are correlated and integrated according to their time series, combining power supply signals, processor operating status signals, and graphics component operating status signals acquired at the same time point. During integration, a precise signal acquisition time stamp is added to each data point. The final generated load timing data not only contains the specific values ​​of each signal but also the temporal correlation information between them.

[0021] Step S200: Perform pattern recognition on the load time series data, separate the power consumption signatures corresponding to different computing tasks, and obtain a power consumption signature set associated with the load tasks. Each power consumption signature in the power consumption signature set has a unique correspondence with a specific computing task.

[0022] In some embodiments, step S200 may specifically include the following steps S210-S260:

[0023] Step S210: Perform event-triggered segmentation on the load timing data, monitor the abrupt change points of the power supply signal, processor running status signal and graphics component working status signal respectively, comprehensively determine the load state switching time, and segment the continuous load timing data at the state switching time to obtain multiple independent event-driven segment units. Each event-driven segment unit contains a complete signal change cycle from the start of one state switching time to the end of the next state switching time.

[0024] In load timing data, power supply signals, processor operating status signals, and graphics component operating status signals change as the computing task changes. When the computing task switches, these signals often exhibit significant abrupt changes. For example, the power supply signal may suddenly increase or decrease, the processor utilization may rise or fall sharply, and the frame rate of the graphics component may fluctuate instantaneously.

[0025] To monitor these abrupt changes, appropriate monitoring methods are needed. For power supply signals, abrupt changes can be determined by calculating the rate of change of power. When the rate of change of power exceeds a pre-set threshold, an abrupt change is considered to have occurred. For processor operating status signals and graphics component operating status signals, the magnitude of changes in their key indicators can be monitored. For example, when the processor utilization rate changes beyond a certain range, it is considered an abrupt change; when the frame rate fluctuation of the graphics component exceeds a specific value, it is also considered an abrupt change.

[0026] When comprehensively determining the load state switching moment, it is insufficient to rely solely on the abrupt change of a single signal. While a single signal abrupt change might be due to noise or other accidental factors, the simultaneous abrupt changes of multiple signals are more likely caused by a switching of computing tasks. Therefore, it is necessary to consider the abrupt changes of power supply signals, processor operating status signals, and graphics component operating status signals simultaneously. When multiple signals from these three categories simultaneously exhibit abrupt changes, a load state switching moment can be determined; this point in time is the load state switching moment. After determining the load state switching moment, the continuous load timing data is segmented at these moments to obtain multiple independent event-driven segment units. Each event-driven segment unit contains a complete signal change cycle from the start of one state switching moment to the end of the next. Within this cycle, the signal changes reflect the power consumption characteristics during the execution of the computing task.

[0027] Step S220: Perform multi-domain feature fusion on each event-driven segment unit, simultaneously extract time-domain waveform features, frequency-domain spectrum features, and signal correlation features, and generate a three-dimensional feature tensor. The dimension of the three-dimensional feature tensor is consistent with the number of signal types, the number of feature types, and the segment length.

[0028] In some embodiments, step S220 may specifically include the following steps S221-S225:

[0029] Step S221: Extract time-domain waveform features from various signals in the event-driven segment unit. By calculating the rise time, fall time, pulse width, and duty cycle of the signal sequence, a time-domain feature vector describing the time-domain characteristics of the signal is generated. The number of elements in the time-domain feature vector is the same as the number of signal types.

[0030] Time-domain waveform feature extraction involves in-depth analysis of the power supply signals, processor operating status signals, and graphics component operating status signals within the event-driven segment unit to obtain their characteristics in the time domain. Rise time is the time required for a signal to rise from its initial value to a specific value (usually a certain percentage of the peak value), reflecting the signal's rise speed and crucial for determining its response characteristics. Fall time is the time required for a signal to fall from its peak value to a specific value, reflecting the signal's decay rate. Pulse width is the duration of a signal being at a high level, reflecting the effective duration of the signal. Duty cycle is the ratio of pulse width to the signal period, representing the proportion of the signal that is at a high level within one period.

[0031] For each signal type, its rise time, fall time, pulse width, and duty cycle need to be calculated separately. Taking power supply signals as an example, the initial value and peak value of the signal must first be determined. By analyzing the signal waveform, the starting point of the signal rising from the initial value and the ending point of reaching the peak value are found, and the time difference between these two points is calculated; this is the rise time. Similarly, the starting point and ending point of the signal falling from the peak value to a specific value are found, and the time difference is calculated to obtain the fall time. The pulse width is determined by observing the time period during which the signal is at a high level. Then, based on the signal period and pulse width, the duty cycle is calculated. A similar method is used to calculate processor running status signals and graphics component operating status signals.

[0032] These calculated parameters are combined to form a time-domain feature vector. The number of elements in the time-domain feature vector is the same as the number of signal types; that is, each signal type corresponds to one time-domain feature vector element. These elements represent the time-domain characteristics of different signals, and the dynamic changes of each signal can be intuitively understood through the time-domain feature vector. For example, the time-domain feature vector elements of a power supply signal can reflect the stability and response speed of the power supply output; the time-domain feature vector elements of a processor operating status signal can reflect the processor's working rhythm and load changes; and the time-domain feature vector elements of a graphics component operating status signal can show the graphics component's working efficiency and performance fluctuations.

[0033] Step S222: Extract frequency domain spectral features from various signals in the event-driven segment unit. By converting the time-domain signal to the frequency domain, calculate the amplitude and phase of the signal at multiple preset frequency points, and generate a frequency domain feature vector describing the frequency domain characteristics of the signal. The number of elements in the frequency domain feature vector is twice the number of preset frequency points.

[0034] Frequency domain spectral feature extraction is the process of converting power supply signals, processor operating status signals, and graphics component operating status signals from the event-driven segment unit from the time domain to the frequency domain for analysis. Time-domain signals describe how a signal changes over time, while frequency-domain signals show the distribution of the signal across different frequency components. Converting a time-domain signal to the frequency domain requires specialized algorithms, such as the Fourier transform. The Fourier transform decomposes a time-domain signal into a superposition of multiple sine and cosine signals of different frequencies, thus obtaining the frequency-domain representation of the signal. In performing the Fourier transform, the signals from the event-driven segment unit are used as input, and after a series of calculations and processing, the frequency-domain representation of the output signal is obtained.

[0035] After obtaining the frequency domain signal, it is necessary to calculate the amplitude and phase of the signal at multiple preset frequency points. These preset frequency points are a pre-determined set of frequency values, typically selected based on actual requirements and signal characteristics. For each preset frequency point, the amplitude and phase of the signal at that point are calculated. The amplitude represents the signal strength at that frequency point, reflecting the importance of that frequency component in the signal. The phase represents the signal's relative position at that frequency point, and is crucial for determining the signal's waveform and phase relationship.

[0036] The calculated amplitude and phase at each preset frequency point are combined to generate a frequency domain feature vector. Since each frequency point corresponds to one amplitude and one phase, the number of elements in the frequency domain feature vector is twice the number of preset frequency points. The frequency domain feature vector can reflect the frequency composition and distribution of a signal. For example, for power supply signals, the frequency domain feature vector can show the content of different frequency components in the power supply output, helping to detect problems such as harmonic interference in the power supply; for processor operating status signals and graphics component operating status signals, the frequency domain feature vector can reveal the periodic changes and frequency characteristics during their operation.

[0037] Step S223: Perform cross-correlation analysis on different types of signals in the event-driven fragment unit, calculate the cross-correlation coefficients between the power supply signal and the processor running status signal, the graphics component working status signal, and the processor running status signal and the graphics component working status signal, respectively, and generate a correlation feature vector describing the correlation characteristics between signals. The number of elements in the correlation feature vector is the number of signal type combinations.

[0038] In event-driven fragment units, power supply signals, processor operating status signals, and graphics component operating status signals are interconnected and correlated. The degree of correlation between these signals can be quantified by calculating their cross-correlation coefficients. The cross-correlation coefficient ranges from -1 to 1. A cross-correlation coefficient of 1 indicates that the two signals are perfectly positively correlated, meaning their trends are completely identical; a cross-correlation coefficient of -1 indicates that the two signals are perfectly negatively correlated, meaning their trends are completely opposite; and a cross-correlation coefficient of 0 indicates that there is no linear correlation between the two signals.

[0039] This step requires calculating the cross-correlation coefficients between the power supply signal and the processor's operating status signal, the graphics component's operating status signal, and the processor's operating status signal and the graphics component's operating status signal, respectively. Specifically, the signals in the event-driven segment unit are processed, and the power supply signal and the processor's operating status signal are compared and analyzed to calculate their cross-correlation coefficients. Similarly, the power supply signal and the graphics component's operating status signal are compared and their cross-correlation coefficients are calculated; the processor's operating status signal and the graphics component's operating status signal are compared and their cross-correlation coefficients are calculated.

[0040] The three calculated cross-correlation coefficients are combined to generate an association feature vector. The number of elements in the association feature vector is the number of signal type combinations, i.e., 3 elements. This vector can reflect the correlation characteristics between different signals. For example, the degree of correlation between the power supply signal and the processor operating status signal can reflect the impact of the processor's operation on the power supply, and the degree of correlation between the processor operating status signal and the graphics component's operating status signal can show the collaborative work between the processor and the graphics component.

[0041] Step S224: Organize the time-domain feature vector, frequency-domain feature vector, and associated feature vector according to feature type, and use them as independent feature groups to generate a combined feature matrix containing the three feature types. The row dimension of the combined feature matrix corresponds to the number of signal types, and the column dimension is divided into three independent regions, which correspond to the element sets of the time domain, frequency domain, and associated feature groups, respectively. The number of columns in each region is equal to the number of elements in the corresponding feature group.

[0042] In the preceding steps, we have extracted time-domain feature vectors, frequency-domain feature vectors, and correlation feature vectors, respectively. To better integrate these features, they need to be organized according to feature type. Time-domain feature vectors represent the characteristics of a signal in the time domain, frequency-domain feature vectors reflect the characteristics of a signal in the frequency domain, and correlation feature vectors describe the correlation relationships between different signals.

[0043] These three feature vectors are treated as independent feature groups. The time-domain feature group contains the feature information of the power supply signal, the processor running status signal, and the graphics component working status signal in the time domain; the frequency-domain feature group contains the feature information of these signals in the frequency domain; and the correlation feature group contains the correlation feature information between different signals.

[0044] These three feature groups are combined to generate a combined feature matrix containing three feature types. The row dimension of the combined feature matrix corresponds to the number of signal types, i.e., there are three rows, corresponding to power supply signals, processor operating status signals, and graphics component operating status signals, respectively. The column dimension is divided into three independent regions, corresponding to the time domain, frequency domain, and the element set of the associated feature group, respectively. The number of columns in each region is equal to the number of elements in the corresponding feature group.

[0045] In this way, the combined feature matrix can clearly display the characteristic information of different signals under different feature types, facilitating subsequent analysis and processing. For example, by observing the element values ​​in different regions of the combined feature matrix, we can understand the characteristics of the signal in the time domain, frequency domain, and correlation relationships, providing a foundation for further feature selection and cluster analysis.

[0046] Step S225: Reconstruct the dimensions of each independent feature group in the combined feature matrix, adjust the feature arrangement according to the order of the number of signal types, the number of feature types, and the number of sampling points of the event-driven segment unit, generate a three-dimensional feature tensor, and make each dimension of the three-dimensional feature tensor correspond to the signal type, feature type, and time sampling point, and keep the dimension size consistent with the number of corresponding attributes.

[0047] After obtaining the combined feature matrix, it is necessary to reconstruct its dimensions in order to more comprehensively represent the feature information of the event-driven segment units. The purpose of dimensional reconstruction is to convert the two-dimensional combined feature matrix into a three-dimensional feature tensor, so as to more clearly show the relationship between different signals, different feature types, and different time sampling points.

[0048] Specifically, the adjustments are made in the following order: number of signal types, number of feature types, and number of sampling points in the event-driven segment unit. The number of signal types is three: power supply signals, processor operating status signals, and graphics component operating status signals. The number of feature types is three: time domain, frequency domain, and associated features. The number of sampling points in the event-driven segment unit depends on the signal sampling frequency and the segment duration.

[0049] During the reconstruction process, the elements in the combined feature matrix are rearranged so that the first dimension of the three-dimensional feature tensor corresponds to the signal type, the second dimension to the feature type, and the third dimension to the time sampling point. In this way, each element of the three-dimensional feature tensor can be uniquely identified by its indices in the three dimensions, representing the feature value of a certain signal type or feature type at a certain time sampling point.

[0050] The dimensionality of the 3D feature tensor is consistent with the number of corresponding attributes. Specifically, the first dimension has a size of 3, representing the three signal types; the second dimension has a size of 3, representing the three feature types; and the third dimension represents the number of sampling points for the event-driven fragment unit. In this way, the 3D feature tensor can more intuitively display the feature information of the event-driven fragment unit, providing richer input for subsequent feature selection and clustering analysis.

[0051] Step S230: Perform feature selection on the three-dimensional feature tensor. By calculating the mutual information value between each feature dimension and the computation task type, retain the feature dimensions with mutual information values ​​higher than a preset threshold, and generate a dimension-reduced feature matrix. The row dimension of the dimension-reduced feature matrix corresponds to the number of event-driven fragment units, and the column dimension corresponds to the number of filtered feature dimensions.

[0052] Feature selection aims to filter out features from a three-dimensional feature tensor that play an important role in distinguishing different types of computational tasks, thereby reducing the dimensionality of the data. Mutual information is used to calculate the correlation between each feature dimension and the type of computational task.

[0053] Calculate the mutual information value between each feature dimension and the computation task type. That is, for each feature dimension in the three-dimensional feature tensor, analyze the degree of correlation between it and the known computation task type. The higher the mutual information value, the stronger the correlation between the feature dimension and the computation task type, which means that the feature is more helpful in distinguishing different computation task types.

[0054] The preset threshold is a pre-defined standard used to determine whether a certain feature dimension is worth retaining. When the mutual information value of a feature dimension is higher than the preset threshold, it is considered that the feature dimension is of great significance in distinguishing the types of computing tasks, and it is retained; conversely, if the mutual information value is lower than the preset threshold, it is considered that the feature dimension is not very useful in distinguishing the types of computing tasks, and it is discarded.

[0055] After filtering out feature dimensions with mutual information values ​​higher than a preset threshold, these retained feature dimensions are combined to generate a dimensionality-reduced feature matrix. The row dimension of the dimensionality-reduced feature matrix corresponds to the number of event-driven fragment units, meaning each event-driven fragment unit corresponds to one row in the dimensionality-reduced feature matrix. The column dimension corresponds to the number of filtered feature dimensions, that is, it only includes those feature dimensions that are considered to be of great importance in distinguishing the types of computational tasks.

[0056] Step S240: Perform density peak clustering on the dimensionality-reduced feature matrix. By identifying regions in the feature space where the local feature point distribution density exceeds a preset density threshold as cluster centers, assign event-driven fragment units that are less than the truncation distance from the cluster center to the corresponding clusters to obtain multiple clusters. Each cluster contains a set of event-driven fragment units with similar features.

[0057] In some embodiments, step S240 may specifically include the following steps S241-S246:

[0058] Step S241: Perform pairwise distance calculations on the feature vectors of all event-driven fragment units in the dimensionality-reduced feature matrix. Obtain the distance values ​​between all feature point pairs by calculating the Euclidean distance, and generate a symmetric distance matrix. The rows and columns of the distance matrix correspond to event-driven fragment units, and the matrix elements are the feature distances between corresponding two fragment units.

[0059] Before performing density peak clustering, it is necessary to calculate the distance between the feature vectors of all event-driven fragment units in the dimensionality-reduced feature matrix. To calculate the Euclidean distance, corresponding elements of two feature vectors are subtracted, the difference is squared, all squared values ​​are summed, and the square root is taken to obtain the distance value. Specifically, for two feature vectors, their dimensions are first determined, and then the elements in each dimension are subtracted sequentially. The resulting difference is squared, and all squared values ​​are summed. Finally, the square root of the summation result gives the Euclidean distance between the two feature vectors.

[0060] For each event-driven fragment unit eigenvector in the reduced feature matrix, pairwise combinations are performed to calculate the Euclidean distance between them. The calculated distance values ​​are stored in a matrix, which is the distance matrix. Rows and columns of the distance matrix correspond to event-driven fragment units, and each element in the matrix represents the feature distance between two corresponding fragment units. Because distance is symmetric—that is, the distance between two eigenvectors is equal to the distance after swapping their order—the distance matrix is ​​a symmetric matrix.

[0061] Step S242: Determine the cutoff distance based on the distance matrix. By statistically analyzing the frequency distribution of all distance values, select the distance value that gives each feature point a preset number of neighbors on average as the cutoff distance. The cutoff distance is used to define the local neighborhood range.

[0062] The purpose of the cutoff distance is to define the local neighborhood range, which is used in subsequent clustering processes to determine which feature points belong to the neighborhood of a certain cluster center. To determine the cutoff distance, it is necessary to statistically analyze the frequency distribution of all distance values ​​in the distance matrix, that is, to count the number of times each distance value appears. Then, a frequency distribution curve is plotted based on these counts. By observing the frequency distribution curve, we can understand the distribution of distance values.

[0063] The preset number is a pre-defined standard representing the desired average number of neighbors per feature point. Based on this preset number, a suitable distance value is sought within the frequency distribution curve. Specifically, starting from the smaller end of the distance value range, the distance value is gradually increased, and the number of neighbors for each feature point within that distance range is counted. When the average number of neighbors per feature point reaches the preset number, the corresponding distance value is selected as the cutoff distance. Once the cutoff distance is determined, it can be used to define the local neighborhood range.

[0064] Step S243: Calculate the local feature point distribution density for each event-driven fragment unit feature vector, count the number of other feature vectors whose distance from the feature vector is less than the truncation distance, use the counted number as the local feature point distribution density of the event-driven fragment unit, and generate a density vector. The number of elements in the density vector is the same as the number of event-driven fragment units.

[0065] After determining the cutoff distance, the local feature point distribution density of each event-driven fragment unit's feature vector is calculated. This local feature point distribution density reflects the density of feature points surrounding a given feature point. For each event-driven fragment unit's feature vector, using it as the center and the cutoff distance as the radius, the number of other feature vectors within this range is counted. Specifically, the rows or columns corresponding to the feature vector in the distance matrix are traversed, elements with distance values ​​less than the cutoff distance are identified, and the number of these elements is counted. This count represents the local feature point distribution density of that event-driven fragment unit.

[0066] The local feature point distribution density of each event-driven fragment unit is combined to generate a density vector. The number of elements in the density vector is the same as the number of event-driven fragment units, and each element corresponds to the local feature point distribution density of an event-driven fragment unit.

[0067] Step S244: Perform relative distance calculation on the feature vector of each event-driven fragment unit, find all other fragment units with higher density than the current fragment unit, calculate the minimum distance between the current fragment unit and these high-density fragment units as the relative distance, and if the current fragment unit has the highest density, use the maximum distance value as its relative distance to generate a relative distance vector.

[0068] After obtaining the density vector, the relative distance of each event-driven fragment unit's feature vector is calculated. The relative distance measures the distance of a feature point relative to other high-density feature points. For each event-driven fragment unit's feature vector, all other fragment units with higher densities are first identified by comparing the values ​​of each element in the density vector. The fragment units corresponding to elements with densities greater than the current fragment unit's density are then determined. Next, the distances between the current fragment unit and these high-density fragment units are calculated. These distances can be obtained from the distance matrix. The minimum value among these distances is determined and used as the relative distance of the current fragment unit. If the current fragment unit has the highest density among all fragment units, then it has no fragment units with higher densities. In this case, the maximum distance value is used as its relative distance. The maximum distance value can be the largest element value in the distance matrix.

[0069] The relative distances of each event-driven fragment unit are combined to generate a relative distance vector. The number of elements in the relative distance vector is the same as the number of event-driven fragment units, and each element corresponds to the relative distance of one event-driven fragment unit.

[0070] Step S245: Perform joint filtering on the density vector and the relative distance vector, and select event-driven fragment units whose density vector values ​​exceed a preset density threshold and whose relative distance vector values ​​exceed a preset distance threshold as cluster centers. Each cluster center represents an initial cluster.

[0071] For each event-driven fragment unit, its value in the density vector and its value in the relative distance vector are checked to see if they exceed a preset density threshold and a preset distance threshold, respectively. Only when both conditions are met is the event-driven fragment unit selected as a cluster center. Each selected cluster center represents an initial cluster. These cluster centers have a high local feature point distribution density in the feature space and a large relative distance to other high-density feature points, indicating that they have unique characteristics and can represent different clusters.

[0072] Step S246: Perform cluster assignment processing on event-driven fragment units that are not cluster centers. Assign each non-center fragment unit to the cluster of the nearest cluster center with a higher density than itself, resulting in multiple clusters containing fragment units with similar features, so that each event-driven fragment unit belongs to only one cluster.

[0073] After determining the cluster centers, the event-driven fragment units that are not cluster centers are subjected to cluster assignment. For each non-center fragment unit, the nearest cluster center with a higher density is determined.

[0074] Specifically, the process iterates through all cluster centers, calculating the distances between non-central fragment units and each cluster center; these distances can be obtained from the distance matrix. Simultaneously, the density of each cluster center is compared with the density of non-central fragment units, identifying cluster centers with higher densities. Among these high-density cluster centers, the one closest to a non-central fragment unit is selected. The non-central fragment unit is then assigned to the cluster containing this nearest cluster center with higher density. In this way, all event-driven fragment units without central cluster centers are assigned to different clusters. Ultimately, multiple clusters containing fragment units with similar features are obtained, and each event-driven fragment unit belongs to only one cluster.

[0075] Step S250: Construct a feature prototype for each cluster by calculating the feature median vector of all event-driven fragment units within the cluster as the feature prototype of the cluster. Each element of the feature prototype corresponds to the median value of a feature type.

[0076] After clustering, a feature prototype needs to be constructed for each cluster. The feature prototype is a vector representing the typical features of a cluster. For each cluster, the median vector of all event-driven fragment units within the cluster is calculated. Specifically, for each feature type, the values ​​of all event-driven fragment units within the cluster for that feature type are sorted. Then, the median value after sorting is determined; this median value is the median value for that feature type. Combining the median values ​​of all feature types forms the feature prototype for that cluster. Each element of the feature prototype corresponds to the median value of a feature type. Through the feature prototype, a concise vector can represent the features of the entire cluster.

[0077] Step S260: Perform similarity matching between the feature prototype and the pre-stored computing task feature library. By calculating the similarity between the feature prototype and the task feature template, determine the computing task type corresponding to each cluster and mark the feature prototype as the power consumption signature of the computing task type. All marked power consumption signatures together form the power consumption signature set associated with the load task, and each power consumption signature establishes a correspondence with the corresponding computing task through a unique task type identifier.

[0078] After obtaining the feature prototypes of each cluster, a similarity match is performed with a pre-stored feature library for computational tasks. The feature library stores feature templates for various known computational tasks, which represent typical features of different computational tasks.

[0079] By calculating the similarity between the feature prototype and the task feature template, the computational task type corresponding to each cluster can be determined. Similarity can be calculated using various methods, such as calculating the distance between the feature prototype and the task feature template (the smaller the distance, the higher the similarity), or using correlation analysis to calculate the correlation coefficient between them (the higher the correlation coefficient, the higher the similarity).

[0080] When a task feature template with the highest similarity to the feature prototype is found, the computational task type corresponding to that cluster is determined as the computational task type represented by that template. Then, the feature prototype is labeled with a power consumption signature for that computational task type. All labeled power consumption signatures together form a power consumption signature set associated with the workload task. In this set, each power consumption signature is associated with a corresponding computational task through a unique task type identifier.

[0081] Step S300: Input the power consumption signature set associated with the load task into the pre-trained load prediction model, and obtain the load intensity change profile in the future power supply cycle through time series analysis. The load intensity change profile contains the time series pattern information of load change.

[0082] In some embodiments, step S300 may specifically include the following steps S310-S360:

[0083] Step S310: Arrange the power signature set associated with the load task in chronological order to generate a power signature sequence with a time sequence relationship. Each element of the power signature sequence corresponds to a power signature vector at a time point, and the order of the elements is consistent with the chronological order.

[0084] To enable better time-series analysis of the load forecasting model, the set of power consumption signatures associated with the load tasks needs to be arranged chronologically. Each power consumption signature corresponds to a specific point in time, representing the power consumption characteristics of the computational task at that time. Arranging these power consumption signatures in chronological order yields a power consumption signature sequence, where each element is a power consumption signature vector containing various power consumption characteristics of the computational task at that time. The order of the elements maintains the chronological order, clearly demonstrating how the power consumption signatures change over time.

[0085] Step S320: Extend the power signature sequence with temporal features. For the power signature vector at each moment in the sequence, calculate the difference between the vector and the power signature vectors at the previous preset number of historical moments to obtain multiple historical difference vectors. Concatenate the original power signature vector at the current moment with the multiple historical difference vectors in chronological order to generate the extended feature vector at that moment. Arrange the extended feature vectors at all moments in chronological order to form an extended power signature sequence. The dimension of each vector in the extended power signature sequence is equal to the dimension of the original power signature vector multiplied by a preset value, where the preset value is the sum of a preset number of historical vectors and 1.

[0086] The purpose of temporal feature extension is to incorporate historical power signature change information to more comprehensively describe the current load state. For each power signature vector in the power signature sequence, the difference between it and the power signature vectors of the previous preset number of historical moments is calculated. The preset number is a predetermined value representing the number of historical moments to consider. By calculating the differences, multiple historical difference vectors are obtained to determine the changes in the power signature across different moments. The original power signature vector at the current moment is concatenated with the multiple historical difference vectors in chronological order, that is, the power signature vector at the current moment is placed first, and then the difference vectors of each previous historical moment are concatenated in sequence. This generates the extended feature vector for that moment. The extended feature vectors of all moments are arranged in chronological order to form the extended power signature sequence. The dimension of each vector in the extended power signature sequence is equal to the dimension of the original power signature vector multiplied by a preset value, which is the preset number of historical moments plus 1. This is because the extended feature vector not only contains the original power signature vector at the current moment but also contains the preset number of historical difference vectors. In this way, the extended power signature sequence contains more historical information, which helps to improve the accuracy of prediction.

[0087] Step S330: Input the extended power signature sequence into the feature encoding layer of the load prediction model, and map the input features to a high-dimensional feature space through multi-layer nonlinear transformation to generate an encoded feature sequence containing deep feature representation. The time length of the encoded feature sequence is the same as that of the extended power signature sequence, and the feature dimension is the encoding dimension preset by the model.

[0088] After the extended power signature sequence is input into the feature encoding layer, the feature encoding layer performs multiple nonlinear transformations on the input features. These nonlinear transformations increase the model's expressive power, enabling it to capture more complex patterns and relationships within the input features. Through these multiple nonlinear transformations, the input features are mapped from the original feature space to a high-dimensional feature space.

[0089] In a high-dimensional feature space, features are more discriminative, and the model can more easily learn deep representations of the features. This ultimately generates an encoded feature sequence containing deep feature representations. The time length of the encoded feature sequence is the same as the extended power signature sequence because each extended feature vector at each time point corresponds to an encoded feature vector.

[0090] Step S340: Perform temporal dependency modeling on the encoded feature sequence. By capturing the long-term and short-term dependencies between encoded features at different time points, a dependent feature sequence with temporal correlation is generated. Each feature vector of the dependent feature sequence contains feature information of the corresponding time point and its associated time points.

[0091] After obtaining the encoded feature sequence, it is necessary to perform temporal dependency modeling. The purpose of temporal dependency modeling is to capture the long-term and short-term dependencies between encoded features at different time points. In actual load changes, the current load state is not only related to the current computational task, but may also be affected by computational tasks over a past period. This influence may be short-term or long-term.

[0092] In some embodiments, step S340 may specifically include the following steps S341-S346:

[0093] Step S341: Divide the encoded feature sequence into time windows. Divide the continuous encoded feature sequence into multiple overlapping time window feature matrices according to a preset window size and sliding step size. Each time window feature matrix contains a preset number of continuous encoded feature vectors.

[0094] Time window partitioning divides a continuous sequence of encoded features into multiple smaller time windows. The preset window size and sliding step are predetermined parameters. The window size represents the number of encoded feature vectors contained in each time window, determining its length. The sliding step represents the distance the time window is moved in each iteration, determining the degree of overlap between time windows. By partitioning the continuous sequence of encoded features according to the preset window size and sliding step, multiple overlapping time window feature matrices are obtained. Each time window feature matrix contains a preset number of continuous encoded feature vectors. Through time window partitioning, the encoded feature sequence can be transformed into multiple relatively independent time windows.

[0095] Step S342: Perform self-attention calculation on the feature matrix of each time window. Generate an attention weight matrix by calculating the attention weights between the encoded feature vectors at different positions within the window. The row and column dimensions of the attention weight matrix are the same as the window size, and the matrix element values ​​represent the correlation strength between the corresponding two position features.

[0096] After obtaining the feature matrices for each time window, self-attention is calculated for each feature matrix. For each feature matrix, attention weights are calculated between encoded feature vectors at different positions within the window. These attention weights represent the correlation strength between feature vectors at different positions. Specifically, each encoded feature vector within the window is compared and calculated with all other encoded feature vectors. Through a series of calculation steps, the correlation degree between each feature vector and other feature vectors is obtained, and these correlation degrees are converted into attention weights. All attention weights are combined to generate an attention weight matrix. The row and column dimensions of the attention weight matrix are the same as the window size, and each element value in the matrix represents the correlation strength between two corresponding feature vectors.

[0097] Step S343: The attention weight matrix and the time window feature matrix are weighted and summed to generate a window-dependent feature vector containing feature association information within the window. The dimension of the window-dependent feature vector is the same as the dimension of the encoded feature vector.

[0098] After obtaining the attention weight matrix, it is weighted and summed with the time window feature matrix. The purpose is to combine the encoded feature vectors at different positions within the window according to their correlation strength, generating a window-dependent feature vector that contains information about feature correlation within the window. Specifically, each element in the attention weight matrix is ​​multiplied by the element of the corresponding encoded feature vector in the time window feature matrix, and then all the multiplication results are summed. This weighted summation operation is performed for each element in each dimension. The final window-dependent feature vector has the same dimension as the encoded feature vector.

[0099] Window-dependent feature vectors contain the correlation information between encoded feature vectors at different locations within a time window, and can more comprehensively represent the feature state within the time window.

[0100] Step S344: Arrange the window dependency feature vectors of all time windows in order to generate an intermediate dependency feature sequence. Each element of the intermediate dependency feature sequence corresponds to the window dependency feature vector of a time window, and its sequence length is equal to the number of time windows.

[0101] Arranging the window-dependent feature vectors of all time windows in chronological order generates an intermediate dependency feature sequence, where each element corresponds to a window-dependent feature vector of a time window. Since the number of time windows is determined by the length of the encoded feature sequence, the window size, and the sliding step, the length of the intermediate dependency feature sequence equals the number of time windows. By generating the intermediate dependency feature sequence, the feature association information of each time window can be integrated in chronological order, providing ordered input for bidirectional cyclic processing.

[0102] Step S345: Perform bidirectional loop processing on the intermediate dependency feature sequence. The forward loop and backward loop capture the forward and backward time dependencies of the sequence respectively to generate a bidirectional dependency feature sequence. The feature dimension of the bidirectional dependency feature sequence is twice that of the intermediate dependency feature sequence.

[0103] The forward loop starts from the beginning of the intermediate dependency feature sequence and processes each element sequentially, capturing the positive temporal dependencies of the sequence. That is, it analyzes the dependency relationship between the current element and the previous elements. The backward loop, on the other hand, starts from the end of the intermediate dependency feature sequence and processes each element sequentially, capturing the negative temporal dependencies of the sequence. That is, it analyzes the dependency relationship between the current element and the subsequent elements.

[0104] By using forward and backward loops, positive and negative dependency features are obtained, respectively. These two dependency features are combined to generate a bidirectional dependency feature sequence. The feature dimension of the bidirectional dependency feature sequence is twice that of the intermediate dependency feature sequence because it simultaneously includes both positive and negative dependency features. Through bidirectional loop processing, long-term and short-term dependencies in the intermediate dependency feature sequence can be captured more comprehensively.

[0105] Step S346: Perform feature fusion between the bidirectional dependency feature sequence and the feature representation of the corresponding time window extracted from the encoded feature sequence to generate the final dependency feature sequence, which contains both the original encoded features and the temporal dependency features.

[0106] The purpose of feature fusion is to combine the original encoded features and temporal dependency features to generate a more comprehensive and accurate dependency feature sequence. Feature representations corresponding to the corresponding time windows are extracted from the encoded feature sequence. These feature representations retain information from the original encoded features. The bidirectional dependency feature sequence and these feature representations are then fused using various methods, such as concatenation or weighted summation. The final dependency feature sequence contains both the original encoded features and the temporal dependency features. The original encoded features provide basic feature information for each time point, while the temporal dependency features reflect the correlations between different time points.

[0107] Step S350: Input the dependent feature sequence into the prediction output layer of the load prediction model, generate a load intensity prediction sequence by predicting the features at future time points, the time length of the load intensity prediction sequence corresponds to the time length of the future power supply cycle, and each element represents the load intensity prediction value at the corresponding time point.

[0108] In some embodiments, step S350 may specifically include the following steps S351-S355:

[0109] Step S351: Extract the dependent feature sequence from the dependent feature sequence and use a preset number of dependent feature vectors at the end of the sequence as the prediction input features. The time length of the prediction input features is related to the preset prediction step size of the model, and the feature dimension is the same as the feature dimension of the dependent feature sequence.

[0110] The purpose of feature sequence truncation is to select appropriate dependent feature vectors as prediction input features to improve prediction accuracy. The preset number is a pre-determined value representing the number of dependent feature vectors to be truncated; this number is related to the model's preset prediction step size. The prediction step size is the time span considered by the model during prediction, and the preset number is chosen based on the prediction step size to ensure that the truncated dependent feature vectors provide sufficient information for prediction. Starting from the end of the dependent feature sequence, the preset number of dependent feature vectors are extracted as prediction input features. The time length of the prediction input features is related to the model's preset prediction step size, and the feature dimension is the same as the feature dimension of the dependent feature sequence.

[0111] Step S352: Reorganize the predicted input features by dimensional substitution of the time dimension and feature dimension and adjusting the arrangement order of features to generate a reorganized feature tensor that meets the input requirements of the predicted output layer.

[0112] Feature reorganization transforms the predicted input features into a format that meets the requirements of the prediction output layer. The original dimensional order of the predicted input features may not conform to the requirements of the prediction output layer, necessitating a dimensionality swap between the time and feature dimensions. Simultaneously, the feature arrangement needs to be adjusted to align with the desired configuration of the prediction output layer. This results in a reorganized feature tensor, with the dimensional order being [feature dimension, time length]. This dimensionality adjustment allows the prediction output layer to more easily process and analyze the input features, improving prediction efficiency and accuracy.

[0113] Step S353: Input the recombined feature tensor into the first fully connected layer of the prediction output layer. Through linear transformation and nonlinear activation function processing, map the recombined feature tensor to the hidden feature space to generate a hidden feature vector. The dimension of the hidden feature vector is the preset number of hidden layer neurons.

[0114] The first fully connected layer of the prediction output layer performs preliminary processing and transformation on the input reconstructed feature tensor. After the reconstructed feature tensor is input into the first fully connected layer, the network performs a linear transformation on it. This linear transformation is achieved through a series of weight matrices and bias vectors, which linearly combine and transform the input reconstructed feature tensor. Then, it is processed by a non-linear activation function to introduce non-linearity and increase the model's expressive power. Through linear transformation and non-linear activation function processing, the reconstructed feature tensor is mapped to the hidden feature space. The hidden feature space is an abstract feature space where the features better reflect the essence and latent patterns of the input features, ultimately generating hidden feature vectors. The dimension of the hidden feature vectors is the preset number of hidden layer neurons. The preset number of hidden layer neurons is a fixed value determined during model training, which determines the dimensionality of the hidden feature vectors.

[0115] Step S354: Input the hidden feature vector into the second fully connected network of the prediction output layer, and map the hidden feature vector to the prediction output space through linear transformation to generate a preliminary prediction vector containing the predicted load intensity values ​​at multiple future time points. The length of the preliminary prediction vector is the same as the time length of the future power supply cycle.

[0116] After obtaining the hidden feature vectors, they are input into the second fully connected layer of the prediction output layer. The main task of the second fully connected layer is to map the hidden feature vectors to the prediction output space, generating preliminary load intensity predictions. The second fully connected layer performs a linear transformation on the hidden feature vectors, using a series of weight matrices and bias vectors to linearly combine and transform them, mapping them from the hidden feature space to the prediction output space. The vectors in the prediction output space represent the predicted load intensity values ​​at multiple future time points.

[0117] Finally, a preliminary prediction vector containing load intensity predictions for multiple future time points is generated. The length of the preliminary prediction vector is the same as the length of the future power supply cycle, which means that each element in the preliminary prediction vector corresponds to the load intensity prediction value for a time point within the future power supply cycle.

[0118] Step S355: Correct the preliminary prediction vector by weighting the preliminary prediction value with the historical average load intensity value to generate the final load intensity prediction sequence. Each element of the load intensity prediction sequence is the corrected load intensity prediction value at the corresponding time point, and the number of elements is consistent with the time length of the future power supply cycle.

[0119] The purpose of revising the initial forecast vector after obtaining it is to improve forecast accuracy and reduce forecast error. The historical average load intensity value is an average calculated based on load intensity data over a past period. It reflects the long-term average level of load intensity. Weighting the initial forecast value with the historical average load intensity value is to comprehensively consider historical data and current forecast results.

[0120] Specifically, different weights are assigned to the preliminary forecast values ​​and historical average load intensity values, and then they are summed with weights. The weight allocation can be adjusted according to actual conditions, for example, based on the reliability of historical data and the credibility of the current forecast. Finally, a final load intensity forecast sequence is generated, where each element is the corrected load intensity forecast value at the corresponding time point, and the number of elements is consistent with the length of the future power supply cycle. Through forecast correction, information from historical data can be fully utilized to optimize the preliminary forecast results, making the final load intensity forecast sequence more accurate and reliable, and better reflecting the actual changes in load intensity within the future power supply cycle.

[0121] Step S360: Perform trend smoothing on the load intensity prediction sequence. By calculating the rate of change of adjacent prediction values, perform moving average correction on prediction values ​​whose rate of change exceeds the preset range to generate a continuous and smooth load intensity change profile. The load intensity change profile includes the corrected load intensity prediction value and the corresponding time stamp information.

[0122] The purpose of trend smoothing is to eliminate fluctuations and noise in the predicted values, making the changes in load intensity more continuous and stable. First, the rate of change of adjacent predicted values ​​in the load intensity prediction sequence is calculated. The rate of change reflects the magnitude of load intensity changes between adjacent time points. By comparing the magnitude of each rate of change with a preset range, it can be determined whether the change is abnormal. The preset range is a reasonable interval pre-set based on practical experience and the characteristics of load changes. For predicted values ​​whose rate of change exceeds the preset range, a moving average correction method is used. The moving average replaces the original data by calculating the average value of data within a certain window. Specifically, for a predicted value that needs correction, a certain number of adjacent predicted values ​​are selected, and the average of these values ​​is calculated. This average is used as the corrected predicted value. In this way, abnormal changes can be smoothed to a certain extent, making the changes in load intensity more natural.

[0123] After correction using a moving average, a continuous and smooth load intensity variation profile is generated. This profile not only includes the corrected load intensity predictions but also adds corresponding time stamp information to each prediction. The time stamp information accurately indicates the time point corresponding to each prediction, thus clearly demonstrating the trend of load intensity changes over time.

[0124] Step S400: Perform feedback calibration between the load intensity change profile and the current dual-power supply status information of the PC power supply to generate a power allocation strategy that adapts to the timing pattern of the load intensity change profile.

[0125] In some embodiments, step S400 may specifically include the following steps S410-S460:

[0126] Step S410: Divide the load intensity change profile into time granularity, and discretize the continuous load intensity change profile into load intensity values ​​at multiple time nodes according to the preset control cycle. Each time node corresponds to a control moment, and the intervals between time nodes are equal.

[0127] The load intensity variation profile is a continuous curve describing how the load intensity changes over time. By dividing the curve into time granularities, this continuous curve is discretized into multiple discrete time nodes. A preset control period is a predetermined time interval that determines the size of the intervals between time nodes. According to this control period, the continuous load intensity variation profile is divided, with each division point corresponding to a time node, and each time node corresponding to a control moment when the power distribution needs to be adjusted. The intervals between time nodes are equal, ensuring regularity and consistency in subsequent analysis and calculations. Through time granularity division, the load intensity variation profile is transformed into a series of discrete load intensity values, each corresponding to a specific time node.

[0128] Step S420: Extract status parameters from the current dual-power supply status information of the PC power supply, obtain the current output power, voltage fluctuation range, current stability and efficiency parameters of the main power supply circuit and the auxiliary power supply circuit respectively, and generate a parameter set characterizing the dual-power supply status. The parameter set contains a number of parameter items, which is the product of the number of power supply circuits and the number of parameter types.

[0129] The dual-power supply status information includes various operating states and performance parameters of the main power supply circuit and the auxiliary power supply circuit. Parameters are extracted for both the main and auxiliary power supply circuits. For each circuit, its current output power, voltage fluctuation range, current stability, and efficiency parameters are obtained. The current output power reflects the actual power output of the circuit; the voltage fluctuation range indicates the magnitude of voltage change over a certain period, reflecting voltage stability; current stability reflects the stability of the current output; and the efficiency parameter indicates the energy conversion efficiency of the circuit during power supply. These parameters from the main and auxiliary power supply circuits are combined to generate a parameter set characterizing the dual-power supply status. Since there are two power supply circuits, each with four parameter types, the parameter set contains the product of the number of power supply circuits and the number of parameter types, i.e., eight parameter items. This parameter set comprehensively describes the current dual-power supply status of the PC power supply.

[0130] Step S430: Correlate the load intensity value of the time node with the dual power supply status parameter set and generate a load power supply ratio sequence by calculating the ratio of the load intensity value of each time node to the total power of the dual power supply. Each element of the load power supply ratio sequence represents the ratio of the load demand to the power supply capacity of the corresponding time node.

[0131] The purpose of correlation modeling is to determine the inherent relationship between load demand and power supply capacity in order to better allocate power. First, the total power of the dual power supply is calculated. The total power of the dual power supply is the sum of the current output power of the main power supply circuit and the auxiliary power supply circuit. Then, for the load intensity value at each time point, its ratio to the total power of the dual power supply is calculated. This ratio represents the proportional relationship between load demand and power supply capacity at that time point.

[0132] The load-to-supply ratios at each time point are combined to generate a load-to-supply ratio sequence. Each element in the sequence corresponds to a time point, and its value represents the ratio of load demand to power supply capacity at that time point. The load-to-supply ratio sequence provides a clear understanding of the matching between load demand and power supply capacity at different time points.

[0133] Step S440: Perform threshold comparison on the load power supply ratio sequence, mark the time nodes with ratios greater than the preset threshold as power adjustment nodes, and mark the remaining time nodes as power maintenance nodes, and generate a node type marking sequence. The length of the node type marking sequence is the same as the number of time nodes.

[0134] A preset threshold is a pre-defined standard used to determine whether the ratio of load demand to power supply capacity is reasonable. For each element in the load-to-power-capacity ratio sequence, it is compared to the preset threshold. If the ratio is greater than the preset threshold, it indicates that the load demand is too high relative to the power supply capacity at that time point, requiring adjustment of the power allocation; therefore, this time point is marked as a power adjustment node. If the ratio is less than or equal to the preset threshold, it indicates that the current power supply capacity can meet the load demand, and no power adjustment is needed; this time point is marked as a power maintenance node. The marking results of all time points are combined to generate a node type marking sequence. The length of the node type marking sequence is the same as the number of time points, and each element corresponds to a marking type for a time point. Through the node type marking sequence, it is clear which time points require power adjustment and which time points can maintain the current power allocation.

[0135] Step S450: Perform power allocation calculation for each power adjustment node. Based on the load intensity value and the set of dual power supply status parameters, determine the target power values ​​of the main power supply circuit and the auxiliary power supply circuit at that time node by solving the optimization problem that maximizes the power supply efficiency. The target power values ​​satisfy the constraint that the total power of the dual power supply is equal to the load intensity value.

[0136] The goal of power allocation calculation is to maximize the efficiency of dual power supply while meeting load demand. Based on the load intensity value and the set of dual power supply state parameters at this time point, an optimization problem is constructed. The objective of the optimization problem is to find the optimal power allocation scheme between the main power supply circuit and the auxiliary power supply circuit, maximizing the overall efficiency of the dual power supply. Simultaneously, the constraint that the total power of the dual power supply equals the load intensity value must be satisfied; that is, the sum of the output power of the main power supply circuit and the auxiliary power supply circuit must equal the load intensity value at this time point.

[0137] In some embodiments, step S450 may specifically include the following steps S451-S456:

[0138] Step S451: Perform efficiency curve fitting on the dual power supply state parameter set, extract the efficiency parameters of the main power supply circuit and the auxiliary power supply circuit respectively, and generate the main power supply efficiency curve and the auxiliary power supply efficiency curve by fitting the efficiency values ​​at different power points into a continuous function. The independent variable of the efficiency curve is the power value and the dependent variable is the efficiency value.

[0139] Efficiency parameters for the main power supply circuit and the auxiliary power supply circuit are extracted from the dual-power supply state parameter set. These efficiency parameters record the power supply efficiency values ​​at different power points. These discrete efficiency values ​​are fitted to form a continuous function. Various fitting methods can be used, such as polynomial fitting and linear fitting, and the appropriate fitting method is selected based on the actual data characteristics. After fitting, the main power supply efficiency curve and the auxiliary power supply efficiency curve are generated. The independent variable of the efficiency curve is the power value, and the dependent variable is the efficiency value, meaning that for each given power value, the corresponding power supply efficiency can be directly obtained from the efficiency curve.

[0140] Step S452: Obtain the preset power allocation optimization objective function, which uses the power value of the main power supply circuit and the power value of the auxiliary power supply circuit as optimization variables, and takes the highest total efficiency of dual power supply as the optimization objective. The total efficiency is calculated by weighted average of the efficiency curves of the main power supply circuit and the auxiliary power supply circuit, with the weight being the power proportion of the corresponding circuit.

[0141] The objective function is the core function guiding power allocation calculations, aiming to find the power allocation scheme that maximizes the overall efficiency of the dual-power supply. The objective function uses the power values ​​of the main power supply circuit and the auxiliary power supply circuit as optimization variables. That is, the optimal power allocation scheme is found by adjusting these two power values. The overall efficiency of the dual-power supply is calculated as a weighted average of the efficiency curves of the main power supply and the auxiliary power supply. Specifically, for each power value of the main power supply circuit and the auxiliary power supply circuit, the corresponding power supply efficiency is obtained from the main power supply efficiency curve and the auxiliary power supply efficiency curve, respectively. Then, a weighted average is performed based on the power proportion of the corresponding circuit to obtain the overall efficiency of the dual-power supply.

[0142] By optimizing the objective function, under the constraint that the total power of the dual power supply is equal to the load strength, we can find the power values ​​of the main power supply circuit and the auxiliary power supply circuit that maximize the overall efficiency of the dual power supply, thereby achieving optimal power allocation.

[0143] Step S453: Obtain preset power allocation constraints, including the equality constraint that the total power of the dual power supply is equal to the load strength value, and the inequality constraints that the power values ​​of the main power supply circuit and the auxiliary power supply circuit are respectively within their minimum and maximum output power ranges. The minimum and maximum output power ranges are extracted from the dual power supply state parameter set.

[0144] The equality constraint states that the total power of the dual power supply equals the load intensity value. This means that the sum of the output power of the main power supply circuit and the auxiliary power supply circuit must equal the load intensity value at that time point to ensure that the load demand is met. The inequality constraint states that the power values ​​of the main power supply circuit and the auxiliary power supply circuit are respectively within their minimum and maximum output power ranges. The minimum and maximum output power ranges can be extracted from the set of dual power supply status parameters. Each power supply circuit has its own allowable minimum and maximum output power, and power allocation must be carried out within these ranges to ensure the normal operation and safety of the power supply circuit. By considering these constraints, it is possible to ensure that power allocation is within a reasonable range while meeting load demand, avoiding over-power or under-power situations, and ensuring the stable operation of the power supply.

[0145] Step S454: Solve the objective function under constraints. By finding the optimal variable values ​​that maximize the objective function, obtain the optimal power values ​​of the main power supply circuit and the auxiliary power supply circuit. The optimal power values ​​satisfy all constraints and maximize the overall efficiency.

[0146] After obtaining the objective function and constraints, the goal of solving the objective function under the constraints is to find the power values ​​of the main power supply circuit and the auxiliary power supply circuit that maximize the objective function, i.e., the optimal power values.

[0147] Various optimization algorithms can be used to solve this optimization problem, such as gradient descent and genetic algorithms. These algorithms gradually approach the optimal solution by continuously adjusting the values ​​of the optimization variables. During the solution process, it is necessary to always satisfy the equality constraint that the total power of the dual power supply equals the load intensity value, and the inequality constraint that the power values ​​of the main power supply circuit and the auxiliary power supply circuit are within their minimum and maximum output power ranges. Through continuous iteration and calculation, the optimal variable values ​​that maximize the objective function are finally found, namely the optimal power values ​​of the main power supply circuit and the auxiliary power supply circuit. These optimal power values ​​satisfy all constraints and maximize the overall efficiency of the dual power supply, achieving optimal power allocation.

[0148] Step S455: Verify the feasibility of the optimal power value by checking whether the optimal power value falls within the corresponding power adjustment range. If it exceeds the range, adjust it to the nearest boundary value to generate a feasible target power value.

[0149] The purpose of feasibility verification is to ensure that the optimal power value is within the actual power regulation range and can be implemented by the power supply. It checks whether the optimal power value falls within the corresponding power regulation range, i.e., the minimum and maximum output power range of the main power supply circuit and the auxiliary power supply circuit. If the optimal power value exceeds this range, it indicates that the value cannot be achieved in practice and needs adjustment. For optimal power values ​​that exceed the range, they are adjusted to the nearest boundary value. For example, if the optimal power value of the main power supply circuit is greater than its maximum output power, it is adjusted to the maximum output power; if it is less than its minimum output power, it is adjusted to the minimum output power.

[0150] Step S456: Compare the feasible target power value with the current power value in the dual power supply status parameter set, calculate the power adjustment amount, and if the adjustment amount is greater than the preset maximum allowable adjustment amount, then correct the target power value according to the maximum allowable adjustment amount to improve the stability of the power adjustment process and finally determine the target power value at this time node.

[0151] After obtaining the feasible target power value, it is compared with the current power value in the dual-power supply status parameter set to calculate the power adjustment amount. The power adjustment amount is the difference between the feasible target power value and the current power value, which reflects the magnitude of the power adjustment required.

[0152] The calculated power adjustment is compared with the preset maximum allowable adjustment, which is a maximum adjustment range pre-set based on the power supply's performance and stability requirements. If the power adjustment exceeds the preset maximum allowable adjustment, it indicates that the adjustment range is too large and may affect the power supply's stability. In this case, the target power value is corrected according to the maximum allowable adjustment. Specifically, the feasible target power value is adjusted towards the current power value by the maximum allowable adjustment. This method ensures the stability of the power adjustment process and avoids power supply failures or instability caused by excessive adjustment. Finally, the target power value at that time point is determined, which satisfies both load requirements and ensures stable power supply operation.

[0153] Step S460: Arrange the target power values ​​or maintenance power values ​​of all time nodes in chronological order to generate a power allocation strategy that includes a dual-power supply power allocation scheme for each control moment. The power value change pattern in the power allocation strategy is consistent with the temporal pattern of the load intensity change profile.

[0154] In some embodiments, step S460 may specifically include the following steps S461-S466:

[0155] Step S461: Determine the maintenance power value for each power maintenance node, and use the target power value of the previous power adjustment node as the maintenance power value of the current node to ensure the continuity of power adjustment.

[0156] To ensure continuous power regulation, a maintenance power value needs to be determined for each power maintenance node. Since power maintenance nodes do not require power regulation, the target power value of the preceding power regulation node is used as the maintenance power value for the current node. This ensures that power distribution remains continuous over time, avoiding sudden power jumps or interruptions. By using the target power value of the preceding power regulation node, it is ensured that the power supply's output power remains relatively stable during power maintenance, providing a continuous and stable power supply to the load.

[0157] Step S462: Smoothly transition the power values ​​of adjacent time nodes by performing linear interpolation between the maintenance power values ​​before and after the power adjustment node to generate a power value sequence of the transition interval. The length of the transition interval corresponds to the response time of the power adjustment.

[0158] Before and after the power adjustment node, in order to make the power change smoother, it is necessary to smoothly transition the power values ​​of adjacent time nodes. The purpose is to avoid sudden power changes from impacting the load and power supply, and to make the power adjustment process more natural and stable.

[0159] A power value sequence for the transition interval is generated by linear interpolation between the sustained power values ​​before and after the power adjustment node. Linear interpolation calculates the value at the midpoint based on a linear relationship between the two endpoints. The length of the transition interval corresponds to the response time of the power adjustment, which is the time required for the power supply to adjust from one power state to another. The power value sequence for the transition interval generated by linear interpolation allows the power to change gradually during the adjustment process, avoiding abrupt power changes and improving the stability of the power supply and load compatibility.

[0160] Step S463: Integrate the target power value, maintenance power value and transition interval power value sequence of all time nodes in chronological order to generate a continuous dual-path power allocation curve, which includes the main power supply circuit power allocation curve and the auxiliary power supply circuit power allocation curve.

[0161] The target power value, maintenance power value, and transition range power value sequences for all time points are integrated and processed in chronological order. First, the power values ​​of the main power supply circuit and the auxiliary power supply circuit at each time point are extracted separately and then arranged sequentially. The arranged power values ​​are connected to generate a continuous dual-path power distribution curve, which includes the power distribution curve of the main power supply circuit and the power distribution curve of the auxiliary power supply circuit. These two curves respectively show the power distribution of the main power supply circuit and the auxiliary power supply circuit at different time points. Through the dual-path power distribution curve, the trend of power change over time can be visually observed.

[0162] Step S464: Perform power balance verification on the dual-path power distribution curve, calculate the sum of the power value of the main power supply circuit and the power value of the auxiliary power supply circuit at each time point, and compare it with the load intensity value at the corresponding time point. If the deviation is greater than the preset allowable deviation, use the current dual-path power value as the initial value, and re-perform the power distribution calculation based on the load intensity value and power supply circuit parameters at that time point to update and correct the dual-path power distribution curve until the deviation between the total power and the load intensity value is within the allowable range.

[0163] The purpose of power balance verification is to check whether the sum of the power values ​​of the main power supply circuit and the auxiliary power supply circuit at each time point matches the load intensity value at the corresponding time point.

[0164] Calculate the sum of the power values ​​of the main power supply circuit and the auxiliary power supply circuit at each time point, and then compare this sum with the load intensity value at the corresponding time point. The preset allowable deviation is a reasonable deviation range pre-set based on the actual situation. If the deviation is greater than the preset allowable deviation, it indicates that there is a problem with the power distribution and it needs to be corrected.

[0165] Using the current dual-circuit power value as the initial value, the power allocation calculation is re-performed based on the load intensity value and power supply circuit parameters at that time point. Through recalculation, a new power allocation scheme for the main and auxiliary power supply circuits is obtained, updating and correcting the dual-circuit power allocation curve. This process is repeated until the deviation between the total power and the load intensity value is within the allowable range. Through power balance verification and correction, the accuracy of the dual-circuit power allocation can be guaranteed, ensuring that the power supply output power accurately meets the load demand.

[0166] Step S465: Perform time stamping processing on the corrected dual-path power distribution curve, add a corresponding control time timestamp to each power value on the curve, and generate a power distribution timetable containing the time-power correspondence.

[0167] The timestamp accurately records the time point corresponding to each power value, and the correspondence between time and power can be clearly established through the timestamp. All time-power correspondences are compiled together to generate a power allocation schedule containing the time-power correspondences. The power allocation schedule records in detail the power allocation of the main power supply circuit and the auxiliary power supply circuit at each control moment.

[0168] Step S466: Format the power allocation schedule according to the control cycle to generate a data format that meets the requirements of the PC power control interface, and obtain the final power allocation strategy. The power value in the power allocation strategy changes over time in line with the trend of load intensity change profile.

[0169] Formatting converts the power allocation schedule into a data format that conforms to the requirements of the PC power control interface. The control interface has its specific data format and communication protocol; only by converting the power allocation strategy to this format can the power supply correctly receive and execute power allocation commands. After formatting, the final power allocation strategy is obtained. The power value in the power allocation strategy changes over time in line with the trend of load intensity changes. This means that when the load intensity increases, the power supply's output power will increase accordingly; when the load intensity decreases, the power supply's output power will decrease accordingly. In this way, dynamic matching of power supply output power and load changes is achieved, improving the power supply's efficiency and stability.

[0170] Step S500: Adjust the operating points of the primary and secondary conversion paths inside the PC power supply according to the power distribution strategy so that the power output power keeps dynamically matched with load changes.

[0171] In some embodiments, step S500 may specifically include the following steps S510-S560:

[0172] Step S510: Synchronize the power distribution strategy by calibrating the timestamp of the adjustment time in the strategy with the internal clock of the PC power supply, so that the execution time of the power adjustment command is consistent with the strategy setting time.

[0173] The purpose of time synchronization is to ensure that the timestamps of the control moments in the strategy are consistent with the internal clock of the PC power supply. The internal clock of the PC power supply is the clock system used for timing and control within the power supply, determining the power supply's operating state at different points in time. Calibrating the timestamps of the control moments in the strategy with the internal clock means matching each control moment specified in the strategy with the time of the internal clock. Through time synchronization, the execution time of the power adjustment command is ensured to be consistent with the strategy's set time. This ensures that at each control moment, the power supply can accurately adjust the operating points of the primary and secondary conversion paths according to the strategy's requirements, achieving precise control of the power supply's output power.

[0174] Step S520: Perform path allocation for the dual-path power allocation value at each control moment, allocate the power value of the main power supply circuit to the primary conversion path, allocate the power value of the auxiliary power supply circuit to the secondary conversion path, and generate the target power sequence for the primary conversion path and the secondary conversion path. Each element of the target power sequence corresponds to the target power value at a control moment.

[0175] After time synchronization is completed, path allocation is performed on the dual-path power distribution values ​​at each control moment. Based on the power supply design and operating principle, the power of the main power supply circuit is output through the primary conversion path, and the power of the auxiliary power supply circuit is output through the secondary conversion path. The power value of the main power supply circuit at each control moment is allocated to the primary conversion path, and the power value of the auxiliary power supply circuit is allocated to the secondary conversion path. These allocated power values ​​are arranged in the order of the control moments to generate target power sequences for the primary and secondary conversion paths. Each element of the target power sequence corresponds to a target power value at a control moment, representing the power output required by the primary and secondary conversion paths at that moment.

[0176] Step S530: Perform working parameter mapping processing on the target power sequence of the primary conversion path. Based on the preset power-parameter mapping relationship, convert each target power value into the corresponding PWM duty cycle, switching frequency and inductor current threshold to generate the target working parameter sequence of the primary conversion path. The length of the target working parameter sequence is the same as the length of the target power sequence, and each element contains three parameter values: PWM duty cycle, switching frequency and inductor current threshold.

[0177] The purpose of the operating parameter mapping process is to convert the target power value into specific operating parameters in order to precisely control the primary conversion path. The preset power-parameter mapping relationship is a correspondence established in advance based on the circuit characteristics and performance requirements of the primary conversion path, describing the operating parameters such as PWM duty cycle, switching frequency, and inductor current threshold corresponding to different power values.

[0178] In some embodiments, step S530 may specifically include the following steps S531-S536:

[0179] Step S531: Perform power range division processing on the target power sequence of the primary conversion path, divide the target power value into multiple power ranges according to the preset power level, each power range corresponds to a set of standard operating parameters, and the boundary values ​​between the power ranges correspond to the efficiency inflection point of the PC power supply.

[0180] The purpose of power range segmentation is to divide different target power values ​​into different ranges, each corresponding to a set of standard operating parameters. Preset power levels are power values ​​pre-defined based on the performance and efficiency characteristics of the primary conversion path. Each target power value in the target power sequence is compared with these power levels, and the target power value is divided into different power ranges based on the comparison results. Each power range corresponds to a set of standard operating parameters, which are typical parameter values ​​obtained through optimization and testing within that power range. The boundary values ​​between power ranges correspond to the efficiency inflection point of the PC power supply. The efficiency inflection point is the power point where the power supply efficiency changes significantly. Near these points, the power supply efficiency fluctuates considerably. By setting the boundary values ​​of the power ranges as efficiency inflection points, the mapping of operating parameters can be made more reasonable, improving the power supply efficiency.

[0181] Step S532: Search for the power range where each target power value is located and determine the corresponding standard operating parameter group. The standard operating parameter group includes the typical PWM duty cycle reference value, switching frequency reference value and inductor current threshold reference value for that power range.

[0182] After dividing the power range, for each target power value in the target power sequence of the primary conversion path, it is necessary to find the power range it belongs to and determine the corresponding standard operating parameter set. The power range to which the target power value belongs is determined by comparing it with the boundary values ​​of each power range. For the determined power range, the pre-set standard operating parameter set is then searched. The standard operating parameter set includes typical PWM duty cycle reference values, switching frequency reference values, and inductor current threshold reference values ​​for that power range.

[0183] Step S533: Calculate the fine-tuning coefficients for the standard operating parameter group. Based on the deviation between the target power value and the center value of the power range, calculate the PWM duty cycle fine-tuning coefficient, switching frequency fine-tuning coefficient, and inductor current threshold fine-tuning coefficient using a preset linear adjustment formula. The range of the fine-tuning coefficients is the preset coefficient adjustment range.

[0184] The deviation between the target power value and the center value of the power range is calculated. The center value of the power range is the middle power value within that range. The deviation is obtained by comparing the target power value with the center value. Based on this deviation, the PWM duty cycle adjustment coefficient, switching frequency adjustment coefficient, and inductor current threshold adjustment coefficient are calculated using preset linear adjustment formulas. For example, each adjustment coefficient is obtained by multiplying the pre-set (e.g., determined through extensive experiments and testing) PWM duty cycle adjustment coefficient, switching frequency adjustment coefficient, and inductor current threshold adjustment coefficient by the deviation. The range of the adjustment coefficients is a preset adjustment range. This range is designed to ensure that the adjustment amplitude is within a reasonable range, avoiding excessive adjustment that could affect the stability of the power supply.

[0185] Step S534: Based on the deviation between the target power value and the center value of the power range, as well as the adjustment characteristics of each parameter, calculate the PWM duty cycle adjustment amount, the switching frequency adjustment coefficient, and the inductor current threshold adjustment coefficient respectively. Based on the reference value in the standard working parameter group, apply the corresponding adjustment amount or adjustment coefficient to generate the adjusted PWM duty cycle, switching frequency, and inductor current threshold.

[0186] After obtaining the fine-tuning coefficients, the PWM duty cycle adjustment, switching frequency adjustment coefficient, and inductor current threshold adjustment coefficient are calculated based on the deviation between the target power value and the center value of the power range, as well as the adjustment characteristics of each parameter.

[0187] Different parameters have different adjustment characteristics. For example, the adjustment of the PWM duty cycle may be linearly related to the deviation, while the adjustment of the switching frequency may require consideration of more factors. Based on these characteristics, and combining the deviation and the fine-tuning coefficient, the specific adjustment amount or adjustment coefficient for each parameter can be calculated.

[0188] Based on the reference values ​​in the standard operating parameter set, the corresponding adjustment amount or adjustment coefficient is applied to generate the adjusted PWM duty cycle, switching frequency, and inductor current threshold. For example, for the PWM duty cycle, the reference value is added to the adjustment amount to obtain the adjusted PWM duty cycle; for the switching frequency, the reference value is multiplied by the adjustment coefficient to obtain the adjusted switching frequency. In this way, the operating parameters are more accurately matched to the target power value.

[0189] Step S535: Perform boundary checks on the adjusted parameter values. If the parameter values ​​exceed the preset safe working range, restrict them to the corresponding boundary values ​​and generate safe working parameter values.

[0190] The preset safe operating range is a parameter value range pre-set based on the circuit design and performance requirements of the primary conversion path. Check whether the adjusted parameter value exceeds this safe operating range. If the parameter value exceeds the safe operating range, it indicates that the parameter value may damage the primary conversion path or affect the stability of the power supply, and needs to be limited.

[0191] Out-of-range parameter values ​​are limited to their corresponding boundary values. For example, if the PWM duty cycle exceeds the maximum allowable value, it is adjusted to the maximum allowable value; if the switching frequency is lower than the minimum allowable value, it is adjusted to the minimum allowable value. Through boundary check processing, safe operating parameter values ​​are generated to ensure the safe operation of the primary conversion path.

[0192] Step S536: Arrange the safe operating parameter values ​​corresponding to each target power value in chronological order to generate the target operating parameter sequence of the primary conversion path. The length of the target operating parameter sequence is the same as the length of the target power sequence, and each element contains three parameter values: PWM duty cycle, switching frequency, and inductor current threshold.

[0193] The PWM duty cycle, switching frequency, and inductor current threshold corresponding to each target power value are combined to form an element containing three parameter values. All elements corresponding to the target power values ​​are arranged sequentially in chronological order to generate the target operating parameter sequence for the primary conversion path. The length of the target operating parameter sequence is the same as the length of the target power sequence. Each element contains three parameter values: PWM duty cycle, switching frequency, and inductor current threshold. This target operating parameter sequence clearly defines the specific operating parameters of the primary conversion path at different control moments, providing an accurate basis for subsequent instruction generation and adjustment.

[0194] Step S540: Perform operating parameter mapping processing on the target power sequence of the secondary conversion path. Based on the same mapping method as the primary conversion path, generate the target operating parameter sequence of the secondary conversion path. The target operating parameter sequence contains the operating parameter types corresponding to those of the primary conversion path.

[0195] The method for mapping the target power sequence of the secondary conversion path to working parameters is similar to that of the primary conversion path. Based on the same preset power-parameter mapping relationship as the primary conversion path, each target power value of the secondary conversion path is processed.

[0196] The process involves dividing the power range, finding the standard operating parameter group, calculating the fine-tuning coefficient, adjusting the parameters, and performing boundary checks. Through these steps, the target power value of the secondary conversion path is converted into the corresponding PWM duty cycle, switching frequency, and inductor current threshold.

[0197] The PWM duty cycle, switching frequency, and inductor current threshold corresponding to each target power value are arranged in chronological order according to the target power sequence to generate a target operating parameter sequence for the secondary conversion path. This target operating parameter sequence includes the operating parameter types corresponding to the primary conversion path, namely, PWM duty cycle, switching frequency, and inductor current threshold. In this way, specific operating parameters are provided for the precise control of the secondary conversion path.

[0198] Step S550: Package the target working parameter sequences of the primary conversion path and the secondary conversion path into instructions in chronological order to generate a working parameter adjustment instruction stream containing the adjustment time, target parameter value, and parameter adjustment step size. The instructions in the adjustment instruction stream are arranged in chronological order.

[0199] In some embodiments, step S550 may specifically include the following steps S551-S556:

[0200] Step S551: Perform time alignment processing on the target working parameter sequences of the primary conversion path and the secondary conversion path, so that the target working parameters of the two paths have a corresponding relationship at the same control time. If there is a deviation in the control time, the parameter interpolation is performed based on the earlier control time.

[0201] Check if the control times in the target operating parameter sequences of the primary and secondary conversion paths are consistent. If there is a discrepancy in control times, perform parameter interpolation using the earlier control time. Parameter interpolation is a method of estimating unknown data points using known data points. For control times with time discrepancies, calculate the target operating parameter value at that time based on the known target operating parameter values ​​before and after the discrepancy. Through time alignment, ensure that the target operating parameters of the primary and secondary conversion paths correspond at the same control times.

[0202] Step S552: Calculate the parameter change for the target working parameter value at each control moment. Calculate the difference between the current parameter value and the target parameter value for the primary conversion path and the secondary conversion path respectively, and generate the parameter change. The sign of the parameter change indicates the direction of parameter adjustment, and the absolute value indicates the adjustment magnitude.

[0203] After completing the time alignment process, for each control moment's target operating parameter value, the parameter change needs to be calculated. This parameter change reflects the magnitude and direction of the change required to adjust from the current parameter value to the target parameter value. The differences between the current and target parameter values ​​are calculated for both the primary and secondary conversion paths. For each operating parameter (PWM duty cycle, switching frequency, and inductor current threshold), the difference between its current and target values ​​is calculated. The sign of the difference indicates the direction of parameter adjustment: a positive value indicates an increase in parameter value, and a negative value indicates a decrease in parameter value; the absolute value of the difference represents the magnitude of the adjustment. All parameter changes at each control moment are combined to generate the parameter change value. This parameter change value provides specific data for subsequent adjustment step size calculations and instruction splitting.

[0204] Step S553: ​​Calculate the step size for adjusting the parameter change. Based on the preset maximum single adjustment amount and the absolute value of the parameter change, calculate the number of steps required for parameter adjustment. If the absolute value of the parameter change is less than the maximum single adjustment amount, the number of steps is 1; otherwise, the number of steps is the rounded-up value of the absolute value of the change divided by the maximum single adjustment amount.

[0205] After obtaining the parameter change, to ensure the stability and safety of parameter adjustment, it is necessary to calculate the adjustment step size. The purpose of the adjustment step size calculation is to determine the number of steps required to adjust the current parameter value to the target parameter value. The preset maximum single adjustment amount is a maximum adjustment range pre-set based on the performance and stability requirements of the primary and secondary conversion paths. The absolute value of the parameter change is compared with the maximum single adjustment amount. If the absolute value of the parameter change is less than the maximum single adjustment amount, it means that the parameter adjustment can be completed in one adjustment, and the step number is 1. If the absolute value of the parameter change is greater than the maximum single adjustment amount, the adjustment process needs to be divided into multiple steps, with the step number being the floor value of the absolute value of the change divided by the maximum single adjustment amount. In this way, it is ensured that the magnitude of each adjustment does not exceed the maximum single adjustment amount, thus guaranteeing the stability of parameter adjustment.

[0206] Step S554: Perform instruction splitting for parameter adjustment at each control moment. The parameter change that requires multiple steps of adjustment is split into multiple adjustment instructions. Each adjustment instruction contains a single adjustment step size, and the number of adjustment instructions is the same as the calculated number of steps.

[0207] For parameter adjustments at each control moment, if multi-step adjustments are required (i.e., more than one step), the parameter change needs to be split into commands. The purpose of command splitting is to divide a large parameter change into multiple smaller adjustment commands, each containing a single adjustment step size. Based on the calculated number of steps, the parameter change is evenly distributed across each adjustment command. Each adjustment command contains a single adjustment step size, calculated based on the parameter change and the number of steps. The number of adjustment commands is the same as the calculated number of steps. By splitting commands, a large parameter adjustment process can be decomposed into multiple smaller adjustment steps, making parameter adjustments smoother and more precise, and avoiding damage to the power supply or affecting stability due to excessive adjustment amplitude.

[0208] Step S555: Perform timestamp allocation processing on the split adjustment instructions, assign a specific execution timestamp to each adjustment instruction, and divide the control period evenly or according to preset rules into multiple sub-intervals equal to the number of adjustment steps, so that the adjustment instructions are distributed within the control period, so that the parameters are smoothly adjusted within the control period.

[0209] Each adjustment command is assigned a specific execution timestamp, and the control period is divided into multiple sub-intervals equal to the number of adjustment steps, either evenly or according to preset rules. If evenly divided, each sub-interval is of equal length; if using preset rules, the length of the sub-intervals can be adjusted according to actual conditions. Each adjustment command is assigned to its corresponding sub-interval, and the start time of that sub-interval is assigned as its execution timestamp. This method ensures that adjustment commands are evenly distributed within the control period, guaranteeing smooth parameter adjustments. The timestamp allocation ensures that adjustment commands are executed in the correct time sequence, avoiding adjustment chaos or delays.

[0210] Step S556: Sort all adjustment instructions according to the order of their execution timestamps to generate a working parameter adjustment instruction stream. Each instruction in the instruction stream contains an execution timestamp, path identifier, parameter type, target parameter value, and single adjustment step size, and the instructions are arranged in chronological order.

[0211] After completing the timestamp allocation process, all adjustment instructions are sorted according to their execution timestamp order. This sorting ensures the instructions are arranged chronologically, guaranteeing that the control unit executes parameter adjustments in the correct order at each control moment. All adjustment instructions are arranged in ascending order of execution timestamps, generating a working parameter adjustment instruction stream. Each instruction in the stream contains information such as execution timestamp, path identifier, parameter type, target parameter value, and single adjustment step size. The execution timestamp specifies the execution time of the instruction; the path identifier indicates whether the instruction targets the primary or secondary conversion path; the parameter type specifies whether the parameter to be adjusted is the PWM duty cycle, switching frequency, or inductor current threshold; the target parameter value is the specific parameter value to be adjusted at that moment; and the single adjustment step size indicates the magnitude of each adjustment.

[0212] Through sorting, the generated operating parameter adjustment command stream has a clear time sequence. The control unit can execute parameter adjustment operations sequentially according to the commands in the command stream, thereby achieving precise adjustment of the operating points of the primary and secondary conversion paths and ensuring dynamic matching between the power supply output power and load changes.

[0213] Step S560: Input the operating parameter adjustment command stream into the control unit of the PC power supply. The control unit executes the parameter adjustment operation sequentially according to the time sequence in the command stream, changes the operating point of the primary conversion path and the secondary conversion path, so that the actual output power tracks the change of the target power sequence and achieves dynamic matching with the load change.

[0214] The control unit is the core component of a PC power supply. It is responsible for receiving and executing various control commands to adjust and control the power supply's operating state. The control unit executes parameter adjustment operations sequentially according to the time order in the command stream. For each command, the control unit accurately adjusts the operating parameters of the primary and secondary conversion paths based on information such as the execution timestamp, path identifier, parameter type, target parameter value, and single adjustment step size contained in the command.

[0215] By changing the operating points of the primary and secondary conversion paths, the actual output power of the power supply can track changes in the target power sequence. When load demand changes, the target power sequence adjusts accordingly, and the control unit adjusts the operating parameters in a timely manner according to the instructions in the command stream, so that the power supply output power changes accordingly, achieving dynamic matching with load changes.

[0216] This application provides a computer-readable storage medium storing computer-executable instructions or a computer program. When the computer-executable instructions or the computer program are executed by a processor, the processor will execute the AI-predictive PC power load regulation method provided in this application. For example, ... Figure 2 The PC power load regulation method based on AI prediction is shown.

[0217] In some embodiments, the computer-readable storage medium may be a read-only memory (ROM), random access memory (RAM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), flash memory, magnetic surface memory, optical disk, or CD-ROM, etc.; or it may be a device that includes one or any combination of the above-mentioned memories.

[0218] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of this application are included within the scope of protection of this application.

Claims

1. A PC power load regulation method based on AI prediction, characterized in that, The method includes: The system continuously collects power supply signals from the PC power output terminal, processor operating status signals, and graphics component working status signals. The collected signals are correlated and integrated according to time sequence to generate load time sequence data. The load time sequence data includes signal acquisition time markers and time sequence change correlation information of each signal. Pattern recognition is performed on the load time series data to separate the power consumption signatures corresponding to different computing tasks, thereby obtaining a set of power consumption signatures associated with the load tasks. The power consumption signature set associated with the load task is input into the pre-trained load prediction model, and the load intensity change profile in the future power supply cycle is obtained through time series analysis. The load intensity change profile contains the time series pattern information of load change. The load intensity change profile is fed back and calibrated with the current dual-power supply status information of the PC power supply to generate a power allocation strategy. This power allocation strategy adapts to the temporal pattern of the load intensity change profile. Specifically, this includes: dividing the load intensity change profile into time granularity; discretizing the continuous load intensity change profile into load intensity values ​​at multiple time nodes according to a preset control cycle, with each time node corresponding to a control moment and equal intervals between time nodes; extracting state parameters from the current dual-power supply status information of the PC power supply, obtaining the current output power, voltage fluctuation range, current stability, and efficiency parameters of the main power supply circuit and the auxiliary power supply circuit respectively, generating a parameter set characterizing the dual-power supply status, the parameter set containing a parameter item consisting of the product of the number of power supply circuits and the number of parameter types; and modeling the correlation between the load intensity values ​​at each time node and the dual-power supply status parameter set, generating a load-to-power ratio by calculating the ratio of the load intensity value at each time node to the total power of the dual power supply. The sequence, where each element of the load-to-supply ratio sequence represents the ratio of load demand to power supply capacity at the corresponding time node; a threshold comparison is performed on the load-to-supply ratio sequence, and time nodes with ratios greater than a preset threshold are marked as power adjustment nodes, while the remaining time nodes are marked as power maintenance nodes, generating a node type marking sequence, the length of which is the same as the number of time nodes; power allocation calculation is performed on each power adjustment node, based on the load intensity value and the set of dual-supply status parameters, by solving an optimization problem that maximizes power supply efficiency, the target power values ​​of the main power supply circuit and the auxiliary power supply circuit at that time node are determined, the target power values ​​satisfying the constraint that the total power of the dual-supply circuit equals the load intensity value; the target power values ​​or maintenance power values ​​of all time nodes are arranged in chronological order to generate a power allocation strategy containing the dual-supply power allocation scheme for each control moment, the power value change pattern in the power allocation strategy being consistent with the temporal pattern of the load intensity change profile; The operating points of the primary and secondary conversion paths within the PC power supply are adjusted according to the power distribution strategy to ensure that the power output power dynamically matches the load changes.

2. The method according to claim 1, characterized in that, The process of performing pattern recognition on the load time-series data to separate the power consumption signatures corresponding to different computing tasks yields a set of power consumption signatures associated with the load tasks, including: The load timing data is segmented into event-triggered segments. The sudden change points of the power supply signal, processor running status signal and graphics component working status signal are monitored respectively. The load state switching time is determined by comprehensively judging the load state switching time. The continuous load timing data is segmented at the state switching time to obtain multiple independent event-driven segment units. Each event-driven segment unit contains a complete signal change cycle from the start of one state switching time to the end of the next state switching time. For each event-driven segment unit, multi-domain feature fusion is performed to simultaneously extract time-domain waveform features, frequency-domain spectrum features, and signal correlation features to generate a three-dimensional feature tensor. The dimension of the three-dimensional feature tensor is consistent with the number of signal types, the number of feature types, and the segment length. Feature selection is performed on the three-dimensional feature tensor. By calculating the mutual information value between each feature dimension and the computation task type, feature dimensions with mutual information values ​​higher than a preset threshold are retained to generate a dimensionality-reduced feature matrix. The row dimension of the dimensionality-reduced feature matrix corresponds to the number of event-driven fragment units, and the column dimension corresponds to the number of filtered feature dimensions. Density peak clustering is performed on the reduced feature matrix. By identifying regions in the feature space where the local feature point distribution density exceeds a preset density threshold as cluster centers, event-driven fragment units that are less than the truncation distance from the cluster center are assigned to the corresponding clusters to obtain multiple clusters. Each cluster contains a set of event-driven fragment units with similar features. For each cluster, a feature prototype is constructed. The feature prototype of the cluster is obtained by calculating the feature median vector of all event-driven fragment units within the cluster. Each element of the feature prototype corresponds to the median value of a feature type. The feature prototype is matched with a pre-stored computing task feature library for similarity matching. By calculating the similarity between the feature prototype and the task feature template, the computing task type corresponding to each cluster is determined. The feature prototype is marked as the power consumption signature of the computing task type. All marked power consumption signatures together form a power consumption signature set associated with the load task. Each power consumption signature is associated with the corresponding computing task through a unique task type identifier.

3. The method according to claim 2, characterized in that, The process of performing multi-domain feature fusion on each event-driven segment unit, simultaneously extracting time-domain waveform features, frequency-domain spectral features, and inter-signal correlation features, and generating a three-dimensional feature tensor includes: Time-domain waveform features are extracted from various signals in the event-driven segment unit. By calculating the rise time, fall time, pulse width, and duty cycle of the signal sequence, a time-domain feature vector describing the time-domain characteristics of the signal is generated. The number of elements in the time-domain feature vector is the same as the number of signal types. Frequency domain spectral features are extracted from various signals in the event-driven segment unit. By converting the time-domain signal to the frequency domain, the amplitude and phase of the signal at multiple preset frequency points are calculated to generate a frequency domain feature vector describing the frequency domain characteristics of the signal. Cross-correlation analysis is performed on different types of signals in the event-driven segment unit. The cross-correlation coefficients between the power supply signal and the processor running status signal and the graphics component working status signal are calculated, as well as the cross-correlation coefficients between the processor running status signal and the graphics component working status signal. A correlation feature vector describing the correlation characteristics between signals is generated. The number of elements in the correlation feature vector is the number of signal type combinations. The time-domain feature vector, frequency-domain feature vector, and associated feature vector are organized according to feature type and treated as independent feature groups to generate a combined feature matrix containing the three feature types. The row dimension of the combined feature matrix corresponds to the number of signal types, and the column dimension is divided into three independent regions, which correspond to the element sets of the time-domain, frequency-domain, and associated feature groups, respectively. The number of columns in each region is equal to the number of elements in the corresponding feature group. The dimensions of each independent feature group in the combined feature matrix are reconstructed. The feature arrangement is adjusted according to the order of the number of signal types, the number of feature types, and the number of sampling points of the event-driven fragment unit to generate a three-dimensional feature tensor. Each dimension of the three-dimensional feature tensor corresponds to the signal type, feature type, and time sampling point, and the size of the dimension is consistent with the number of corresponding attributes.

4. The method according to claim 3, characterized in that, The density peak clustering of the dimensionality-reduced feature matrix is ​​performed by identifying regions in the feature space where the local feature point distribution density exceeds a preset density threshold as cluster centers. Event-driven fragment units that are less than the truncation distance from the cluster center are assigned to the corresponding clusters, resulting in multiple clusters, including: The pairwise distances are calculated for all event-driven fragment unit feature vectors in the reduced feature matrix. The distance values ​​between all feature point pairs are obtained by calculating the Euclidean distance, generating a symmetric distance matrix. The rows and columns of the distance matrix correspond to event-driven fragment units, and the matrix elements are the feature distances between two corresponding fragment units. The cutoff distance is determined based on the distance matrix. By statistically analyzing the frequency distribution of all distance values, the distance value that gives each feature point a preset number of neighbors on average is selected as the cutoff distance. The cutoff distance is used to define the local neighborhood range. For each event-driven fragment unit feature vector, the local feature point distribution density is calculated, and the number of other feature vectors that are less than the cutoff distance from the feature vector is counted. The counted number is used as the local feature point distribution density of the event-driven fragment unit, and a density vector is generated. The number of elements in the density vector is the same as the number of event-driven fragment units. For each event-driven fragment unit feature vector, perform relative distance calculation processing, find all other fragment units with higher density than the current fragment unit, calculate the minimum distance between the current fragment unit and these high-density fragment units as the relative distance, and if the current fragment unit has the highest density, use the maximum distance value as its relative distance to generate a relative distance vector. The density vector and the relative distance vector are jointly filtered to select event-driven fragment units whose density vector values ​​exceed a preset density threshold and whose relative distance vector values ​​exceed a preset distance threshold as cluster centers. Each cluster center represents an initial cluster. For event-driven fragment units that are not cluster centers, cluster assignment is performed. Each non-center fragment unit is assigned to the cluster of the nearest cluster center with a higher density than itself, resulting in multiple clusters containing fragment units with similar features. This ensures that each event-driven fragment unit belongs to only one cluster.

5. The method according to claim 1, characterized in that, The power consumption signature set associated with the load task is input into a pre-trained load prediction model, and the load intensity change profile within future power supply cycles is obtained through time series analysis. The load intensity change profile contains time series information on load changes, including: The power signature set associated with the load task is arranged in chronological order to generate a power signature sequence with a time sequence relationship. Each element of the power signature sequence corresponds to a power signature vector at a time point, and the order of the elements is consistent with the chronological order. The power consumption signature sequence is extended with temporal features. For the power consumption signature vector at each moment in the sequence, the difference between the vector and the power consumption signature vectors at a predetermined number of previous historical moments is calculated to obtain multiple historical difference vectors. The original power consumption signature vector at the current moment is concatenated with the multiple historical difference vectors in chronological order to generate the extended feature vector at that moment. The extended feature vectors at all moments are arranged in chronological order to form an extended power consumption signature sequence. The dimension of each vector in the extended power consumption signature sequence is equal to the dimension of the original power consumption signature vector multiplied by a predetermined value, where the predetermined value is the sum of a predetermined number of historical moments and 1. The extended power signature sequence is input into the feature encoding layer of the load prediction model. The input features are mapped to a high-dimensional feature space through multi-layer nonlinear transformation to generate an encoded feature sequence containing deep feature representation. The time length of the encoded feature sequence is the same as that of the extended power signature sequence, and the feature dimension is the encoding dimension preset by the model. Temporal dependency modeling is performed on the encoded feature sequence. By capturing the long-term and short-term dependencies between encoded features at different time points, a dependency feature sequence with temporal correlation is generated. Each feature vector of the dependency feature sequence contains feature information of the corresponding time point and its associated time points. The dependent feature sequence is input into the prediction output layer of the load prediction model. By predicting the features at future time points, a load intensity prediction sequence is generated. The time length of the load intensity prediction sequence corresponds to the time length of the future power supply cycle, and each element represents the load intensity prediction value at the corresponding time point. The load intensity prediction sequence is trend smoothed by calculating the rate of change of adjacent prediction values ​​and applying a moving average correction to prediction values ​​whose rate of change exceeds a preset range, thereby generating a continuous and smooth load intensity change profile. The load intensity change profile includes the corrected load intensity prediction value and the corresponding time stamp information.

6. The method according to claim 5, characterized in that, The temporal dependency modeling of the encoded feature sequence involves capturing the long-term and short-term dependencies between encoded features at different time points to generate a temporally correlated dependency feature sequence. Each feature vector of the dependency feature sequence contains feature information of the corresponding time point and its associated time points, including: The encoded feature sequence is divided into time windows, and the continuous encoded feature sequence is divided into multiple overlapping time window feature matrices according to a preset window size and sliding step size. Each time window feature matrix contains a preset number of continuous encoded feature vectors. Self-attention calculation is performed on the feature matrix of each time window. An attention weight matrix is ​​generated by calculating the attention weights between the encoded feature vectors at different positions within the window. The row and column dimensions of the attention weight matrix are the same as the window size, and the matrix element values ​​represent the correlation strength between the corresponding two position features. The attention weight matrix and the time window feature matrix are weighted and summed to generate a window-dependent feature vector containing feature association information within the window. The dimension of the window-dependent feature vector is the same as the dimension of the encoded feature vector. The window dependency feature vectors of all time windows are arranged in order to generate an intermediate dependency feature sequence. Each element of the intermediate dependency feature sequence corresponds to the window dependency feature vector of a time window, and the sequence length is equal to the number of time windows. The intermediate dependency feature sequence is subjected to bidirectional loop processing. The forward loop and the backward loop capture the forward time dependency and the reverse time dependency of the sequence respectively, generating a bidirectional dependency feature sequence. The bidirectional dependency feature sequence is fused with the feature representation of the corresponding time window extracted from the encoded feature sequence to generate the final dependency feature sequence, which contains both the original encoded features and the temporal dependency features. The dependent feature sequence is input into the prediction output layer of the load prediction model. A load intensity prediction sequence is generated by predicting features at future time points. The time length of the load intensity prediction sequence corresponds to the time length of the future power supply cycle. Each element represents the predicted load intensity value at the corresponding time point, including: The dependent feature sequence is truncated, and a preset number of dependent feature vectors at the end of the sequence are extracted as prediction input features. The time length of the prediction input features is related to the preset prediction step size of the model, and the feature dimension is the same as the feature dimension of the dependent feature sequence. The predicted input features are reorganized by dimensional substitution of the time dimension and feature dimension and adjustment of the feature arrangement order to generate a reorganized feature tensor that meets the input requirements of the prediction output layer. The recombined feature tensor is input into the first fully connected layer of the prediction output layer. Through linear transformation and nonlinear activation function processing, the recombined feature tensor is mapped to the hidden feature space to generate a hidden feature vector. The dimension of the hidden feature vector is the preset number of hidden layer neurons. The hidden feature vector is input into the second fully connected network of the prediction output layer. The hidden feature vector is mapped to the prediction output space through linear transformation to generate a preliminary prediction vector containing the predicted load intensity values ​​at multiple future time points. The length of the preliminary prediction vector is the same as the time length of the future power supply cycle. The preliminary prediction vector is modified by weighting the preliminary prediction value with the historical average load intensity value to generate the final load intensity prediction sequence. Each element of the load intensity prediction sequence is the modified load intensity prediction value at the corresponding time point, and the number of elements is consistent with the time length of the future power supply cycle.

7. The method according to claim 1, characterized in that, The power allocation calculation for each power adjustment node, based on the load intensity value and the set of dual-power supply status parameters, determines the target power values ​​for the main power supply circuit and auxiliary power supply circuit at that time node by solving an optimization problem that maximizes power supply efficiency. The target power values ​​satisfy the constraint that the total power of the dual power supply equals the load intensity value, including: Efficiency curve fitting is performed on the set of dual power supply status parameters. The efficiency parameters of the main power supply circuit and the auxiliary power supply circuit are extracted respectively. By fitting the efficiency values ​​at different power points into a continuous function, the main power supply efficiency curve and the auxiliary power supply efficiency curve are generated. The independent variable of the efficiency curve is the power value and the dependent variable is the efficiency value. Obtain a preset power allocation optimization objective function, which uses the power values ​​of the main power supply circuit and the auxiliary power supply circuit as optimization variables, and takes the highest total efficiency of dual power supply as the optimization objective. The total efficiency is calculated by weighted average of the efficiency curves of the main power supply circuit and the auxiliary power supply circuit, with the weight being the power proportion of the corresponding circuit. Obtain preset power allocation constraints, including the equality constraint that the total power of the dual power supply is equal to the load strength value, and the inequality constraints that the power values ​​of the main power supply circuit and the auxiliary power supply circuit are respectively within their minimum and maximum output power ranges. The minimum and maximum output power ranges are extracted from the dual power supply state parameter set. The optimization objective function is solved under constraints. By finding the optimal variable values ​​that maximize the objective function, the optimal power values ​​of the main power supply circuit and the auxiliary power supply circuit are obtained. The optimal power values ​​satisfy all constraints and maximize the overall efficiency. The optimal power value is verified for feasibility by checking whether it falls within the corresponding power adjustment range. If it exceeds the range, it is adjusted to the nearest boundary value to generate a feasible target power value. The feasible target power value is compared with the current power value in the dual-power supply status parameter set, and the power adjustment amount is calculated. If the adjustment amount is greater than the preset maximum allowable adjustment amount, the target power value is corrected according to the maximum allowable adjustment amount to improve the stability of the power adjustment process, and finally the target power value at this time node is determined.

8. A computer device, characterized in that, include: Memory is used to store executable instructions or computer programs. When a processor executes computer-executable instructions or computer programs stored in the memory, it implements the PC power load regulation method based on AI prediction as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which is loaded and executed by a processor to implement the AI-predictive PC power load regulation method as described in any one of claims 1 to 7.

Images