Method and system for dynamic energy-saving scheduling of computing power center for multi-modal load prediction

Abstract

The application discloses a power center dynamic energy-saving scheduling method and system for multi-modal load prediction, relates to the technical field of energy-saving scheduling, and comprises the following steps: based on the business load data of the power center in a stable state for a long time, performing long-tail fractal filtering, extracting the steady-state energy spectrum, and constructing a zero-drift reference surface; applying a small random phase disturbance to the zero-drift reference surface, and combining an entropy gradient amplification strategy to enhance the weak disturbance characteristics in the steady-state energy spectrum; and converting the enhanced precursor microwave fluctuation characteristics into a time series difference energy cloud chart. The application enhances micro-disturbance identification by constructing a zero-drift reference surface, combines multi-scale energy analysis and Bayesian confidence correction, realizes accurate prediction and quantification of sudden load, adopts fractal slicing and low-power parallel scheduling strategies when the load surges, ensures high response while reducing energy consumption, and is suitable for power center scenes with high performance and energy efficiency requirements.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of energy-saving scheduling, in particular to a dynamic energy-saving scheduling method and system for an algorithm center based on multi-modal load prediction. BACKGROUND

[0002] The dynamic energy-saving scheduling of an algorithm center based on multi-modal load prediction refers to constructing a multi-modal load prediction model by fusing multiple data sources (such as user request behavior, system running state, external environment parameters, and business type characteristics) to accurately predict the change trend of the computing resource demand of the algorithm center in the future period. On this basis, the switching strategy of the server resource, the load balancing scheme, and the energy consumption optimization configuration are dynamically adjusted to realize the fine matching of the algorithm resource and the business demand, thereby significantly reducing the energy consumption and operation cost on the premise of guaranteeing the service performance and response time. This method not only improves the timeliness and accuracy of the scheduling strategy, but also has the self-adaptive response ability to sudden load fluctuations, and is suitable for the green energy-saving management scene of edge computing, cloud computing, and large-scale data centers.

[0003] The prior art has the following disadvantages:

[0004] The existing multi-modal load prediction model often lacks the prediction ability of sudden load surge in an extremely short time when dealing with business types that are in a stable state for a long time. Since such abnormal forms are extremely rare in historical samples, the model is difficult to accurately capture the transient change characteristics, thereby causing the dynamic energy-saving scheduling mechanism to incorrectly execute the algorithm unit shutdown operation in the load surge stage. This problem will directly cause the high-energy-consumption computing unit to fail to restart in time at a critical moment, thereby causing the task queue to continue to backlog, the service response time delay to rapidly enlarge, and the business continuity and system stability of the algorithm center to be seriously affected.

[0005] The above information disclosed in the background section is only used to strengthen the understanding of the background of the present disclosure, and therefore it can include information that does not constitute the prior art known to those of ordinary skill in the art. SUMMARY

[0006] The purpose of the present application is to provide a dynamic energy-saving scheduling method and system for an algorithm center based on multi-modal load prediction, which realizes the accurate prediction and quantification of sudden load by constructing a zero-drift reference surface to enhance perturbation recognition, combining multi-scale energy analysis and Bayesian confidence correction, and adopting fractal slicing and low-power parallel scheduling strategy to ensure high response while reducing energy consumption, so as to solve the problems in the background technology.

[0007] In order to achieve the above purpose, the present application provides the following technical scheme: a dynamic energy-saving scheduling method for an algorithm center based on multi-modal load prediction, comprising the following steps:

[0008] Based on the business load data of the computing center in a stable state for a long time, long-tail fractal filtering is performed, the steady-state energy spectrum is extracted, and a zero-drift control base is constructed. The control base is used as a reference for subsequent amplification of abnormal load signals;

[0009] A small random phase disturbance is applied to the zero-drift control base, and an entropy gradient amplification strategy is combined to enhance the weak disturbance characteristics in the steady-state energy spectrum, thereby identifying the precursor fluctuation characteristics of potential burst loads;

[0010] The enhanced precursor fluctuation characteristics are converted into a time series difference energy cloud chart, and a multi-scale discrete saddle point detection method is used in the energy cloud chart to determine the trigger time window of potential load surge, which is used to define the boundary conditions for subsequent surge intensity quantification calculation;

[0011] In the surge trigger time window, a dynamic cumulative potential well integral operation is performed to obtain the intensity change trend of the load energy change process in this time period, which is mapped to an instantaneous burst intensity vector, and a preliminary load surge factor sequence is generated accordingly;

[0012] An adaptive Bayesian confidence rollback mechanism is applied to the load surge factor sequence for multiple rounds of confidence correction and periodic correction to obtain a burst load surge index comparable across operating periods;

[0013] The burst load surge index is compared with the preset load surge threshold item by item. If the prediction result meets the surge trigger condition, a fractal slicing operation is performed on the target business computing task, which is split into multiple parallel executable slice units and scheduled to multiple low-power computing units for parallel operation, realizing distributed load bearing processing of burst loads and avoiding full start of high-energy computing units.

[0014] Preferably, long-tail fractal filtering is performed based on the business load data of the computing center in a stable state for a long time, specifically including the following steps:

[0015] The load data of the target business is collected at 5-second intervals for at least 90 natural days, and the load data is subjected to band-pass filtering, Z-score normalization, and sliding window mean smoothing;

[0016] A long-tail fractal filtering convolution kernel based on a Pareto distribution is applied to the smoothed load curve, a segmented adaptive convolution operation is performed, and a Savitzky-Golay method is used for smoothing correction, combined with symmetric mirror padding to supplement the data at both ends;

[0017] The processed curve is converted to the frequency domain by fast Fourier transform, the main frequency energy component is extracted, and the statistical indicators are calculated to form a multi-feature energy spectrum line;

[0018] A two-dimensional matrix composed of daily energy features is constructed, and a LOWESS regression is performed in the time dimension to obtain a denoised and smoothed zero-drift control base surface as a reference for subsequent surge identification.

[0019] Preferably, the step of applying a small random phase perturbation on the zero-drift control base surface and combining with the entropy gradient amplification strategy specifically includes:

[0020] A normal distribution random perturbation sequence is generated based on the phase values of each frequency component in the zero-drift control base surface, and the weighted distribution is applied to the phase. An inverse Fourier transform is performed to generate a perturbed load curve sequence;

[0021] Frame sliding energy window analysis is performed on the load curve sequence, the local energy entropy and its time gradient are calculated, and an enhancement weight is applied to the entropy gradient value higher than the upper limit of the standard deviation;

[0022] The precursor response point group that meets the condition is extracted from the entropy gradient enhancement curve, the perturbation intensity of the precursor response point group is calculated, the response intensity curve is formed, and the low-frequency and high-frequency feature parameters are extracted by wavelet packet decomposition for subsequent burst load surge trend identification.

[0023] Preferably, the step of converting the precursor fluctuation feature into a time difference energy cloud chart and determining the surge trigger time window by using a multi-scale discrete saddle point detection method specifically includes:

[0024] The precursor perturbation response intensity curve is converted into a two-dimensional time-energy heat map, and the contrast of the low-energy area is enhanced by gamma transformation to construct a continuous time difference energy cloud chart;

[0025] A plurality of difference scale lengths are set, and a multi-scale energy difference sequence is calculated in a sliding window manner. After normalization, a time-scale-energy three-dimensional difference body is formed;

[0026] Second derivative operation is performed on the three-dimensional difference body, and the local extreme conjugate point group that repeatedly appears at multiple scales is extracted as a multi-scale saddle point, and a saddle point density map is constructed to screen a continuous saddle point group;

[0027] According to the head and tail positions of the saddle point group and the disturbance acceleration and energy mutation amplitude, it is determined whether the time interval exceeds the preset threshold value to determine the potential surge trigger time window.

[0028] Preferably, the step of performing dynamic cumulative potential well integral operation in the surge trigger time window specifically includes:

[0029] The perturbation response curve in the surge trigger time window is extracted, and a cumulative energy curve is constructed based on the square value of the perturbation intensity. Bezier interpolation function is used for fitting processing;

[0030] The accumulated energy process is mapped to an asymmetric hyperbolic tangent potential well function, the velocity and acceleration of the particles at each time are calculated, and the energy transition section with a sharp change in derivative is identified;

[0031] An instantaneous burst intensity vector is constructed in the energy transition section, the energy change rate per unit time is calculated, and a trend weight coefficient proportional to the acceleration is introduced;

[0032] The maximum value is extracted based on the burst intensity vector in a sliding time window, a preliminary load surge factor sequence is constructed, and a nonlinear enhancement mapping process is performed through a Sigmoid function.

[0033] Preferably, the step of applying an adaptive Bayesian confidence rollback mechanism to the load surge factor sequence specifically includes:

[0034] A prior model of Gaussian mixture distribution is constructed based on the load surge factor sequence of multiple historical periods, and is normalized to the [0, 1] interval;

[0035] The preliminary load surge factor in the current period is input into the prior model, the posterior probability is calculated, and the first round of confidence scaling correction is performed;

[0036] A sliding rollback window is constructed to perform multiple rounds of decreasingly weighted rollback iteration on the correction factor sequence, generating a continuous and smooth confidence factor sequence;

[0037] The sequence after rollback is standardized with zero mean and unit variance, and is scale-aligned with reference to the historical mean curve, to obtain a sudden load surge index.

[0038] Preferably, the step of comparing the sudden load surge index with the preset load surge threshold item by item and triggering task scheduling specifically includes:

[0039] The sudden load surge index is compared with the set load surge threshold at a second-level granularity, and when any sudden load surge index exceeds the corresponding load surge threshold, the surge triggering time is determined.

[0040] Preferably, the following operations are performed when the surge is triggered:

[0041] Based on the input dependency graph and concurrent structure of the task, a fractal slicing operation is performed, and the target task is recursively split into multiple-level slicing units with independent execution capability;

[0042] According to the delay sensitivity and power consumption weight of the slicing unit, combined with the current load and resource capability vector of the low-power computing unit, the minimum difference priority matching and distribution are completed;

[0043] During the parallel execution of the slices, the sudden load surge index and response delay are dynamically monitored, and if the lag or surge expansion is triggered, the task migration and resource expansion are triggered, otherwise the redundant nodes are released to realize energy saving optimization.

[0044] The power center dynamic energy-saving scheduling system of multi-modal load prediction comprises a steady-state modeling module, a disturbance enhancement module, a trigger window identification module, an explosive enhancement modeling module, a confidence calibration module, and a response scheduling module.

[0045] The steady-state modeling module, based on the long-term stable state of the business load data in the power center, performs long-tail fractal filtering, extracts its steady-state energy spectrum, and constructs a zero-drift reference surface.

[0046] The disturbance enhancement module applies a small random phase disturbance to the zero-drift reference surface, and combines an entropy gradient amplification strategy to enhance the weak disturbance characteristics in the steady-state energy spectrum.

[0047] The trigger window identification module converts the enhanced precursor microwave fluctuation characteristics into a time series difference energy cloud map, and uses a multi-scale discrete saddle point detection method in the energy cloud map to determine the trigger time window of potential load surge.

[0048] The explosive enhancement modeling module performs dynamic cumulative potential well integral operation within the explosive trigger time window to obtain the intensity change trend of the load energy change process in this period, maps it to an instantaneous explosive intensity vector, and generates a preliminary load surge factor sequence accordingly.

[0049] The confidence calibration module applies an adaptive Bayesian confidence rollback mechanism to the load surge factor sequence for multiple rounds of confidence correction and periodic correction to obtain a burst load surge index comparable across operating cycles.

[0050] The response scheduling module compares the burst load surge index with the preset load surge threshold item by item, and if the prediction result meets the surge trigger condition, performs fractal slicing operation on the target business computing task immediately, splits it into multiple parallel executable slice units, and schedules them to multiple low-power computing units for parallel operation.

[0051] In the above technical solution, the technical effects and advantages provided by the present application are as follows:

[0052] The present application introduces a zero-drift reference surface based on long-tail fractal filtering and energy spectrum construction, which enhances the identification ability of weak disturbances in stable business load. Compared with the existing method which relies on sliding window or short-term mean value judgment, the present application combines frequency domain and time domain analysis to extract the real energy structure, and amplifies the precursor signal through the micro-disturbance excitation mechanism to realize the early perception of the surge trend, and improves the burst response ability of the prediction model.

[0053] The present application constructs a multi-scale differential energy cloud chart and a discrete saddle point detection method, accurately determines the burst trigger time window, and performs dynamic cumulative potential well integration in the window, quantifies the energy aggregation process, and generates a strength vector and a burst factor sequence.

[0054] The present application avoids the centralized restart of traditional large-scale high-energy consumption computing resources based on the fractal slicing triggered by the sudden load burst index and the low-power parallel scheduling strategy. BRIEF DESCRIPTION OF DRAWINGS

[0055] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0056] Figure 1 The method flow chart of the power center dynamic energy-saving scheduling method for multi-modal load prediction of the present application.

[0057] Figure 2 The module schematic diagram of the power center dynamic energy-saving scheduling system for multi-modal load prediction of the present application. DETAILED DESCRIPTION

[0058] Example implementations will now be described more fully with reference to the accompanying drawings. Example implementations may, however, be implemented in many different forms and should not be construed as limited to the examples set forth herein; rather, these examples are provided so that this disclosure will be thorough and complete, and will fully convey the gist of each example to those skilled in the art.

[0059] The present application provides a power center dynamic energy-saving scheduling method for multi-modal load prediction as shown in Figure 1 The present application provides a power center dynamic energy-saving scheduling method for multi-modal load prediction as shown in

[0060] Based on the business load data in the stable state of the power center in the long term, long-tail fractal filtering is performed, the steady-state energy spectrum is extracted, and a zero-drift reference surface is constructed, and the reference surface is used as a reference for subsequent burst load anomaly signal amplification;

[0061] Based on long-tail fractal filtering, the steady-state energy spectrum of service load is extracted and a zero-drift control base surface is constructed. This method processes a large amount of load data generated by continuous running services in the computing power center, deeply describes the steady-state characteristics, and provides an accurate benchmark for subsequent surge identification. The specific operation steps are as follows:

[0062] First, select the average load collection data of a certain type of service in a natural day for not less than 90 seconds from the historical records of the target computing power center, and get a complete service load time series. In order to ensure that the analysis results are not affected by occasional anomalies, a band-pass filter is used to remove all components with a sampling frequency higher than 0.5Hz or lower than 0.0001Hz, and only the service fluctuation data in the middle frequency band is retained. Subsequently, the time series data is subjected to Z-score normalization, that is, the local mean of each sampling period is taken as the center, the data is centralized, and the local standard deviation is scaled, so that the data in different time periods can be compared uniformly. Then, a 5-minute sliding window method is used to smooth the normalized data, and the sliding step length in each window is 30 seconds. The data in the window is generated by a cubic spline interpolation method to generate a smooth curve, thereby maximizing the elimination of occasional jumps and short-term anomalies on the global analysis results.

[0063] After completion, the above processed smooth service load curve is input into the pre-defined long-tail fractal filtering calculation process. Specifically, first, a weight sequence with heavy-tailed distribution characteristics is constructed as a filtering convolution kernel, and a specific Pareto distribution is selected as the tail control factor, so that the filtering kernel has high response capability in the low frequency interval and quickly decays in the high frequency interval. One-dimensional convolution operation is performed on the original smooth curve using this distribution kernel to extract the long-term trend component point by point. Next, a segmented adaptive filtering control is applied to the trend curve, which automatically identifies the variance range of local fluctuations for each continuous time period and dynamically adjusts the heavy-tail parameters of the convolution kernel to further enhance the retention capability of self-similar detail structures. After the convolution operation is completed, the curve is modified once by Savitzky-Golay smoothing filter to eliminate the step-type fluctuations that may be introduced in the filtering process. When supplementing the insufficient samples at both ends of the curve, symmetric mirror filling is used, that is, the mirror values of the first 20 data points are used to supplement the starting end, and the reverse order values of the last 20 points are symmetrically supplemented to prevent boundary distortion.

[0064] The above-mentioned smooth long-tail fractal curve is converted to the frequency domain to obtain its steady-state energy spectrum. The conversion process uses fast Fourier transform (FFT) to map the time series to the frequency space to obtain the power intensity value corresponding to each frequency component. In order to prevent leakage error introduced by discrete Fourier transform at the cycle boundary, the original data is weighted using a Hamming window function. After conversion, the energy intensity of each frequency point is recorded, and all frequency energy values are linearly scaled to the same order of magnitude. Then, the energy distribution spectrum is drawn in order of frequency from low to high, the main frequency component is extracted, and the total energy percentage of the main frequency component in each day cycle is calculated. Further, the energy spectrum obtained for each day is averaged, and four statistical indicators of spectral peak position, spectral peak width, spectral peak skewness, and spectral peak kurtosis are calculated. Moving average and two-way exponential smoothing are performed on the above statistical indicators to keep them stable across cycles, and a multi-feature spectrum line representing the energy distribution rule of the service in the steady state is obtained.

[0065] Finally, a zero-drift control base is constructed based on the steady-state energy spectrum. The specific operation is as follows: a set of feature vectors is constructed for the energy spectrum values of each time period, and all feature vectors are aligned by day and stacked to form a two-dimensional matrix structure. In the matrix, each column represents the energy feature value of the same time point in different cycles. For each column, a local weighted regression method (LOWESS) is used to fit a smooth curve as the minimum offset reference value of the time point in multiple cycles, i.e., the energy spectrum control base point. The control base points of all time points are spliced into a complete curve in time sequence, which is the final constructed zero-drift control base. In order to ensure that the control base has good steady-state response ability, median filtering and five-point difference denoising are performed on the base to make it naturally robust to abnormal interference in all frequency bands. In the subsequent anomaly precursor identification process, the control base can be used as a reference line for amplitude comparison and disturbance offset detection to determine whether the current energy change exceeds the disturbance threshold allowed by the historical steady-state reference.

[0066] Through the above steps, a steady-state feature extraction mechanism for long-term stable service load is provided, which is significantly different from the existing technology which only uses the mean or maximum value in a fixed time window as the control reference. In contrast, the present method constructs a dynamic and responsive energy reference base in both the frequency and time domains, improving the resolution of abnormal fluctuations and laying the foundation for subsequent precursor identification and energy-saving scheduling strategy execution of sudden load surge.

[0067] A small random phase disturbance is applied to the zero-drift control base, and combined with an entropy gradient amplification strategy, the weak disturbance features in the steady-state energy spectrum are enhanced to identify the precursor micro-fluctuation features of potential sudden load;

[0068] To improve the ability to identify the precursor of sudden load surge, the method based on the combination of micro-random phase perturbation and entropy gradient amplification strategy is used to enhance the weak disturbance characteristics of steady-state energy spectrum. This method is based on the previously constructed zero-drift control base surface, and the controllable disturbance is introduced to activate the potential energy anomaly response channel. Then, the disturbance response is enhanced by the information entropy change trend, so as to extract the precursor fluctuation characteristics that are difficult to observe in the original steady state. The method includes the following steps:

[0069] Based on the phase value of each frequency component in the zero-drift control base surface, a group of random disturbance sequences with normal distribution characteristics is constructed. Specifically, the phase perturbation amplitude of each frequency point is randomly fluctuated in the range of [-0.03π, 0.03π], and the normal distribution N(0, σ²) is used to generate the perturbation value, where σ is the perturbation intensity control factor and is initially set to 0.01. In order to avoid the concentration of perturbation in the low or high frequency area, the frequency weighting processing is performed on the perturbation sequence, so that the perturbation energy is evenly distributed in the whole frequency spectrum. Then, the perturbation sequence is applied to the original phase value to obtain the perturbed complex frequency spectrum representation. Then, the inverse Fourier transform is used to restore the perturbed frequency spectrum to the time domain waveform, forming a new load curve sequence containing micro-perturbation. The above phase perturbation process is the key part of the active disturbance excitation mechanism, and its purpose is to break the masking effect of weak fluctuations in the steady-state energy spectrum, so that the slight mutations hidden in the original smooth curve can be visualized.

[0070] Secondly, the energy response analysis is performed on the load curve sequence obtained after the disturbance processing. The frame sliding energy window method is used to divide the curve into analysis windows with a fixed length of 200 data points, and the sliding step is set to 20 points. In each window, the local energy value of the curve is calculated, and the entropy value of the energy in the window is obtained. The calculation of entropy value is based on the definition of Shannon entropy. After normalizing the energy to probability distribution, the information entropy index is obtained. Then, the entropy gradient between consecutive windows is calculated in the time axis direction, that is, the entropy difference between adjacent windows divided by the window sliding time interval, forming an entropy gradient sequence. The greater the entropy gradient, the faster the information uncertainty hidden in the local load waveform rises, which usually indicates that there is a disturbance response anomaly in this area. In the experiment, the initial entropy gradient amplification factor is set to 1.5, and all entropy gradient values greater than the upper limit of the standard deviation are enhanced by weight to further highlight the weak disturbance signal.

[0071] Then, in the energy response curve enhanced by entropy gradient, the abnormal disturbance aggregation area is identified. The specific operation is: first derivative processing is performed on the enhanced curve, local extreme points are extracted, and points that meet the following conditions at the same time in 10 continuous sliding windows are selected: ① the entropy gradient is greater than 2 times the average value, ② the energy growth rate is greater than 1.2 times the historical average, ③ the disturbance phase continuously rises within 3 cycles. If the above three conditions are met, it is determined that the point is a precursor response point of potential sudden load surge. In the embodiment, in order to reduce the false positive probability, only the response point group that appears continuously is retained, and the spacing between the response points is required to be not greater than 30 data points, so as to ensure that the abnormal area presents aggregation, trend and continuity.

[0072] Finally, for the identified precursor response point group, the disturbance amplitude index of the corresponding period is calculated, and a comparable disturbance response strength curve is generated through standardization processing. The response curve takes time as the horizontal axis and the disturbance amplified strength value at each time as the vertical axis, forming a monitoring curve that can dynamically describe the change of the precursor characteristic strength. Further, the wavelet packet decomposition operation is performed on the curve, the low-frequency trend component and the high-frequency oscillation component are extracted, and the variation coefficient and the period fluctuation ratio are calculated, so as to use these parameters as discriminant inputs of the precursor characteristic quantity in subsequent time series modeling. Unlike the existing technology which relies on static abnormal value screening or maximum-minimum interval change method, the present application introduces phase disturbance and entropy gradient joint analysis, does not rely on any fixed threshold judgment logic, but relies on the energy structure response induced by active disturbance for discrimination, so it has stronger sensitivity, robustness and adaptability, and can identify the time region where the future surge behavior may occur at an early and small amplitude load change stage.

[0073] Through the complete process of disturbance excitation + entropy gradient enhancement + abnormal point group extraction + multi-feature structure analysis, the accurate identification of the precursor micro-fluctuation characteristics of potential sudden load surge is realized, which provides a reference basis for subsequent time window locking and dynamic scheduling response. It is different from the traditional load monitoring method based on experience threshold determination in technology, and has obvious creativity and engineering implementability.

[0074] The enhanced precursor micro-fluctuation characteristics are converted into a time series difference energy cloud chart, and a multi-scale discrete saddle point detection method is used in the energy cloud chart to determine the trigger time window of potential load surge, which is used to limit the boundary conditions of subsequent surge intensity quantification calculation;

[0075] To further mine the starting rules of potential sudden load surges from the precursor microwave fluctuation characteristics extracted in the previous stage, an analysis method is proposed, which converts the micro-disturbance intensity into a time-difference energy cloud map and accurately locks the potential surge trigger time window by combining a multi-scale discrete saddle point detection method. This method realizes local period boundary identification using the structural evolution rules of disturbance energy without relying on rule model assumptions. The specific steps include:

[0076] The precursor disturbance response intensity curve obtained after the phase disturbance and entropy gradient amplification in the previous stage is converted into a time-energy two-dimensional mapping. The specific steps are as follows: The intensity value of the precursor microwave fluctuation characteristic is segmented into horizontal axis at equal interval time, and the disturbance response intensity value in each time period is coded as vertical pixel value with color gradient. A two-dimensional energy spectrum is generated in the form of color scale heat map. In order to enhance the contrast of small disturbances in the visual level, gamma transformation is used to compress the low energy region of the heat map nonlinearly, and linear stretching is performed on the high energy region, so that its dynamic range is expanded to within the set visual threshold range. In the middle, in order to ensure time continuity, the time axis is smoothed with 10 seconds as the step, ensuring that there is no horizontal fault or visual jump in the structure of the image, thus forming a time-difference energy cloud map with time correlation and energy density readability.

[0077] Perform multi-scale discrete difference operation on the time-difference energy cloud map constructed above to capture energy change trends at different time length scales. The specific method is as follows: Set multiple difference scale lengths, such as 30 seconds, 60 seconds, 90 seconds, and 120 seconds, and calculate the local energy difference curve in the form of sliding window at each scale. In the difference calculation at each scale, the average energy of the next window is subtracted from the average energy of the previous window to obtain the local trend difference sequence, and the sign transformation and standardization normalization are performed to ensure that different scales can be compared across scales. Then, the difference results at multiple scales are stacked in three-dimensional data to build a time-scale-energy change three-dimensional difference body, where each section represents the energy mutation trend at a certain scale, facilitating subsequent comprehensive judgment of abnormal boundary.

[0078] In the generated three-dimensional difference body, a multi-scale discrete saddle point detection method is used to identify the potential starting position of the sudden surge. The specific steps include: first, the second derivative operation is performed on each scale section respectively to calculate the local slope change rate; then, the key turning points where the derivative value changes from positive to negative or from negative to positive are selected as the candidate point set; then, among all the candidate points, the conjugate point group that appears repeatedly in multiple scales and has a reversed energy gradient change direction is identified, which is defined as a multi-scale saddle point. In the embodiment, the saddle point is defined as a point that has local extreme values in both time and scale dimensions in a two-dimensional surface, i.e., the point is at the intersection of the change of energy structure in two dimensions. By constructing a saddle point density distribution map and setting a minimum density threshold, isolated saddle points are further screened out, and only the continuous saddle point community is retained as the warning signal area of potential surge triggering.

[0079] According to the position of the multi-scale saddle point group identified above, the triggering time window of the potential load surge is determined. The specific operation is: taking the first and last time coordinates in the saddle point group as the start and end boundaries, the disturbance response curve is traced back in the time period, and the disturbance growth rate and energy jump amplitude are further calculated. If the set threshold is met (such as the disturbance growth rate is 1.8 times the average rate within 3 seconds, and the energy mutation amplitude is more than 25% of the reference control value), the time period is formally defined as the triggering time window of the surge. This time window will be used as a boundary condition in the subsequent dynamic energy intensity modeling and burst index calculation, ensuring that the energy mapping process strictly corresponds to the actual surge period, thereby improving the credibility and accuracy of the load surge factor. Unlike the traditional method of using a fixed time slice or a mean sliding window to determine the analysis interval, the present application uses a multi-scale saddle point superposition analysis method to construct a highly adaptive burst boundary determination mechanism from the intersection point of energy structure changes, which has significant sensitivity and time domain recognition ability.

[0080] Through the steps of graphic transformation of precursor disturbance characteristics, multi-scale difference analysis, saddle point detection and time period determination, the whole process from disturbance enhancement to surge time boundary extraction is completely realized. This method does not depend on rule models or fixed threshold judgments, and can adapt to the burst feature recognition of multiple types of services in different load evolution paths, and is particularly suitable for identifying the surge starting point with no obvious quantity change characteristics but significant energy structure change signals.

[0081] In the surge triggering time window, dynamic cumulative potential well integral operation is performed to obtain the intensity change trend of the load energy change process in the time period, which is mapped into an instantaneous burst intensity vector, and a preliminary load surge factor sequence is generated accordingly;

[0082] To achieve the accuracy of the identified load energy change characteristics in the identified surge trigger time window, the specific solution is: based on dynamic cumulative potential well integral operation for intensity modeling, this method extracts the non-stationary cumulative characteristics in the change path by modeling the process of energy gradually gathering to the critical state in the microscopic nonlinear environment, and further converts it into a quantifiable and comparable instantaneous burst intensity vector, thereby constructing a preliminary load surge factor sequence. The technical process specifically includes the following sub-steps:

[0083] Extract the disturbance response curve in the surge trigger time window and construct the corresponding energy accumulation curve. The specific method is: the disturbance intensity value in this time period is constructed into an energy function at a sampling rate of once per second, and the value of each sampling point in the energy function is the square of the current disturbance amplitude, which is used to measure the instantaneous energy density of the disturbance. Then, starting from the beginning of the time window, the energy function is cumulatively summed to form an accumulated energy curve that grows over time. In order to improve the smoothness and analyticality of the curve, a five-order Bezier interpolation function is used to continuously fit the discrete cumulative curve, so that it has a clear first derivative and second derivative, in order to analyze the trend change in the subsequent.

[0084] On the basis of the above-mentioned cumulative energy curve, perform dynamic potential well mapping operation. The specific operation is: the energy accumulation process is equivalent to the trajectory of a virtual particle running along the bottom of the potential well, and the potential well function form is set as an asymmetric hyperbolic tangent surface, which is used to describe the behavior characteristics of slow energy growth before mutation and sharp energy rise at mutation. Each data point in the fitted curve is taken as a position node on the potential well curve, and its first derivative in the time axis direction is calculated as the particle velocity, and the second derivative is calculated as the particle acceleration. The relative relationship between the acceleration and the corresponding potential well shape is used to deduce the turning property of the current energy trend. Through this mechanism, the potential energy collapse point or the energy steep wall segment in the energy curve can be identified, that is, the key time position where the energy rises sharply and the derivative changes sharply, as a candidate area of the intensity transition section.

[0085] In the above-mentioned energy transition section, the instantaneous change rate corresponding to each time is extracted, and an instantaneous burst intensity vector is generated. Each component in the vector represents the absolute value of the energy density change rate per unit time at the corresponding time point, and the calculation method is the difference between adjacent sampling points of the cumulative energy curve divided by the sampling interval time. In order to enhance the ability of the intensity vector to describe the mutation trend, a trend weight function is introduced in the vector generation process, and the weight coefficient is proportional to the acceleration, so that the energy change of the time section is greater. After obtaining the complete burst intensity vector, it is normalized to limit its value range to 0~1, which is convenient for subsequent sequence calculation. The vector as a time distribution expression of the burst intensity process can clearly describe the intensity and concentration of the surge process.

[0086] According to the above-mentioned burst intensity vector, a preliminary load surge factor sequence is constructed. The factor sequence is based on the burst intensity vector, combined with the time axis for time coding, and adopts a sliding time window integration method. The maximum burst intensity in every 10 seconds is taken as the load surge factor corresponding to the window, and a continuous time-intensity sequence is obtained. In the present application, in order to make the load surge factor more discriminant, a nonlinear enhancement mapping is further applied to the factor sequence. The Sigmoid function is used to compress the small disturbance to below 0.3 and expand the mutation segment to above 0.7, so as to realize the prominent expression of the surge segment in visualization judgment and model training. The final obtained load surge factor sequence can not only be used for subsequent confidence assessment and threshold comparison judgment, but also can be used as the early warning input of asynchronous control signal to realize the preparation of feedforward resource scheduling.

[0087] Unlike the commonly used sliding average rate change or fixed slope analysis in the prior art, the dynamic potential well integration method adopted in the present application fully introduces nonlinear cumulative characteristics and trend acceleration judgment mechanism, avoids the problems of misjudgment in slow changing period and late judgment in rapid mutation in traditional methods, and can realize the prospective modeling of energy change trend and the accurate quantification of burst intensity. In actual application, the method shows stronger trend response and higher surge recognition rate, and can well adapt to the change characteristics of different business load curves. By constructing a nonlinear interaction model and a dynamic trend mapping mechanism, the dynamic measurement of energy change in the load surge stage is effectively completed, and the structure is successfully formed into a controllable intensity vector and a factor sequence, which provides accurate and quantitative support for the next stage of confidence judgment and scheduling strategy execution.

[0088] An adaptive Bayesian confidence rollback mechanism is applied to the load surge factor sequence to perform multiple rounds of confidence correction and periodic correction to obtain a burst load surge index that is comparable across operating periods.

[0089] To overcome the problem of non-comparability of load surge factors in distribution form and amplitude range under different operating periods, the present application introduces a multi-round confidence correction method based on an adaptive Bayesian confidence rollback mechanism. The method adaptively performs multi-stage rollback update by fusing the dynamic difference between the prior confidence model and the current observation distribution, thereby constructing a burst load surge index that is stable, continuous and discriminant in multiple periods. The specific implementation steps include the following links:

[0090] Firstly, a prior distribution model of the burst load surge factor is constructed. The model is based on the load surge factor sequence recorded in the past multiple business operation cycles, and Gaussian mixture distribution is used for fitting. The specific operation is as follows: the load surge factor sequence recorded in the past 30 natural days is subjected to time standardization processing, and the factor sequences in different cycles are uniformly mapped to the standardized time axis; then all the data are normalized, and the load surge factor is mapped to the closed interval [0, 1]; then the expectation maximization algorithm is used to model the load surge factor by Gaussian mixture, and the initial number of mixtures is set to 3, that is, three types of potential distribution of weak surge, medium surge and strong surge are fitted. The Gaussian mixture prior distribution is used as the basic probability model in the subsequent Bayesian update, which is used to measure the deviation between the current observation value and the historical statistical structure.

[0091] The preliminary load surge factor sequence actually observed in the current cycle is subjected to the first round of Bayesian confidence update. The specific method is as follows: the current load surge factor value is input into the prior distribution one by one, and the posterior probability of each mixed distribution component is calculated, that is, the confidence degree of each load surge factor in the “weak-medium-strong” three states. Then based on the maximum a posteriori estimation method, the surge state to which the load surge factor belongs at the current time is determined; and according to the confidence degree in this state, the original load surge factor is subjected to confidence scaling, that is, using the confidence probability as the weight, the load surge factor is multiplied to generate the first round of confidence load surge factor. The purpose of this stage of correction is to weaken the fluctuation signal with low confidence degree, and to strengthen the trend signal with high prior agreement degree, so as to improve the stability of burst identification.

[0092] Based on the above confidence load surge factor sequence, multiple rounds of rollback correction are performed to realize the smooth transition of the confidence structure on the time axis. The specific operation is as follows: a sliding rollback window is constructed, and the window length is set to 15 seconds. The weighted mean value of the load surge factor sequence in the window is calculated by using the reverse decreasing weighting method, which is used as the rollback target value of the current load surge factor. Each round of rollback is driven by the deviation between the target value and the current load surge factor value, and gradually approaches according to the set rollback learning rate. This step can be regarded as a time sequence self-adjusting process, so that the burst load surge index does not appear sharp fluctuation in continuous period, and the reliability of scheduling decision is avoided due to short-time surge false alarm. On this basis, not less than three rounds of iterative rollback are performed, the mean value in the sliding window is recalculated in each round, and the current load surge factor is updated, and finally a confidence rollback smooth sequence is formed.

[0093] The load surge factor sequence completing the confidence rollback is periodically standardized to generate a final burst load surge index. The correction process is designed to improve cross-period comparability, using daily mean alignment and variance scaling. Specifically, the mean and standard deviation of the current rollback load surge factor sequence are calculated and adjusted to a zero-mean unit-variance distribution. Then, a historical multi-day mean standard curve is introduced as a reference, and the current standardized factor is mapped to the same scale as the historical curve using linear interpolation. The final output burst load surge index is a comparable index curve with a value range of [0, 1], a stable trend, preserved mutation structure, and multiple rounds of confidence correction, which can be directly used for dynamic threshold comparison and strategy triggering. Unlike existing methods that use fixed moving averages or threshold judgments to determine surge behavior, the present invention uses a confidence framework based on Bayesian inference, introducing prior learning ability and dynamic rollback adjustment mechanism, so that the surge index not only has strong recognition ability within a single period, but also can achieve horizontal alignment and stable structure expression in multiple periods, with higher practicality and engineering value in dynamic scheduling scenarios.

[0094] This embodiment successfully builds a burst load surge index construction method with time series stability, statistical consistency and trend sensitivity by establishing a prior mixed distribution model, performing multiple rounds of confidence posterior update, constructing a sliding rollback window and performing cross-period standardization, which is significantly different from the coarse-grained detection mechanism in existing load prediction algorithms, and reflects high innovation and precision control ability.

[0095] The burst load surge index is compared with the preset load surge threshold item by item. If the prediction result meets the surge trigger condition, the fractal slicing operation is performed on the target business computing task, which is divided into multiple parallel execution slice units and scheduled to multiple low-power computing units for parallel running, realizing distributed bearing processing of burst load and avoiding full start of high-energy consumption computing units.

[0096] To solve the problem that the existing computing power scheduling mechanism can only respond to burst load surge by starting large-scale high-energy consumption computing resources, resulting in decreased energy efficiency and delayed resource response, the present embodiment provides a specific solution: task fractal slicing and low-power resource collaborative scheduling driven by burst load surge index. This method structures the load surge behavior as a quantifiable index, and when the burst load surge index exceeds the preset load surge threshold, it triggers a lightweight parallel task decomposition and differentiated resource matching mechanism, thereby achieving energy optimization of burst load bearing without sacrificing task response performance. The technical process includes the following steps:

[0097] The burst load surge index obtained by the Bayesian confidence rollback mechanism is compared with the preset load surge threshold. The comparison process traverses each value in the burst load surge index sequence with a time resolution of 1 second, and compares it with the load surge threshold determined by the business category, task priority, and resource level at the corresponding time point. The load surge threshold is dynamically adjusted based on multiple strategy parameters such as the response tolerance of the target business in historical surge behavior, safety redundancy, and maximum allowed scheduling delay. In this embodiment, the value range is usually between [0.65, 0.85]. If the burst load surge index at a certain time point exceeds the load surge threshold, it is judged that the load has entered the non-delayable scheduling stage at that time, and the fractal slicing and parallel scheduling process is triggered immediately. This method is different from the traditional fixed threshold alarm mechanism. Its comparison mechanism is based on the index-threshold mapping function and the strategy fusion function, and has high adaptability and elastic adjustment ability.

[0098] When the surge trigger condition is met, the fractal slicing operation is performed on the target business currently in the scheduling queue. This operation first analyzes the input dependency graph and internal call topology of the target task, identifies its splittable boundary and internal parallel channel; then according to the business execution logic, the entire task is divided into several input-output closed computing subunits according to the function structure, data dependency degree and logical concurrency, each subunit is called a first-level slicing unit. Then, according to the complexity and resource granularity requirement of the first-level slicing unit in the internal structure, the calculation intensive part is further sliced recursively to form a second-level slicing unit. In this embodiment, the depth-first slicing method is preferred to expand the sub-task, so as to convert the high-complexity task into a low-delay and weak-dependent micro-task slicing structure in the shortest time. The entire fractal slicing operation ensures that all slicing units still retain complete input-output chains after slicing and have independent execution ability, ensuring that subsequent parallel running on distributed resources does not appear scheduling blockage.

[0099] After completing the task fractal slicing, a matching sorting based on delay sensitivity and energy consumption weight is performed on all slice units, which are sequentially distributed to multiple low-power computing units for parallel scheduling. The specific distribution strategy is as follows: first, according to the current load state, the predicted available window and the processing capacity of each low-power computing unit, a resource capability vector set is constructed; then, the slice units are sorted according to the unit power consumption reachable computing amount, and are paired with the resource vector with the smallest difference; in the matching process, high response requirement slices are preferentially distributed to low delay computing units, and non-critical path slices are scheduled to the standby execution node with moderate performance but extremely low power consumption. To improve the execution efficiency, a short period re-distribution strategy is introduced when making resource scheduling decisions. Whenever the slice queue changes or the execution unit load fluctuation exceeds 10%, the matching table is re-optimized to ensure that resources and tasks are always efficiently coupled. This method is significantly different from the existing technology of pushing all burst tasks to the main computing unit for processing, avoiding the problems of cold start cost and waste of redundant resources.

[0100] During the parallel execution of slice tasks, the trend of sudden load surge index change and task execution feedback are continuously monitored, and the re-distribution strategy of slice tasks is dynamically adjusted. If the sudden load surge index is further increased during parallel processing, or the processing of part of the slices in the low-power unit lags behind the set tolerance (such as the delay exceeds 15% of the expected execution time), the high energy-efficient auxiliary execution unit is activated from the standby resource pool immediately, and the remaining unallocated slices or lagging slices are quickly switched and migrated. The migration strategy uses the head-preserving tail-replacing logic, that is, the processed part is preserved and the unfinished part is re-scheduled as a whole, thereby avoiding execution interruption. On the contrary, if the sudden load surge index shows a clear downward trend within 1 minute, and all current slices are within the controllable response interval, the excess low-power units are gradually released, so that the task bearing naturally converges to the minimum energy consumption in a stable state. The whole process maintains the scheduling balance of slice tasks in a dynamic environment, achieving both the processing capacity of sudden load and the energy saving goal.

[0101] The present embodiment is significantly superior to traditional strategy scheduling technology in terms of sudden task scheduling response timeliness, task resource distribution rationality and overall energy consumption control capability, through the dynamic threshold comparison mechanism based on the sudden load surge index, the fractal slicing processing technology with task structure perception, and the low-power distributed scheduling method driven by multiple resource capabilities. Especially in the edge computing and cloud-edge collaborative environment with limited resources and high response rigidity requirements, this method can achieve millisecond-level scheduling decision response and fragment-level task elastic placement, providing a replicable, scalable and engineered solution for sudden load scheduling in green computing scenarios, with high innovation, adaptability and practical application value.

[0102] To ensure full disclosure and implementability of the method of the present application, the following supplementary explanations are made for key algorithm parameters, embodiments, experimental data and formula variable definitions:

[0103] In the present application, the Pareto distribution parameter alpha value of the long-tailed fractal filter is verified by CEC2017 benchmark test, and the optimal value is determined as 1.5 under the comprehensive evaluation of multiple groups of frequency domain energy fidelity and noise suppression performance;

[0104] In an embodiment, the service load data collected by a certain cloud computing center for 90 consecutive days at an interval of 5 seconds is selected, band-pass filtered, Z-score normalized, smoothed by sliding mean, constructed on a zero-drift base surface, enhanced by perturbation, and detected by multi-scale saddle point to form a complete processing flow and output a time series difference energy cloud map;

[0105] The experimental results show that the F1 score of the multi-scale saddle point detection module reaches 0.92 in a typical burst load scenario, and the false positive rate of the surge prediction is less than 3% after combining with the adaptive Bayesian confidence rollback mechanism; in terms of formula definition, for example, the parameters A=1, B=0.5 and C=0 of the asymmetric hyperbolic tangent potential well function represent the trend weight coefficient, the acceleration weight coefficient and the translation correction, respectively, and all variable symbols are clearly defined and explained in the formula paragraph to ensure that the technical personnel can directly reproduce the present application scheme.

[0106] The present application significantly enhances the ability to identify weak disturbances in stable service load by introducing a zero-drift control base surface based on long-tailed fractal filtering and energy spectrum construction. Compared with existing prediction methods that only rely on fixed sliding windows or short-term mean fluctuations, the present application extracts the true energy structure under stable load from both frequency and time domains, and applies a perturbation excitation mechanism on this structure, so that the pre-surge precursor signal that is easily ignored is quantifiable and identifiable, thereby achieving earlier prediction response before the surge occurs, and improving the sensitivity and forward-looking nature of the prediction model to sudden load.

[0107] The present application precisely locks the trigger time window of load surge by constructing a multi-scale difference energy cloud map and a multi-scale discrete saddle point detection mechanism, and performs dynamic cumulative potential well integration within the window to quantify the load energy aggregation trend during the mutation, and generates an instantaneous surge intensity vector and a surge factor sequence. On this basis, an adaptive Bayesian confidence rollback mechanism is further introduced to realize cross-cycle confidence correction and obtain a sudden load surge index with time continuity and multi-cycle comparability. This combined method not only maintains high prediction accuracy, but also effectively avoids the problem of mischeduling caused by single-cycle overfitting and recognition bias of rare events, and improves the measurement accuracy and judgment robustness of sudden behavior.

[0108] The application is based on the burst load surge index trigger fractal slicing and multi-node low-power parallel scheduling strategy, which breaks the rigid mode of traditional computing power scheduling relying on the overall restart of large-scale main computing resources. When the surge is accurately identified, the task is immediately divided into multiple independent execution units, and is scheduled to multiple low-power nodes for collaborative processing according to the energy-performance trade-off logic, which not only ensures the timely response of the task, but also significantly reduces the overall energy consumption. Compared with the existing scheduling strategy, the application provides a more dynamic and energy-efficient optimization capability of green burst load processing mechanism, which is particularly suitable for edge computing, cloud edge fusion, large-scale resource pool and other scenes with dual requirements of energy efficiency and stability.

[0109] The application provides a multi-modal load prediction computing power center dynamic energy-saving scheduling system as shown in Figure 2 The application provides a multi-modal load prediction computing power center dynamic energy-saving scheduling system as shown in

[0110] The steady-state modeling module performs long-tail fractal filtering based on the business load data of the computing power center in a stable state for a long time, extracts the steady-state energy spectrum, and constructs a zero-drift reference surface.

[0111] The disturbance enhancement module applies a slight random phase disturbance to the zero-drift reference surface, and combines an entropy gradient amplification strategy to enhance the weak disturbance characteristics in the steady-state energy spectrum.

[0112] The trigger window identification module converts the enhanced precursor microwave fluctuation characteristics into a time sequence difference energy cloud map, and determines the trigger time window of the potential load surge in the energy cloud map by using a multi-scale discrete saddle point detection method.

[0113] The surge intensity modeling module performs dynamic cumulative potential well integral operation within the surge trigger time window to obtain the intensity change trend of the load energy change process in the time period, maps it into an instantaneous burst intensity vector, and generates a preliminary load surge factor sequence accordingly.

[0114] The confidence calibration module applies an adaptive Bayesian confidence rollback mechanism to the load surge factor sequence for multiple rounds of confidence correction and periodic correction to obtain a burst load surge index comparable across operating periods.

[0115] The response scheduling module compares the burst load surge index with the preset load surge threshold item by item, and if the prediction result meets the surge trigger condition, the fractal slicing operation is immediately performed on the target business computing task to divide it into multiple parallel executable slice units, and the slice units are scheduled to multiple low-power computing units for parallel running.

[0116] The multi-modal load prediction computing power center dynamic energy saving scheduling method provided by the embodiment of the application is implemented through the multi-modal load prediction computing power center dynamic energy saving scheduling system, and the specific method and process of the multi-modal load prediction computing power center dynamic energy saving scheduling system are described in the embodiment of the multi-modal load prediction computing power center dynamic energy saving scheduling method, which will not be described here.

[0117] The above merely describes certain exemplary embodiments of the application by way of illustration, and it is needless to say that those skilled in the art can modify the described embodiments in various manners without departing from the spirit and scope of the application. Therefore, the above drawings and descriptions are illustrative in nature and should not be understood as limiting the scope of protection of the claims of the application.

Claims

1. A method for dynamic energy saving scheduling of a power center for multi-modal load prediction, characterized in that, The method comprises the following steps: Based on the business load data of the computing center in the medium and long term in the stable state, long tail fractal filtering is performed, the steady state energy spectrum is extracted, and a zero drift control base is constructed; A micro random phase disturbance is applied to the zero drift control base, and an entropy gradient amplification strategy is combined to enhance the weak disturbance characteristics in the steady state energy spectrum; The enhanced precursor microwave fluctuation characteristics are converted into a time series difference energy cloud chart, and a multi-scale discrete saddle point detection method is used in the energy cloud chart to determine the trigger time window of potential load surge; In the trigger time window of the surge, a dynamic cumulative potential well integral operation is performed to obtain the intensity change trend of the load energy change process in the time window, which is mapped into an instantaneous burst intensity vector, and a preliminary load surge factor sequence is generated accordingly; The adaptive Bayesian confidence rollback mechanism is applied to the load surge factor sequence for multiple rounds of confidence correction and periodic correction to obtain a sudden load surge index comparable across operating periods; The sudden load surge index is compared with the preset load surge threshold item by item, and if the prediction result reaches the surge trigger condition, a fractal slicing operation is performed on the target business computing task to split it into multiple parallel executable slice units and schedule them to multiple low-power computing units for parallel operation.

2. The method of claim 1, wherein, Based on the business load data of the computing center in the medium and long term in the stable state, long tail fractal filtering is performed, specifically including the following steps: Collect the load data of the target business at 5-second intervals for at least 90 natural days, and perform band-pass filtering, Z-score normalization, and sliding window mean smoothing on the load data; A long tail fractal filtering convolution kernel based on Pareto distribution is applied to the smoothed load curve, a segmented adaptive convolution operation is performed, and a Savitzky-Golay method is used for smoothing correction, combined with symmetric mirror filling to supplement the data at both ends; The processed curve is converted to the frequency domain by fast Fourier transform, the main frequency energy component is extracted, and the statistical indicators are calculated to form a multi-feature energy spectrum line; A two-dimensional matrix composed of energy features of each day is constructed, LOWESS regression is performed in the time dimension to obtain a denoised and smoothed zero drift control base as a reference for subsequent surge identification.

3. The dynamic energy saving scheduling method of the power center of the multi-modal load prediction according to claim 1, wherein, The step of applying a micro random phase disturbance to the zero drift control base and combining an entropy gradient amplification strategy specifically includes: Based on the phase values of each frequency component in the zero drift control base, a normal distribution random disturbance sequence is generated, and after weighted distribution, it is applied to the phase to perform inverse Fourier transform to generate a disturbed load curve sequence; Frame sliding energy window analysis is performed on the load curve sequence to calculate the local energy entropy and its time gradient, and an enhancement weight is applied to the entropy gradient values higher than the upper limit of the standard deviation; The entropy gradient enhancement curve is extracted to obtain the disturbance intensity of the precursor response point group, form a response intensity curve, and perform wavelet packet decomposition to extract low-frequency and high-frequency feature parameters for subsequent burst load surge trend identification.

4. The method of claim 3, wherein, The step of converting the precursor microwave fluctuation characteristics into a time series difference energy cloud chart and using a multi-scale discrete saddle point detection method to determine the surge trigger time window specifically includes: The precursor disturbance response intensity curve is converted into a two-dimensional time-energy heat map, and the contrast of the low-energy area is enhanced by gamma transformation to construct a continuous time-difference energy cloud image; A plurality of difference scale lengths are set, a plurality of scale energy difference value sequences are calculated in a sliding window manner, and a time-scale-energy three-dimensional difference body is formed after normalization; Second derivative operation is performed on the three-dimensional difference body, and a group of local extreme conjugate points that repeatedly appear at multiple scales are extracted as multi-scale saddle points, and a saddle point density map is constructed to screen continuous saddle point groups; Whether the time interval exceeds the preset threshold is determined according to the head and tail positions of the saddle point group, and the energy mutation amplitude and the disturbance increase rate, and the potential burst triggering time window is determined.

5. The dynamic energy saving scheduling method of the power center of the multi-modal load prediction according to claim 1, wherein, The steps of performing dynamic cumulative potential well integration operation in the burst triggering time window specifically include: Extracting the disturbance response curve in the burst triggering time window, constructing a cumulative energy curve based on the square value of the disturbance intensity, and fitting the cumulative energy curve by using a Bezier interpolation function; Mapping the cumulative energy process to an asymmetric hyperbolic tangent potential well function, calculating the velocity and acceleration of particles at each time, and identifying the energy transition section with a sharp change in derivative; In the energy transition section, an instantaneous burst intensity vector is constructed, the energy change rate per unit time is calculated, and a trend weight coefficient proportional to the acceleration is introduced; Based on the burst intensity vector, the maximum value is extracted in a sliding time window, a preliminary load burst factor sequence is constructed, and a nonlinear enhancement mapping process is performed by using a Sigmoid function.

6. The dynamic energy saving scheduling method of the power center of the multi-modal load prediction according to claim 5, characterized in that, The steps of applying an adaptive Bayesian confidence rollback mechanism to the load burst factor sequence specifically include: A prior model of Gaussian mixture distribution is constructed based on the load burst factor sequences of multiple historical periods, and is normalized to the [0, 1] interval; The preliminary load burst factor in the current period is input into the prior model, the posterior probability is calculated, and the first round of confidence scaling correction is performed; A sliding rollback window is constructed to perform multiple rounds of decreasingly weighted rollback iteration on the correction factor sequence to generate a continuous and smooth confidence factor sequence; The sequence after rollback is standardized to zero mean and unit variance, and the scale is aligned with reference to the historical mean curve to obtain a sudden load burst index.

7. The dynamic energy saving scheduling method of the power center of the multi-modal load prediction according to claim 6, characterized in that, The steps of comparing the sudden load burst index with the preset load burst threshold item by item and triggering task scheduling specifically include: The sudden load burst index is compared with the set load burst threshold at a second-level granularity, and when any sudden load burst index exceeds the corresponding load burst threshold, the burst triggering time is determined.

8. The method of dynamic energy saving scheduling of the power center for multi-modal load prediction according to claim 7, wherein, When the burst is triggered, the following operations are performed: Based on the input dependency graph and concurrent structure of the task, fractal slicing operation is performed, and the target task is recursively split into multiple-level slicing units with independent execution capability; According to the delay sensitivity and power consumption weight of the slicing unit, and combining the current load and resource capability vector of the low-power computing unit, the minimum difference priority matching and distribution are completed; During the slicing parallel execution process, the sudden load burst index and response delay are dynamically monitored, and if the lag or burst expansion is triggered, task migration and resource expansion are triggered, otherwise, redundant nodes are released to realize energy saving optimization.

9. The computing power center dynamic energy-saving scheduling system of multi-modal load prediction, used to realize the multi-modal load prediction computing power center dynamic energy-saving scheduling method in any one of claims 1-8, characterized in that, The method comprises a steady-state modeling module, a disturbance enhancement module, a trigger window identification module, an intensive enhancement modeling module, a confidence calibration module, and a response scheduling module. The steady-state modeling module performs long-tail fractal filtering based on the business load data of the computing center in a stable state for a long time, extracts the steady-state energy spectrum, and constructs a zero-drift reference surface. The disturbance enhancement module applies a small random phase disturbance to the zero-drift reference surface and combines an entropy gradient amplification strategy to enhance the weak disturbance characteristics in the steady-state energy spectrum. The trigger window identification module converts the enhanced precursor microwave fluctuation characteristics into a time series difference energy cloud chart, and determines the trigger time window of potential load surge by using a multi-scale discrete saddle point detection method in the energy cloud chart. The intensive enhancement modeling module performs dynamic cumulative potential well integral operation within the trigger time window of the surge to obtain the intensity change trend of the load energy change process within the time window, maps it to an instantaneous burst intensity vector, and generates a preliminary load surge factor sequence accordingly. The confidence calibration module applies an adaptive Bayesian confidence rollback mechanism to the load surge factor sequence for multiple rounds of confidence correction and periodic correction to obtain a sudden load surge index comparable across operating cycles. The response scheduling module compares the sudden load surge index with the preset load surge threshold item by item. If the prediction result meets the surge trigger condition, the fractal slicing operation is immediately performed on the target business computing task to split it into multiple parallel executable slice units and schedule them to multiple low-power computing units for parallel operation.

Images