Deep learning-based computer power supply ripple interference suppression optimization method and system

Abstract

The application provides a kind of computer power ripple interference suppression optimization method and system based on deep learning, by the ripple signal of computer power output end is carried out time domain-frequency domain joint collection analysis, generates time domain-frequency domain joint ripple spectrum feature mapping and power working state associated information, input ripple countermeasure generation network, generates virtual compensation ripple signal, with load dynamic fluctuation time sequence information input load-ripple dynamic correlation model, generates load adaptive real-time compensation waveform, same power output energy efficiency time sequence feedback information and ripple residual monitoring information input ripple countermeasure generation network carries out hierarchical iterative optimization, based on weight optimized ripple countermeasure generation network, ripple suppression self-adaptive regulation link is built ripple real-time monitoring link linkage power output, iteratively adjusts the phase, amplitude and time sequence configuration information of load adaptive real-time compensation waveform.The application can improve the stability and adaptability of computer power ripple interference suppression.

Description

A Deep Learning-Based Optimization Method and System for Suppressing Computer Power Supply Ripple Interference Technical Field

[0001] This invention relates to the field of deep learning, and more specifically, to a deep learning-based method and system for optimizing computer power supply ripple interference suppression. Background Technology

[0002] Computer power supply ripple interference suppression is used to reduce the adverse effects of power supply output ripple signals on downstream computer hardware, and is an important support for maintaining stable hardware operation and ensuring equipment performance. Existing ripple interference suppression solutions typically acquire the ripple signal at the power supply output through signal acquisition components, generate a compensation signal using preset signal processing logic, and then inject the compensation signal into the output link through a dedicated circuit to cancel the ripple. The compensation signal generated by existing solutions often fails to accurately match the ripple characteristics when the power supply's operating state changes, resulting in fluctuations in ripple suppression effectiveness. In scenarios with dynamic load changes, the adjustment rhythm of the compensation signal is not sufficiently adapted to load fluctuations, making it impossible to maintain a stable suppression effect. At the same time, some solutions do not fully utilize the various types of correlation information generated during power supply operation when optimizing the compensation signal, making it difficult for the optimization direction of the compensation signal to align with the actual operating requirements of the power supply system, thus affecting the overall efficiency and reliability of ripple suppression. Summary of the Invention

[0003] This invention provides a deep learning-based optimization method and system for suppressing computer power supply ripple interference.

[0004] In a first aspect, embodiments of the present invention provide a deep learning-based optimization method for suppressing computer power supply ripple interference. The method includes: jointly acquiring and analyzing the ripple signal at the output of the computer power supply in the time and frequency domains; synchronously associating it with power supply operating status information to enhance its spectral features, generating a joint time-domain and frequency-domain ripple spectral feature map and power supply operating status association information; inputting the joint time-domain and frequency-domain ripple spectral feature map and power supply operating status association information into a ripple adversarial generation network; through the collaborative interaction between the feature encoding branch and the compensation signal generation branch, combined with a cross-branch attention fusion mechanism, performing phase inverse modulation and amplitude adaptive calibration to generate a phase inverse calibration virtual compensation ripple signal; and combining the virtual compensation ripple signal with real-time acquired load dynamic waveforms. The dynamic timing information is input into the load-ripple dynamic correlation model. By matching the correlation between ripple characteristics and load fluctuations, the timing configuration information of the compensation signal is adjusted to generate a load-adaptive real-time compensation waveform. The power supply output energy efficiency timing feedback information and ripple residual monitoring information are acquired and input into the ripple adversarial generation network along with the load-adaptive real-time compensation waveform. Based on the dual-objective optimization criterion, the weights inside the network are iteratively optimized in a hierarchical manner to obtain the weight-optimized ripple adversarial generation network. Based on the weight-optimized ripple adversarial generation network, a ripple suppression adaptive adjustment link is constructed in conjunction with the power supply output ripple real-time monitoring link. Feedback calibration is continuously initiated through ripple residual monitoring information to iteratively adjust the phase, amplitude, and timing configuration information of the load-adaptive real-time compensation waveform.

[0005] Secondly, embodiments of the present invention provide a computer system, including: a memory storing a computer program; and a processor for loading the computer program to implement the above-mentioned deep learning-based computer power supply ripple interference suppression optimization method.

[0006] This invention performs joint time-domain and frequency-domain acquisition and analysis of the ripple signal at the computer power supply output, and synchronously associates it with the power supply operating status information to enhance its spectral features. This allows for a more comprehensive capture of the time-domain variation and frequency-domain distribution characteristics of the ripple signal, providing a more accurate foundation for subsequent ripple processing. The joint time-domain and frequency-domain ripple spectral feature mapping and power supply operating status association information are input into a ripple adversarial generation network. Through the collaborative interaction of the feature encoding branch and the compensation signal generation branch, phase inverse modulation and amplitude adaptive calibration are performed to generate a virtual compensation ripple signal. This allows the phase and amplitude of the generated compensation signal to better match the characteristics of the original ripple, improving the matching degree between the compensation signal and the original ripple. The virtual compensation ripple signal and the real-time acquired load dynamic fluctuation timing information are input into the load-ripple dynamic correlation model to adjust the timing parameters of the compensation signal, generating a load-adaptive real-time compensation signal. The compensation waveform allows for adaptive adjustment to the dynamic fluctuations of the load, avoiding the problem of fixed-sequence compensation failing to adapt to load changes. It acquires power output energy efficiency timing feedback information and ripple residual monitoring information, and inputs them in conjunction with the load-adaptive real-time compensation waveform into a ripple adversarial generation network. This network performs hierarchical iterative optimization of internal weights, allowing the weight adjustments to better align with the actual energy efficiency and ripple suppression requirements of the power supply, improving the adaptability of the generated compensation signal. Based on the weight-optimized ripple adversarial generation network, a ripple suppression adaptive adjustment link is constructed in conjunction with the real-time power output ripple monitoring link. Continuous feedback calibration is triggered by ripple residual monitoring information, allowing for continuous dynamic adjustment of the phase, amplitude, and timing parameters of the compensation waveform. This maintains the stability of the ripple interference suppression effect and avoids the problem of decreased suppression effect due to changes in power supply operating conditions. Attached Figure Description

[0007] Figure 1 is a flowchart of a deep learning-based optimization method for suppressing computer power supply ripple interference, provided by an embodiment of the present invention.

[0008] Figure 2 is a schematic diagram of the composition of a computer system provided in an embodiment of the present invention. Detailed Implementation

[0009] Please refer to Figure 1, which is a flowchart of a deep learning-based optimization method for suppressing computer power supply ripple interference according to an embodiment of the present invention. This method can be executed by a computer system and may include the following steps:

[0010] Step S100: Perform time-domain and frequency-domain joint acquisition and analysis on the ripple signal at the computer power supply output terminal, synchronously associate it with the power supply operating status information to enhance the spectral features, and generate a time-domain and frequency-domain joint ripple spectral feature mapping and power supply operating status association information.

[0011] Time-domain analysis focuses on how the ripple signal changes over time, such as its amplitude, period, rising edge, and falling edge. By observing the time-domain signal, we can understand the intensity and trend of the ripple signal at different times. Frequency-domain analysis, on the other hand, focuses on the frequency components of the signal. By converting the ripple signal from the time domain to the frequency domain, we can determine which frequency components are included in the ripple signal and the intensity of each frequency component.

[0012] To achieve joint time-domain and frequency-domain acquisition and analysis, professional instruments such as oscilloscopes are used to acquire ripple signals in the time domain. The oscilloscope can display the waveform of the ripple signal in real time and record the signal's amplitude and time information. The acquired time-domain signal is then converted to the frequency domain using a Fast Fourier Transform (FFT) algorithm.

[0013] Power supply operating status information includes multiple parameters such as input voltage, output current, and operating temperature, which are closely related to the ripple signal. When the input voltage of the power supply changes, the characteristics of the ripple signal also change; the magnitude of the output current affects the load condition of the power supply, which in turn affects the amplitude and frequency of the ripple signal.

[0014] Spectral feature enhancement involves processing the acquired ripple signal spectrum to highlight its key features for subsequent analysis and processing. Filtering algorithms can be used to remove noise and interference components from the spectrum while retaining the main frequency characteristics of the ripple signal. Feature extraction methods can also be employed to extract features such as peak values ​​and bandwidth from the ripple signal spectrum, further enhancing the signal's characteristic representation.

[0015] When generating the joint time-domain and frequency-domain ripple spectrum feature map, the feature information from the time and frequency domains is integrated to form a comprehensive feature map. The power supply operating state correlation information associates the power supply's operating state parameters with the characteristics of the ripple signal, establishing a correspondence. Through this correlation, the impact of the power supply's operating state on the ripple signal can be better understood, allowing for appropriate measures to suppress ripple.

[0016] Step S200: Input the time-frequency domain joint ripple spectrum feature mapping and power supply operating state correlation information into the ripple adversarial generation network. Through the collaborative interaction between the feature encoding branch and the compensation signal generation branch, and combined with the cross-branch attention fusion mechanism, perform phase inverse modulation and amplitude adaptive calibration to generate a phase inverse calibration virtual compensation ripple signal.

[0017] For example, step S200 may specifically include the following steps S210 to S250:

[0018] Step S210: Input the time-domain-frequency domain joint ripple spectrum feature map into the input layer of the feature coding branch, and perform channel splitting on the time-domain-frequency domain joint ripple spectrum feature map. The channels are evenly split according to the proportion of time-domain dimension and frequency-domain dimension in the feature map to obtain two parallel feature sub-maps. One feature sub-map corresponds to the time-domain dimension information and contains all time-related channel information in the time-domain-frequency domain joint ripple spectrum feature map. Each channel corresponds to a time-domain related feature dimension. The other feature sub-map corresponds to the frequency-domain dimension information and contains all frequency-domain related channel information in the time-domain-frequency domain joint ripple spectrum feature map. Each channel corresponds to a frequency-domain related feature dimension.

[0019] In the input layer, the joint time-frequency domain ripple spectral feature map is split into channels. This is because the features in the time and frequency domains have different properties and functions, and processing them separately allows for better extraction of their respective key information.

[0020] When performing channel splitting, the proportion of channels in the feature map to the proportions of the time-domain and frequency-domain dimensions is first determined. This can be determined based on extensive experimental data and experience to ensure that the split feature map accurately reflects the feature information in both the time and frequency domains. Then, the channels of the feature map are evenly divided into two parts according to this proportion. The resulting time-domain feature map contains all time-related channel information from the time-frequency joint ripple spectrum feature map. Each channel corresponds to a specific time-domain related feature dimension, such as the amplitude variation of the signal over time, the periodicity of the signal, etc.

[0021] The frequency domain dimensional feature sub-mapping contains all channel information related to the frequency domain. Each channel corresponds to a frequency domain-related feature dimension, such as the intensity and frequency distribution of different frequency components. Frequency domain feature information can help us understand the frequency composition of the ripple signal and identify the frequency components that have a significant impact on power supply stability.

[0022] Step S220: Input the power supply operating state association information into the state embedding layer of the feature encoding branch, perform sequence alignment on the power supply operating state association information, adjust the sequence length of the power supply operating state association information to be consistent with the sequence length of the feature sub-mapping, and obtain a state embedding sequence that matches the dimension of the feature sub-mapping. Each element of the state embedding sequence corresponds to a channel position of the feature sub-mapping, and the dimension of each element is consistent with the dimension of the corresponding channel of the feature sub-mapping.

[0023] The power supply operating state association information includes various state parameters of the power supply during operation, such as input voltage, output current, and operating temperature. The role of the state embedding layer of the feature encoding branch is to process the power supply operating state association information so that it can be effectively fused with the feature sub-map. Since the sequence lengths of the power supply operating state association information and the feature sub-map may differ, a sequence alignment operation is required. The purpose of sequence alignment is to adjust the sequence length of the power supply operating state association information to be consistent with the sequence length of the feature sub-map, which can be achieved through interpolation or truncation methods.

[0024] If the sequence length of the power supply operating state association information is less than the sequence length of the feature submap, interpolation can be used to insert new elements into the sequence to increase its length. Interpolation methods can determine the value of the inserted element based on the values ​​of existing elements and their relationships; for example, linear interpolation calculates the linear relationship between existing elements to insert an appropriate value. If the sequence length of the power supply operating state association information is greater than the sequence length of the feature submap, truncation can be used to retain only the first or second part of the sequence, making its length consistent with the feature submap.

[0025] After sequence alignment, a state embedding sequence matching the dimension of the feature map is obtained. Each element of the state embedding sequence corresponds to a channel position of the feature map, and the dimension of the element is consistent with the dimension of the corresponding channel of the feature map.

[0026] Step S230: The feature sub-maps of the time domain dimension information, the feature sub-maps of the frequency domain dimension information, and the fusion layer of the state embedding sequence input feature encoding branch are associated and concatenated element by element. The corresponding elements of the three inputs are concatenated in the order of time domain, frequency domain, and state to obtain the pre-encoding fusion information. The number of channels of the pre-encoding fusion information is the sum of the number of channels of the three inputs, and the length of each element is the sum of the lengths of the corresponding elements of the three inputs.

[0027] The main task of the fusion layer in the feature encoding branch is to fuse the feature sub-maps of the time domain, the feature sub-maps of the frequency domain, and the state embedding sequence to form a comprehensive feature representation. Element-by-element association and concatenation is the specific method to achieve this fusion.

[0028] During concatenation, the corresponding elements of the three inputs are connected sequentially in the order of time domain, frequency domain, and state. For each element, the corresponding elements in the time domain feature sub-map, the frequency domain feature sub-map, and the state embedding sequence are concatenated together in sequence.

[0029] Since the fused information before encoding is composed of three inputs, its channel count is the sum of the channel counts of the three inputs. This means that the fused information contains more feature dimensions and can more comprehensively reflect the combined characteristics of the ripple signal and the power supply operating state. At the same time, the element length at each position is also the sum of the element lengths at the corresponding positions of the three inputs, further enriching the feature information at each position.

[0030] Step S240: Input the pre-encoding fused information into the convolutional processing layer of the feature encoding branch and perform multi-step convolution operations. First, perform preliminary convolution through the first convolution kernel to extract local information, and then perform secondary convolution through the second convolution kernel to extract global correlation information, so as to obtain the encoded feature representation. The number of channels of the encoded feature representation is equal to the sum of the number of the first convolution kernel and the second convolution kernel. Each channel corresponds to one extracted feature.

[0031] The convolutional processing layer of the feature encoding branch plays a crucial role in the entire feature encoding process. It extracts and transforms features from the fused information before encoding through multiple convolutional operations.

[0032] The first step is preliminary convolution using the first convolutional kernel. The main function of the first convolutional kernel is to extract local information from the pre-encoding fused information. Local information includes features such as signal edges, textures, and small-scale fluctuations. During the convolution operation, the first convolutional kernel slides across the pre-encoding fused information, extracting the features of the local region through convolution operations with its elements. This process is similar to identifying edges and textures in an image, determining edges of different intensities and directions by calculating with the convolutional kernel and the local regions of the image.

[0033] After the initial convolution, an intermediate result containing local features is obtained. Next, a second convolution is performed using a second convolution kernel. The size of the second convolution kernel is typically larger than that of the first kernel, and its purpose is to extract global correlation information. This global correlation information reflects the relationships and interactions between different local features. During the second convolution, the second convolution kernel scans a wider area of ​​the intermediate result after the initial convolution, integrating and correlating features from different local regions to obtain a higher-level global feature.

[0034] The final encoded feature representation has the same number of channels as the sum of the number of the first and second convolutional kernels. This is because each convolutional kernel extracts a feature during the convolution process, and different channels correspond to different feature types.

[0035] Step S250: Input the encoded feature representation into the output layer of the feature encoding branch, perform dimensionality compression, and compress the number of channels of the encoded feature representation to be consistent with the receiving dimension of the compensation signal generation branch through pointwise convolution to obtain the output information of the feature encoding branch. Transmit the output information of the feature encoding branch to the receiving layer of the compensation signal generation branch through the data transmission line to start the collaborative interaction between the feature encoding branch and the compensation signal generation branch.

[0036] Pointwise convolution is an effective method for dimensionality reduction, using a 1×1 kernel. It compresses the number of channels by linearly combining each channel of the encoded feature representation. During pointwise convolution, the kernel performs convolution operations on elements of each channel, fusing information from multiple channels into fewer channels, thus reducing the number of channels. After dimensionality compression, the output information of the feature encoding branch is obtained. This information contains the encoded and compressed features, enabling more efficient transmission and processing within the network.

[0037] Then, the output information of the feature-encoded branch is transmitted to the receiving layer of the compensation signal generation branch via a data transmission line. The data transmission line can be an internal computer bus or other data transmission channel to ensure accurate and rapid information transmission. After receiving the output information of the feature-encoded branch, the two branches begin to interact collaboratively. This collaborative interaction enables the compensation signal generation branch to utilize the feature information extracted by the feature-encoded branch to further generate a virtual compensation ripple signal with phase inversion calibration, thereby suppressing power supply ripple interference.

[0038] For example, in step S250, the output information of the feature coding branch is transmitted to the receiving layer of the compensation signal generation branch through the data transmission line, and the collaborative interaction between the feature coding branch and the compensation signal generation branch begins. Specifically, this may include the following steps S251 to S255:

[0039] Step S251: Input the output information of the feature coding branch into the receiving layer of the compensation signal generation branch, perform dimensional transformation on the output information of the feature coding branch, and adjust the dimension of the output information of the feature coding branch to be completely consistent with the internal processing dimension of the compensation signal generation branch through a linear transformation matrix, so as to obtain the received information consistent with the internal processing dimension of the compensation signal generation branch.

[0040] The receiving layer of the compensation signal generation branch is a crucial node for information exchange between the feature encoding branch and the compensation signal generation branch. After receiving the output information from the feature encoding branch, it performs a dimensionality transformation to ensure that the information can be correctly processed in the compensation signal generation branch.

[0041] The linear transformation matrix is ​​the core tool for achieving dimensionality transformation. The elements of the linear transformation matrix are designed based on the dimensionality requirements of the feature-encoded branch output information and the internal processing of the compensation signal generation branch. By performing matrix multiplication on the feature-encoded branch output information and the linear transformation matrix, the dimensionality of the feature-encoded branch output information can be adjusted.

[0042] During matrix multiplication, each row of the linear transformation matrix is ​​multiplied element-wise with each column of the feature encoding branch output information, and the results are summed to obtain new element values. In this way, the dimension of the feature encoding branch output information is transformed to be consistent with the internal processing dimension of the compensation signal generation branch.

[0043] After dimensional transformation, the received information is obtained that is consistent with the processing dimension within the compensation signal generation branch. This received information can better adapt to the processing requirements of the compensation signal generation branch, ensuring the smoothness and effectiveness of information exchange between the two branches.

[0044] Step S252: Input the received information into the cross-branch attention fusion layer of the compensation signal generation branch, call the cross-branch attention fusion mechanism to associate and match the output information of the feature encoding branch with the current internal state information of the compensation signal generation branch, first extract the key association points of the two, then calculate the matching degree of each association point to obtain the cross-branch attention weight distribution, and each element of the cross-branch attention weight distribution corresponds to the matching degree of an association point.

[0045] For example, step S252 may specifically include the following steps S2521 to S2526:

[0046] Step S2521: Extract the key information sequence from the output information of the feature coding branch. Extract the maximum response value in each window as the key information element through a sliding window to obtain the key information sequence. The length of the key information sequence is consistent with the sequence length of the output information of the feature coding branch.

[0047] Within each window, the maximum response value is identified as the key information element. The maximum response value represents the strongest feature of the signal within that window, potentially corresponding to important information such as signal peaks or abrupt changes. By sequentially extracting the maximum response value within each window and arranging these values ​​according to the sliding order of the windows, a key information sequence is obtained. Since each window corresponds to a position in the output information of the feature encoding branch, the length of the key information sequence is consistent with the length of the feature encoding branch output information sequence. This effectively highlights important features in the feature encoding branch output information, reduces interference from noise and irrelevant information, and provides more representative information for association matching and matching degree calculation.

[0048] Step S2522: Extract the query information sequence from the current internal state information of the compensation signal generation branch. Extract the maximum response value in each window as the query information element through the same sliding window to obtain the query information sequence. The length of the query information sequence is consistent with the sequence length of the internal state information.

[0049] This step is similar to step S2521, except that the object of operation is changed to the current internal state information of the compensation signal generation branch. The same sliding window setting is used to slide on the current internal state information of the compensation signal generation branch.

[0050] Within each window, the maximum response value is extracted as the query information element. Similarly, the maximum response value reflects the strongest feature of the internal state information within that window. Arranging these query information elements according to the sliding window sequence yields the query information sequence. Since each window corresponds to a position within the internal state information, the length of the query information sequence is the same as the length of the internal state information sequence. By using the same sliding window to extract both the key information sequence and the query information sequence, consistency in structure and information extraction methods is ensured, providing comparability for association matching and helping to more accurately determine the key correlation points between the output information of the feature encoding branch and the internal state information of the compensation signal generation branch.

[0051] Step S2523: Input the key information sequence and the query information sequence into the computation layer of the cross-branch attention fusion mechanism, perform pairwise similarity calculation on the key information sequence and the query information sequence, calculate the cosine similarity between each key information element and each query information element, and obtain the initial similarity matrix. The number of rows in the initial similarity matrix is ​​equal to the length of the key information sequence, and the number of columns is equal to the length of the query information sequence.

[0052] The computational layer of the cross-branch attention fusion mechanism is responsible for calculating the similarity between the key information sequence and the query information sequence. The pairwise similarity calculation means comparing the similarity between each element in the key information sequence and each element in the query information sequence.

[0053] Cosine similarity is a commonly used similarity metric that measures the degree of similarity between two vectors by calculating the cosine of the angle between them. In this step, each element in the key information sequence and the query information sequence is treated as a vector, and the cosine similarity between them is calculated.

[0054] Specifically, for each element in the key information sequence and each element in the query information sequence, their dot product is calculated, and then divided by the product of their moduli to obtain the cosine similarity value between the corresponding elements. Arranging all these cosine similarity values ​​into a matrix yields the initial similarity matrix. The number of rows in the initial similarity matrix equals the length of the key information sequence, since each row corresponds to one element in the key information sequence; the number of columns equals the length of the query information sequence, since each column corresponds to one element in the query information sequence. Each element in this matrix represents the similarity between the element in the corresponding row of the key information sequence and the element in the corresponding column of the query information sequence, providing the basis for weight calculation and information fusion.

[0055] Step S2524: Normalize the row dimension of the initial similarity matrix by dividing the similarity value of each row by the sum of the row values ​​to obtain the normalized similarity matrix. The sum of the elements in each row of the normalized similarity matrix is ​​1.

[0056] When performing row-level normalization, for each row of the initial similarity matrix, the sum of all elements in that row is calculated. Then, each element in that row is divided by this sum to obtain a new element value. After this processing, the sum of the elements in each row becomes 1, forming the normalized similarity matrix. Row-level normalization allows the elements in each row to reflect their relative importance within the overall matrix. For example, if an element in a row has a high similarity value, after normalization, its proportion in that row will also be larger, indicating a stronger association between that element and its corresponding element in the query information sequence. The normalized similarity matrix is ​​more suitable for subsequent weighted integration operations, as it more accurately reflects the relationship between the key information sequence and the query information sequence.

[0057] Step S2525: Smooth the column dimensions of the normalized similarity matrix by using a weighted average of adjacent columns to smooth the matrix elements, resulting in a smoothed similarity matrix with less fluctuation in the elements.

[0058] Column smoothing aims to reduce fluctuations in elements within the normalized similarity matrix, making the matrix elements more stable and continuous. Weighted averaging of adjacent columns is the primary method for achieving smoothing. During smoothing, for each column of the normalized similarity matrix, the element values ​​of its adjacent columns are considered. Based on preset weights, a weighted average is calculated for the elements in the adjacent columns. For example, for an element in a particular column, its own value, the value of its left adjacent column, and the value of its right adjacent column can be added together according to certain weights to obtain a new element value.

[0059] The preset weights can be set according to specific circumstances. For example, an equal-weighted average can be used, where adjacent columns have the same weight; or a Gaussian weighted average can be used, assigning different weights based on the distance from the current column, with closer columns receiving larger weights. In this way, the element values ​​of adjacent columns influence each other, reducing the fluctuation of matrix elements and resulting in a smoothed similarity matrix. A smoothed similarity matrix is ​​more conducive to subsequent analysis and processing, avoiding errors caused by excessive element fluctuations.

[0060] Step S2526: The smoothed similarity matrix is ​​used as the cross-branch attention weight distribution and output to the subsequent processing layer of the compensation signal generation branch. The subsequent processing layer will perform weighted integration processing based on this weight distribution.

[0061] After receiving the cross-branch attention weight distribution, the subsequent processing layer of the compensation signal generation branch performs a weighted integration of the received information and the current internal state information of the compensation signal generation branch based on this weight distribution. The weighted integration process involves summing each element in the received information and internal state information according to the cross-branch attention weight distribution. For example, for a certain correlation point, the element value of that correlation point in the received information is multiplied by the corresponding weight, and then the element value of that correlation point in the internal state information is multiplied by the corresponding weight. The two products are then added together to obtain the weighted integrated element value. In this way, the output information of the feature coding branch and the internal state information of the compensation signal generation branch can be more effectively fused, improving the quality of the generated virtual compensation ripple signal.

[0062] Step S253: Based on the cross-branch attention weight distribution, the received information and the compensation signal generation branch's current internal state information are weighted and integrated. The received information element and the internal state information element of each associated point are multiplied according to their corresponding weights and then summed to obtain the fused state information. The dimension of the fused state information is consistent with the dimension of the received information.

[0063] During weighted integration, for each associated point, the element of that associated point in the received information and the element of that associated point in the internal state information are multiplied by their respective weights, and then the two products are added together. Specifically, each associated point in the received information and internal state information is traversed, and the corresponding element is processed according to its weight. For example, for the first associated point, the element value of that associated point in the received information is multiplied by its corresponding weight to obtain one product; the element value of that associated point in the internal state information is multiplied by its corresponding weight to obtain another product. These two products are then added together to obtain the fused element value of that associated point.

[0064] By performing this operation on all associated points, the fused state information is obtained. Since the processing of each associated point is based on elements at corresponding positions in the received information and internal state information, the dimension of the fused state information is consistent with the dimension of the received information. The fused state information integrates the characteristics of the received information and the internal state information, and is weighted according to the importance of the associated points, which is more conducive to subsequent signal processing and analysis, and provides more effective information support for generating high-precision virtual compensated ripple signals.

[0065] Step S254: Input the fused state information into the phase modulation layer of the compensation signal generation branch, perform phase inversion modulation, and reverse the phase according to the phase-related information in the fused state information to obtain the phase-modulated intermediate signal. The phase of the phase-modulated intermediate signal is opposite to the phase of the original ripple signal.

[0066] The main task of the phase modulation layer in the compensation signal generation branch is to perform phase inversion modulation on the fused state information to generate a compensation signal with a phase opposite to that of the original ripple signal. Phase inversion modulation is one of the key steps in suppressing power supply ripple interference because when the phase of the compensation signal is opposite to that of the original ripple signal, they can cancel each other out during superposition, thereby effectively reducing the amplitude of the ripple.

[0067] When performing phase-reverse modulation, phase-related information must first be extracted from the fused state information. This phase-related information can be obtained through phase analysis of the signal, for example, by using a Fourier transform-based method to convert the signal to the frequency domain and analyze the phase information of the frequency domain signal. Then, based on the extracted phase information, the phase is reversed. Specifically, the phase value is increased by 180 degrees (in degrees) or by π (in radians) to achieve phase reversal. Based on this, the resulting new phase is opposite to the phase of the original ripple signal.

[0068] Based on the inverted phase information, the fused state information is adjusted to obtain the phase-modulated intermediate signal. This phase-inverse modulation method effectively utilizes the compensation signal to cancel the original ripple signal, improving power supply stability.

[0069] Step S255: Input the phase-modulated intermediate signal into the amplitude calibration layer of the compensation signal generation branch, perform amplitude adaptive calibration, compare and adjust according to the amplitude information of the intermediate signal with the preset amplitude range, and obtain the virtual compensation ripple signal with phase reverse calibration. The amplitude of the virtual compensation ripple signal matches the amplitude of the original ripple signal.

[0070] The preset amplitude range is obtained through statistical analysis of the amplitudes of a large number of original ripple signals, ensuring that the compensation signal is neither too large nor too small. Before amplitude calibration, amplitude information needs to be extracted from the phase-modulated intermediate signal. Amplitude information can be obtained by calculating the absolute value of the signal, which reflects the strength of the signal.

[0071] The amplitude information of the intermediate signal is compared with a preset amplitude range. If the amplitude of the intermediate signal exceeds the preset range, adjustment is required. If the amplitude is greater than the preset maximum value, it can be reduced by a certain proportion; if the amplitude is less than the preset minimum value, it can be increased by a certain proportion. This adjustment process is adaptive, dynamically adjusting according to the actual amplitude of the intermediate signal. After amplitude adaptive calibration, a virtual compensation ripple signal with phase inversion calibration is obtained. This signal not only has the opposite phase to the original ripple signal, but its amplitude also matches the original ripple signal. When superimposed with the original ripple signal, it can more effectively cancel ripple, improving the stability and quality of the power supply output.

[0072] Step S300: Input the virtual compensation ripple signal and the real-time acquired load dynamic fluctuation timing information into the load-ripple dynamic correlation model, and adjust the compensation signal timing configuration information by matching the correlation between ripple characteristics and load fluctuations to generate a load-adaptive real-time compensation waveform.

[0073] Ripple characteristics include the amplitude, phase, and frequency of the ripple signal, reflecting its fundamental properties. Load dynamic fluctuation timing information records dynamic information such as power and current changes in the load at different times. Correlation matching involves analyzing and comparing the ripple and load fluctuation characteristics to determine the correlation points between them. For example, when the load power suddenly increases, the amplitude of the ripple signal may also increase accordingly; this is a correlation point.

[0074] Based on the correlation matching results, the timing configuration information of the compensation signal is adjusted. This timing configuration information includes the start time, duration, and phase adjustment of the compensation signal. By adjusting this information, the compensation signal can appear at the appropriate time and phase and last for a sufficient duration to better cope with load fluctuations. Finally, based on the adjusted compensation signal timing configuration information, a load-adaptive real-time compensation waveform is generated. This waveform can adapt to load changes in real time, providing timely and effective compensation during load fluctuations, thereby improving the power supply's response to load changes, further suppressing ripple interference, and ensuring the stability of the power supply output.

[0075] For example, step S300 may specifically include the following steps S310 to S350:

[0076] Step S310: Input the virtual compensation ripple signal into the signal preprocessing layer of the load-ripple dynamic correlation model, perform time-series splitting on the virtual compensation ripple signal, and split the virtual compensation ripple signal into multiple continuous signal segments at fixed time intervals, with each signal segment having the same time length.

[0077] The choice of a fixed time interval needs to be determined based on the characteristics of the virtual compensation ripple signal and the needs of the analysis. If the time interval is too short, too many signal segments will be obtained, increasing the complexity of subsequent processing; if the time interval is too long, some important signal features may be lost. During time-series decomposition, signal segments are extracted sequentially at fixed time intervals, starting from the beginning time of the virtual compensation ripple signal. Each signal segment has the same duration, facilitating unified processing and analysis of each segment later. By decomposing the complex virtual compensation ripple signal into multiple signal segments, the characteristics of the ripple signal at different time periods can be observed and analyzed more clearly, facilitating ripple characteristic extraction and correlation matching.

[0078] Step S320: Input the real-time collected load dynamic fluctuation time series information into the time series preprocessing layer of the load-ripple dynamic correlation model, adjust the sampling frequency of the load dynamic fluctuation time series information, and adjust the sampling frequency to be consistent with the sampling frequency of the signal segment by interpolation or decimation, so as to obtain a load fluctuation segment with the same sampling frequency as the signal segment. Each load fluctuation segment corresponds to a signal segment.

[0079] The time-series preprocessing layer of the load-ripple dynamic correlation model is responsible for preprocessing the real-time acquired load dynamic fluctuation time-series information to ensure that it is consistent with the signal segment of the virtual compensation ripple signal in terms of sampling frequency, which facilitates subsequent correlation analysis.

[0080] Since the signal segments obtained after the virtual compensation ripple signal is split into time segments have corresponding sampling frequencies, while the sampling frequency of the real-time acquired load dynamic fluctuation timing information may be different, sampling frequency adjustment is required.

[0081] After adjusting the sampling frequency, a load fluctuation segment with the same sampling frequency as the signal segment is obtained. Each load fluctuation segment corresponds to a signal segment, thus enabling a one-to-one correlation analysis between the signal segment and the load fluctuation segment.

[0082] Step S330: Input the signal segment and the corresponding load fluctuation segment into the correlation matching layer of the load-ripple dynamic correlation model, perform correlation matching between ripple characteristics and load fluctuation, extract the characteristics of the two and calculate the correlation degree, and obtain the correlation matching result corresponding to each signal segment. The correlation matching result reflects the degree of correlation between ripple characteristics and load fluctuation.

[0083] For example, step S330 may specifically include the following steps S331 to S336:

[0084] Step S331: Extract ripple characteristic information from the signal segment. Extract ripple characteristic information by analyzing the waveform shape and changing trend of the signal segment. The ripple characteristic information includes the waveform type and changing pattern of the ripple.

[0085] Analyzing the waveform shape of a signal segment can be achieved by observing its peak values, trough values, period, and other characteristics. For example, if the peak and trough values ​​of a signal are relatively fixed and the period is regular, it may be a sine wave; if the rising and falling edges of the signal are very steep, it may be a square wave. By observing and statistically analyzing a large number of signal segments, the waveform type of the ripple can be determined.

[0086] Trend analysis can be performed by calculating the derivative or difference of signal segments. The derivative reflects the rate of change of the signal, while the difference reflects the changes between adjacent data points. If the derivative or difference is positive, it indicates that the signal is in an upward trend; if it is negative, it indicates that the signal is in a downward trend; if it is close to zero, it indicates that the signal is relatively stable. By analyzing signal segments at different time intervals, the patterns of ripple changes can be summarized, such as whether there are periodic rises and falls, or whether there are sudden jumps.

[0087] Step S332: Extract load fluctuation characteristic information from the corresponding load fluctuation segment. The load fluctuation characteristic information is extracted by analyzing the fluctuation shape and change trend of the load fluctuation segment. The load fluctuation characteristic information includes the waveform type and change pattern of the load fluctuation.

[0088] This step is similar to step S331, except that the object of analysis is changed to a load fluctuation segment. The shape of the fluctuation is analyzed by observing the power change curve or current change curve of the load fluctuation segment. For example, if the curve shows periodic fluctuations, it may be a periodic load fluctuation; if the curve has a sudden jump, it may be a momentary change in the load.

[0089] Trend analysis can also be performed by calculating the derivative or difference of load fluctuation segments. Based on the value of the derivative or difference, the upward, downward, or stable trend of load fluctuation can be determined. By analyzing load fluctuation segments over different time periods, the changing patterns of load fluctuations can be summarized, such as whether the load power gradually increases or decreases, or whether intermittent fluctuations exist.

[0090] Step S333: Input the ripple characteristic information and load fluctuation characteristic information into the feature alignment sub-layer of the association matching layer for dimension alignment. Adjust the dimensions of the two to be consistent by adding zero padding or truncation to obtain aligned ripple characteristic information and aligned load fluctuation characteristic information. The aligned information has the same dimension.

[0091] The feature alignment sublayer of the association matching layer aligns the ripple characteristics and load fluctuation characteristics in terms of dimensions to facilitate subsequent association degree calculation. Since the ripple characteristics and load fluctuation characteristics may have different dimensions, such as different numbers of features or different lengths, they need to be adjusted to ensure dimensional consistency.

[0092] If the dimensionality of the ripple characteristic information is smaller than that of the load fluctuation characteristic information, zero-padding can be used. Zero elements are added after the ripple characteristic information to make its dimension the same as that of the load fluctuation characteristic information. This ensures that the two pieces of information can be effectively compared in subsequent correlation calculations.

[0093] If the dimension of the ripple characteristic information is greater than the dimension of the load fluctuation characteristic information, a truncation method can be used. Only the first part of the ripple characteristic information is retained, making its dimension consistent with that of the load fluctuation characteristic information.

[0094] After dimension alignment, aligned ripple characteristics and aligned load fluctuation characteristics are obtained. Since both have the same dimension, correlation calculations can be performed on the same basis, improving the accuracy and effectiveness of correlation analysis.

[0095] Step S334: Input the aligned ripple characteristic information and the aligned load fluctuation characteristic information into the correlation calculation sub-layer of the correlation matching layer, perform correlation degree calculation at each time step, calculate the waveform similarity between the two at each time step, and obtain the correlation degree value at each time step. The correlation degree value ranges from 0 to 1.

[0096] The correlation calculation sublayer of the correlation matching layer is responsible for calculating the correlation degree of the aligned ripple characteristics and the aligned load fluctuation characteristics at each time step. Waveform similarity calculation is the core of the correlation degree calculation. At each time step, the corresponding parts of the ripple characteristics and load fluctuation characteristics are treated as waveforms, and their similarity is calculated using similarity measurement methods, such as cosine similarity and Euclidean distance.

[0097] Taking cosine similarity as an example, the waveform characteristics and load fluctuation characteristics at each time step are represented as vectors. The dot product of these two vectors is calculated, and then divided by the product of their magnitudes to obtain the cosine similarity value, or correlation value, at that time step. The closer the correlation value is to 1, the more similar the waveforms of the two devices are at that time step, and the higher the correlation. The closer the correlation value is to 0, the less similar the waveforms of the two devices are at that time step, and the lower the correlation.

[0098] Step S335: Input the correlation value at each time step into the sequence integration sub-layer of the correlation matching layer for temporal integration, calculate the average correlation value of all time steps within each signal segment, and obtain the correlation sequence corresponding to each signal segment. The length of the correlation sequence is consistent with the time length of the signal segment.

[0099] The task of the sequence integration sublayer of the association matching layer is to perform temporal integration of the association degree values ​​at each time step. Since each signal segment contains association degree values ​​at multiple time steps, these values ​​need to be integrated to obtain an index that can represent the degree of association of the entire signal segment.

[0100] Calculating the average correlation value across all time points within each signal segment is a common time series integration method. The average correlation value for each signal segment is obtained by summing the correlation values ​​and then dividing by the number of time points. Arranging the average correlation values ​​of each signal segment sequentially yields a correlation sequence for each segment. The length of the correlation sequence is the same as the time length of the signal segment, as each segment corresponds to one average correlation value. The correlation sequence more comprehensively reflects the correlation between ripple characteristics and load fluctuations throughout the entire signal segment, providing a more accurate basis for adjusting the timing configuration information of the compensation signal.

[0101] Step S336: The correlation degree sequence is used as the correlation matching result for each signal segment and output to the configuration adjustment layer of the load-ripple dynamic correlation model. The configuration adjustment layer will adjust the timing configuration information of the compensation signal based on the correlation degree sequence.

[0102] The correlation sequence reflects the degree of correlation between ripple characteristics and load fluctuation characteristics. Using it as the correlation matching result for each signal segment can accurately reflect the correlation between ripple and load fluctuation in each signal segment.

[0103] After receiving the correlation degree sequence, the configuration adjustment layer of the load-ripple dynamic correlation model will adjust the timing configuration information of the compensation signal based on the sequence. The higher the correlation degree value, the closer the correlation between the ripple characteristics and the load fluctuation characteristics. At this time, it is necessary to adjust the timing configuration information such as the start time, duration, and phase of the compensation signal more accurately so that the compensation signal can better match the load fluctuation.

[0104] For example, for signal segments with high correlation values, the start time of the compensation signal can be advanced so that it can function at the onset of load fluctuations; the duration can be extended to ensure sufficient compensation signal throughout the entire load fluctuation process; and the phase of the compensation signal can be adjusted according to the phase change of the load to achieve better synchronization. In this way, the adaptability of the compensation signal can be improved, power supply ripple interference can be suppressed more effectively, and the power supply's response to load changes can be enhanced.

[0105] Step S340: Input the correlation matching results corresponding to all signal segments into the configuration adjustment layer of the load-ripple dynamic correlation model, adjust the compensation signal timing configuration information of the corresponding signal segment according to each correlation matching result, adjust the start time and duration of each signal segment, and obtain the signal segment after configuration adjustment.

[0106] For example, step S340 may specifically include the following steps S341 to S344:

[0107] Step S341: Input the correlation sequence corresponding to each signal segment into the weight allocation sub-layer of the configuration adjustment layer, and allocate adjustment weights according to the numerical distribution of the correlation sequence. The higher the correlation value, the greater the weight is allocated. The adjustment weights corresponding to each signal segment are obtained, and the length of the adjustment weights is consistent with the length of the correlation sequence.

[0108] The weight allocation sublayer of the configuration adjustment layer is responsible for assigning adjustment weights to each signal segment based on the numerical distribution of the correlation sequence. The correlation sequence reflects the degree of correlation between ripple characteristics and load fluctuation characteristics at each time step within the signal segment.

[0109] When allocating adjustment weights, the weights are determined based on the correlation value. A higher correlation value indicates a stronger correlation between the ripple and load fluctuations at that moment, making it more important in adjusting the timing configuration of the compensation signal, and therefore assigning it a larger weight. This approach highlights the importance of adjusting the compensation signal timing at moments when the ripple and load are closely correlated.

[0110] To ensure that the length of the adjustment weights matches the length of the correlation sequence, each correlation value corresponds to an adjustment weight. This guarantees that precise adjustments can be made at each moment within the signal segment during subsequent configuration adjustments. Through weight allocation, moments with high correlation have a greater impact during the adjustment process, improving the targeting and effectiveness of the adjustments.

[0111] Step S342: Input the adjustment weight corresponding to each signal segment into the configuration adjustment sub-layer of the configuration adjustment layer, and adjust the compensation signal timing configuration information of the corresponding signal segment in combination with the adjustment weight. Adjust the signal output time at each moment to obtain the configured and adjusted signal segment. The timing of the configured and adjusted signal segment is more in line with the load fluctuation.

[0112] After receiving the adjustment weights corresponding to each signal segment, the configuration adjustment sublayer of the configuration adjustment layer adjusts the timing configuration information of the compensation signal for the corresponding signal segment based on these weights. The key to the adjustment lies in the precise control of the signal output time at each moment.

[0113] For moments with significant adjustment weights, the correlation between ripple and load fluctuations is strong, requiring more substantial adjustments to the signal output time. For example, a large adjustment weight indicates a significant impact of load fluctuations on ripple at that moment; in this case, the signal output time can be advanced or delayed based on the load fluctuation. If the load suddenly increases, the signal output at that moment can be advanced, allowing the compensation signal to take effect earlier.

[0114] For moments with smaller adjustment weights, it indicates that the correlation between ripple and load fluctuations is weaker at that moment, and the signal output time can be adjusted by a smaller margin, or even not adjusted at all.

[0115] Step S343: Input all the signal segments after configuration adjustment into the timing sorting sublayer of the configuration adjustment layer, sort them according to the original timing order, and ensure that the order of the signal segments is consistent with the order of the original virtual compensation ripple signal to obtain the sorted signal segment sequence.

[0116] The timing sorting sublayer of the configuration adjustment layer sorts all signal segments after configuration adjustment to ensure that the order of the signal segments is consistent with the order of the original virtual compensation ripple signals. During the adjustment of the timing configuration information of the compensation signals for the signal segments, the order of the signal segments may become disordered. To ensure the continuity and accuracy of the generated compensation waveform in timing, a sorting operation is required.

[0117] During the sorting process, the signal segments are sorted according to their original timing information within the original virtual compensated ripple signal. Each signal segment has its corresponding time sequence, and by comparing this timing information, the signal segments can be arranged in the correct order.

[0118] Step S344: Input the sorted signal segment sequence into the waveform generation sublayer of the configuration adjustment layer to generate a continuous waveform. Fill the gaps between the signal segments with linear interpolation to obtain a continuous load-adaptive real-time compensation waveform.

[0119] The waveform generation sublayer of the configuration adjustment layer is responsible for generating a continuous waveform from the sorted sequence of signal segments. Since the signal segments are discrete, gaps may exist between them, causing discontinuities in the generated waveform and affecting the compensation effect. Therefore, linear interpolation can be used to fill these gaps, making the waveform continuous.

[0120] For two adjacent signal segments, new data points are inserted at the gap based on their endpoint values ​​and time interval. Specifically, the difference between the two endpoint values ​​is calculated, and then the interpolation value at that time point is calculated linearly based on the distance between the time point within the gap and the endpoint. In this way, a series of smoothly transitioning data points are inserted between signal segments, making the waveform continuous. After filling the gaps between signal segments through linear interpolation, a continuous load-adaptive real-time compensation waveform is obtained. This waveform can adapt to load changes in real time, has better continuity and stability, and can provide effective compensation in a timely manner when the load fluctuates, more accurately suppressing power supply ripple interference and improving the quality of power output.

[0121] Step S350: Input all the configured signal segments into the waveform reconstruction layer of the load-ripple dynamic correlation model and perform timing splicing. Splice all the configured signal segments sequentially according to the original timing sequence to obtain the load-adaptive real-time compensation waveform. The timing of the load-adaptive real-time compensation waveform matches the timing of the load fluctuation.

[0122] The waveform reconstruction layer of the load-ripple dynamic correlation model is responsible for timing-sequentially stitching together all signal segments after configuration adjustments. After a series of preceding processes, the timing configuration information of the compensation signal for each signal segment has been adjusted, and the signal segments have been sorted. The role of the waveform reconstruction layer is to connect these signal segments sequentially according to their original timing order, forming a complete load-adaptive real-time compensation waveform.

[0123] Timing stitching ensures that the timing of the generated compensation waveform matches the timing of load fluctuations. When load fluctuations occur, the compensation waveform responds promptly, providing the appropriate compensation signal at the right time. Through timing stitching operations at the waveform reconstruction layer, the resulting load-adaptive real-time compensation waveform accurately tracks load fluctuations. When load demand increases, the compensation waveform increases output and suppresses ripple interference; when load demand decreases, the compensation waveform correspondingly reduces output, lowering energy consumption. Therefore, this waveform effectively improves power supply stability and reliability, providing a more stable power supply for computer systems.

[0124] Step S400: Obtain power output energy efficiency timing feedback information and ripple residual monitoring information, and input them together with the load-adaptive real-time compensation waveform into the ripple adversarial generation network. Based on the dual-objective optimization criterion, perform hierarchical iterative optimization of the weights inside the network to obtain the weight-optimized ripple adversarial generation network.

[0125] For example, step S400 may specifically include the following steps S410-S470:

[0126] Step S410: Obtain power output energy efficiency timing feedback information. Collect energy efficiency information through the power supply's built-in energy efficiency monitoring device to obtain power output energy efficiency timing feedback information, which includes energy efficiency data at different times.

[0127] The power supply's built-in energy efficiency monitoring device consists of a power sensor and a data acquisition module. The power sensor measures the power supply's input and output power in real time. Input power reflects the energy the power supply draws from the external environment, while output power reflects the energy the power supply actually provides to the load. By dividing the output power by the input power, the power supply's energy efficiency value at a given moment can be obtained. The data acquisition module collects and records the input and output power data measured by the power sensor at set time intervals. Over time, these collected energy efficiency values ​​form the power supply's output energy efficiency time-series feedback information.

[0128] Step S420: Synchronize the power supply output energy efficiency timing feedback information. Align the time of the energy efficiency timing feedback information with the time of the load-adaptive real-time compensation waveform by timestamp alignment to obtain energy efficiency feedback information with the timing of the load-adaptive real-time compensation waveform. The timestamp of the energy efficiency feedback information corresponds one-to-one with the timestamp of the compensation waveform.

[0129] A timestamp is a time identifier recorded for each data point, which can accurately represent the moment of data acquisition. Since the acquisition times of power supply output energy efficiency timing feedback information and load-adaptive real-time compensation waveforms may differ, timing synchronization is required to accurately analyze the relationship between the two.

[0130] When performing timestamp alignment, first determine the timestamp sequence of the load-adaptive real-time compensation waveform. Then, for each data point in the power output energy efficiency timing feedback information, find the data point of the load-adaptive real-time compensation waveform that is closest to it in time. Unify the timestamps of these two data points to make them correspond in time.

[0131] If there are large time intervals and corresponding data points cannot be directly found, interpolation or extrapolation methods can be used. Interpolation estimates the energy efficiency value at a given time point based on data from adjacent time points; extrapolation predicts the energy efficiency value at a given time point based on the trend of existing data. By aligning the timestamps, energy efficiency feedback information consistent with the timing of the load-adaptive real-time compensation waveform is obtained. Based on this, in subsequent processing, the energy efficiency changes of the power supply under different compensation waveforms can be accurately analyzed, providing a more accurate basis for optimizing network weights.

[0132] Step S430: Obtain residual ripple monitoring information. Collect residual ripple information at the power output terminal through a ripple monitoring device to obtain residual ripple monitoring information, which includes residual ripple data at different times.

[0133] Ripple monitoring devices can employ high-precision oscilloscopes or dedicated ripple detection circuits. Oscilloscopes, with their high resolution and high sampling rate, can capture the ripple signal waveform at the power supply output in real time and accurately measure parameters such as ripple amplitude and frequency. Dedicated ripple detection circuits can be integrated into the power supply system for real-time ripple monitoring.

[0134] The ripple monitoring device collects residual ripple information at the power output terminal at set time intervals. Each time it collects data, it records key information such as the ripple amplitude and frequency components. Over time, this collected data forms the residual ripple monitoring information.

[0135] The residual ripple data at different times in the residual ripple monitoring information can reflect the actual situation of the power supply output ripple after compensation at different times.

[0136] Step S440: Filter the residual ripple monitoring information to obtain filtered residual ripple monitoring information, which is smoother.

[0137] Various filtering methods can be employed, such as low-pass filtering and median filtering. When performing filtering, the appropriate filtering method and parameters should be selected based on the characteristics of the ripple residue monitoring information and actual needs. After filtering, the resulting ripple residue monitoring information is smoother and can more accurately reflect the true situation of ripple residue.

[0138] Step S450: Input the energy efficiency feedback information, the filtered ripple residue monitoring information, and the load-adaptive real-time compensation waveform into the optimized input layer of the ripple adversarial generator network. After normalization, align them according to the timestamp and splice them along the feature dimension to obtain the collaborative input information, which includes information on energy efficiency, ripple residue, and compensation waveform.

[0139] Normalization transforms data of different ranges and scales into a unified range, thus avoiding the adverse effects on network training caused by excessive scale differences between data. Appropriate normalization methods are employed for energy efficiency feedback information, filtered ripple residue monitoring information, and load-adaptive real-time compensation waveforms.

[0140] For energy efficiency feedback information, its values ​​can be scaled to between 0 and 1. First, determine the maximum and minimum values ​​in the energy efficiency feedback information. Then, subtract the minimum value from each energy efficiency value and divide by the difference between the maximum and minimum values ​​to obtain the normalized energy efficiency value. A similar method is used to normalize the filtered ripple residue monitoring information. By scaling the amplitude and other data of the ripple residue, it is brought to a suitable range.

[0141] For the load-adaptive real-time compensation waveform, normalization is also performed to ensure its data range is consistent with other information. After normalization, these three pieces of information are aligned according to their timestamps. Since the energy efficiency feedback information and the load-adaptive real-time compensation waveform have already been time-stamp aligned, the filtered ripple residue monitoring information is also matched according to the same timestamp to ensure they correspond in time. Finally, the normalized and timestamp-aligned energy efficiency feedback information, the filtered ripple residue monitoring information, and the load-adaptive real-time compensation waveform are concatenated along the feature dimension. Concatenation combines these three pieces of information along the feature dimension to form a new information vector. This vector contains information on energy efficiency, ripple residue, and the compensation waveform, and is used as collaborative input information to the optimization input layer of the ripple adversarial generative network, providing comprehensive information for dual-objective optimization.

[0142] Step S460: Input the collaborative input information into the weight optimization layer of the ripple adversarial generative network, and call the bi-objective optimization criterion to perform hierarchical iterative optimization of the internal weights of the network. First, optimize the weights of the feature encoding branch, and then optimize the weights of the compensation signal generation branch to obtain the iteratively optimized internal weights of the network.

[0143] For example, step S460 may specifically include the following steps S461-S465:

[0144] Step S461: Input the collaborative input information into the target decomposition sublayer of the weight optimization layer, and decompose it into two optimization targets based on the dual-objective optimization criterion, corresponding to the energy efficiency optimization target and the ripple residual optimization target, respectively.

[0145] During the objective decomposition process, based on the requirements of the dual-objective optimization criterion, the energy efficiency-related and ripple residue-related components of the collaborative input information are analyzed. For the energy efficiency optimization objective, the focus is on the impact of power supply output energy efficiency timing feedback information and compensation waveforms on energy efficiency. By analyzing this information, it is determined how to adjust the network weights to improve the power supply's output energy efficiency under the current conditions, such as reducing losses and improving energy conversion efficiency during energy conversion.

[0146] For ripple residue optimization, the main focus is on ripple residue monitoring information and the ripple suppression effect of the compensation waveform. It involves analyzing whether the compensation signal generated under the current weights can effectively reduce ripple residue. If not, the problem needs to be identified, and the weights need to be adjusted to enhance ripple suppression.

[0147] Step S462: Input the energy efficiency optimization objective into the first optimization sub-layer of the weight optimization layer, iteratively optimize the weights related to energy efficiency within the network, adjust the weights related to energy efficiency information in the feature encoding branch, and obtain the first part of optimized weights.

[0148] The role of the feature encoding branch is to extract key features from the collaborative input information, among which features related to energy efficiency are crucial for improving power supply output energy efficiency. During iterative optimization, the quality of energy efficiency-related features extracted by the feature encoding branch under the current weights is continuously evaluated.

[0149] If the extracted features fail to accurately reflect energy efficiency information or are detrimental to the subsequent generation of compensation signals that improve energy efficiency, the weights related to energy efficiency information are adjusted. The adjustment process involves gradually changing the weight values ​​based on feedback from the energy efficiency optimization objective, enabling the feature encoding branch to extract features that are more conducive to improving energy efficiency.

[0150] In each iteration, by analyzing the power supply output efficiency timing feedback information, it is determined whether the adjusted weights have improved the power supply's output efficiency. If the efficiency has improved, it indicates that the current weight adjustment direction is correct, and fine-tuning can continue in this direction; if the efficiency has not improved or has even decreased, the weight adjustment direction needs to be readjusted.

[0151] Step S463: Input the ripple residue optimization target into the second optimization sub-layer of the weight optimization layer, iteratively optimize the weights related to ripple residue inside the network, adjust the weights related to ripple residue in the compensation signal generation branch, and obtain the second part of optimized weights.

[0152] The second optimization sublayer of the weight optimization layer focuses on optimizing the weights related to ripple residue within the network. When the ripple residue optimization objective is input into this sublayer, it mainly adjusts the weights related to ripple residue in the compensation signal generation branch.

[0153] The compensation signal generation branch generates a compensation signal based on the features extracted by the feature encoding branch. The weights related to ripple residue directly affect the ripple suppression effect of the generated compensation signal. During iterative optimization, the ability of the compensation signal generated by the compensation signal generation branch to suppress ripple residue under the current weights is continuously evaluated.

[0154] If the generated compensation signal cannot effectively reduce ripple residue, the weights related to ripple residue need to be adjusted. Based on ripple residue monitoring information, analyze the deficiencies of the current compensation signal, such as whether the amplitude and phase of the compensation signal are appropriate. Then, based on the analysis results, adjust the weights accordingly so that the compensation signal generation branch can generate a more effective compensation signal.

[0155] In each iteration, by comparing the ripple residue monitoring information before and after the weight adjustment, it is determined whether the adjusted weight has reduced the ripple residue. If the ripple residue has decreased, it indicates that the current weight adjustment direction is correct, and fine-tuning can continue in this direction; if the ripple residue has not decreased or has even increased, the weight adjustment direction needs to be readjusted.

[0156] After multiple iterations and optimizations, the second set of optimized weights was finally obtained. These weights enable the compensation signal generation branch to generate a more effective compensation signal, thereby better suppressing the residual ripple at the power supply output.

[0157] Step S464: Input the first part of the optimized weights and the second part of the optimized weights into the weight integration sub-layer of the weight optimization layer, perform weight integration, and add the two parts of weights according to the set ratio to obtain the integrated internal weights of the network.

[0158] When integrating the weights, first determine the ratio of the two weights. This ratio needs to consider the importance of both power supply output efficiency and residual ripple. If more emphasis is placed on improving power supply output efficiency, the weight of the first part of the optimization can be appropriately increased; if more emphasis is placed on reducing residual ripple, the weight of the second part of the optimization can be increased.

[0159] After determining the ratio, the first part of the optimization weights and the second part of the optimization weights are added together according to this ratio. For each weight position, the value of the corresponding position of the first part of the optimization weights is multiplied by its set ratio, and the value of the corresponding position of the second part of the optimization weights is multiplied by its set ratio. Then the two results are added together to obtain the integrated weight value for that position.

[0160] During the addition process, it is essential to ensure that the dimensions and positions of the two weights correspond. If there is a dimension mismatch, appropriate handling is required, such as dimension alignment or interpolation, to ensure that the two weights can be added correctly.

[0161] By integrating the weights in the sub-layer, the integrated weights of the network are obtained. These weights comprehensively consider both the goals of improving power output efficiency and reducing ripple residue, providing more optimized parameters for improving the performance of ripple adversarial generator networks.

[0162] For example, step S464 may include the following steps S4641-S4646:

[0163] Step S4641: Extract the weight distribution information related to energy efficiency from the first part of the optimization weights, analyze the scope of each weight in the first part of the optimization weights, extract the weight distribution related to energy efficiency optimization, and obtain the weight distribution information related to energy efficiency.

[0164] When extracting energy efficiency-related weight distribution information from the first part of the optimized weights, we first need to clarify the role of each weight in the feature encoding branch. The first part of the optimized weights is obtained during the process of optimizing energy efficiency targets, where different weights play different roles in extracting energy efficiency-related features.

[0165] Analyzing the scope of each weight requires considering the structure and function of the feature encoding branch. A feature encoding branch typically consists of multiple layers and modules, each with its own weights. For example, some weights might be used to extract features related to the power input from the input information, which significantly impact energy efficiency calculations; others might be used to extract features related to the output power.

[0166] By analyzing the role of each weight in the first part of the optimization weights, we identified those weights that have a direct or indirect impact on energy efficiency optimization. These energy efficiency-related weights were then organized according to their position and role in the feature encoding branches to obtain the weight distribution information related to energy efficiency.

[0167] Step S4642: Extract the weight distribution information related to ripple residue from the second part of the optimization weights, analyze the scope of each weight in the second part of the optimization weights, extract the weight distribution related to ripple residue optimization, and obtain the weight distribution information related to ripple residue.

[0168] For the second part of the optimization weights, it is also necessary to analyze the scope of influence of each weight in the compensation signal generation branch. The second part of the optimization weights is obtained in the process of optimizing the ripple residue target, and different weights have different effects on generating the compensation signal that suppresses ripple.

[0169] The compensation signal generation branch generates a compensation signal based on the features extracted by the feature encoding branch. The weights related to ripple residue determine the amplitude, phase, and other parameters of the compensation signal, which directly affect the ripple suppression effect.

[0170] When analyzing the scope of each weight, the structure and function of the compensation signal generation branch should be considered. For example, some weights may be used to adjust the amplitude of the compensation signal to match the amplitude of the ripple signal; other weights may be used to adjust the phase of the compensation signal to be opposite to the phase of the ripple signal, thereby achieving a better cancellation effect.

[0171] By conducting an in-depth analysis of the role of each weight in the second part of the optimization weights, we identified those weights that play a crucial role in ripple residue optimization. These weights related to ripple residue optimization are then organized according to their position and function in the compensation signal generation branch, yielding the weight distribution information related to ripple residue.

[0172] Step S4643: Input the weight distribution information related to energy efficiency and the weight distribution information related to ripple residue into the distribution alignment sublayer of the weight integration sublayer, perform distribution alignment, and adjust the weight positions of the two to be consistent through coordinate transformation, so as to obtain the aligned energy efficiency weight distribution and the aligned ripple residue weight distribution, and the aligned weight distribution positions are the same.

[0173] The distribution alignment sublayer of the weight integration sublayer aligns the weight distribution information related to energy efficiency with that related to ripple residual. Since these two weights originate from the feature encoding branch and the compensation signal generation branch, respectively, their positions may be inconsistent.

[0174] To accurately integrate the two weights, distribution alignment is required. Coordinate transformation is the key method for achieving distribution alignment. Through coordinate transformation, the positions of the weight distribution information related to energy efficiency and the weight distribution information related to ripple residue are adjusted so that they are aligned in the new coordinate system.

[0175] When performing coordinate transformation, a unified coordinate system is first determined. This coordinate system can be determined based on the overall structure and weight distribution of the ripple adversarial generator network. Then, the weight distribution information related to energy efficiency and the weight distribution information related to ripple residue are mapped to this unified coordinate system respectively.

[0176] By adjusting the positions of the weights in the coordinate system, they can be accurately compared and added at their corresponding positions. After distribution alignment, the aligned energy efficiency weight distribution and the aligned ripple residual weight distribution are obtained.

[0177] Step S4644: Input the aligned energy efficiency weight distribution and the aligned ripple residual weight distribution into the weight fusion sublayer of the weight integration sublayer, perform weight-by-weight fusion, add the energy efficiency weight and ripple residual weight at each position to obtain the fused weight distribution, which contains the optimization information of both.

[0178] The weight fusion sublayer of the weight integration sublayer is responsible for fusing the aligned energy efficiency weight distribution and the aligned ripple residual weight distribution. Since these two weight distributions have already been aligned in the distribution alignment sublayer, weight-by-weight fusion can be performed.

[0179] During the weighted fusion process, for each position, the aligned energy efficiency weight and the aligned ripple residue weight are added together. This is done to integrate information from energy efficiency optimization and ripple residue optimization. The energy efficiency weight reflects optimization for improving power supply output efficiency, while the ripple residue weight reflects optimization for reducing ripple residue; adding them together yields a weight value that comprehensively considers both objectives.

[0180] By adding the weights one by one, the fused weight distribution is obtained. This weight distribution includes both energy efficiency optimization information and ripple residue optimization information, enabling the ripple adversarial generator network to simultaneously achieve the goals of improving power output energy efficiency and reducing ripple residue in subsequent operation, thus realizing a more comprehensive performance improvement.

[0181] Step S4645: Input the merged weight distribution into the weight verification sublayer of the weight integration sublayer for rationality verification. Check whether the value of each weight is within a reasonable range, remove unreasonable weight values ​​and make corrections to obtain the verified weight distribution. The verified weight distribution is more reasonable.

[0182] The weight verification sublayer of the weight integration sublayer is responsible for verifying the rationality of the merged weight distribution. During the weight fusion process, some weight values ​​may become unreasonable, such as being too large or too small, which may lead to network instability or performance degradation.

[0183] When performing a rationality check, first determine the reasonable range for each weight. This reasonable range can be determined based on the structure of the ripple adversarial generative network, training experience, and practical application requirements. For example, the values ​​of certain weights should be within a set range; exceeding this range may negatively impact the network's performance.

[0184] For each weight in the merged weight distribution, check if its value is within a reasonable range. If a weight's value exceeds the reasonable range, it is considered an unreasonable weight value. Unreasonable weight values ​​need to be corrected. Correction methods include adjusting it to the boundary values ​​of the reasonable range, or interpolating based on the values ​​of surrounding reasonable weights to obtain a more reasonable weight value.

[0185] After rationality verification and correction, the verified weight distribution is obtained. Each weight value in this weight distribution is within a reasonable range, which can ensure that the ripple adversarial generative network is more stable and reliable in subsequent operation, thus improving its overall performance.

[0186] Step S4646: The verified weight distribution is used as the integrated internal weight of the network and output to the weight update layer of the ripple adversarial generative network. The weight update layer updates the internal weight of the network based on this weight.

[0187] After obtaining the verified weight distribution, it is used as the internal weight of the integrated network. This part of the weight is obtained after optimizing, aligning, fusing, and verifying the rationality of the two objectives of energy efficiency and ripple residue, taking into account the needs of improving power supply output energy efficiency and reducing ripple residue.

[0188] The weight update layer of the ripple adversarial generator network (PAG) is responsible for updating the actual weights within the network based on the integrated weights. During the update process, each weight value in the verified weight distribution is accurately replaced with the corresponding weight within the network. The weight update layer ensures the accuracy and stability of the update process, preventing weight loss or erroneous updates. By updating the weights within the network, the PAG can operate with more optimized parameters, better achieving the dual objectives of improving power supply output efficiency and reducing ripple residue, thus enhancing the overall performance and stability of the power system.

[0189] Step S465: The integrated network internal weights are used as the iteratively optimized network internal weights and output to the weight update layer of the ripple adversarial generative network. The weight update layer updates the network based on these weights.

[0190] After a series of optimization steps, the integrated network internal weights are obtained. These weights are derived by performing hierarchical iterative optimization on the weights of the feature encoding branch and the compensation signal generation branch under a bi-objective optimization criterion, followed by weight integration and rationality verification. They integrate optimization information from two important objectives: improving power output efficiency and reducing ripple residue. The integrated network internal weights are then output as the iteratively optimized network internal weights to the weight update layer of the ripple adversarial generator network. The main task of the weight update layer is to update the actual weights within the network based on these optimized weights.

[0191] During the update process, the weight update layer replaces each value in the iteratively optimized weights within the network with the corresponding weight within the network. To ensure the accuracy and stability of the update, the weight update layer performs rigorous checks and verifications to prevent weight update errors or loss. By updating the weights within the network, the ripple adversarial generative network can operate with more optimized parameters. This makes the network more efficient and accurate in handling power output efficiency and ripple residue issues, enabling it to generate compensation signals that better meet the dual objectives, thereby further improving the performance of the power system and achieving high efficiency and stability in power output.

[0192] Step S470: Update the iteratively optimized network internal weights to the ripple adversarial generative network, replacing the original network internal weights, to obtain the weight-optimized ripple adversarial generative network.

[0193] The original network internal weights were parameters before the dual-objective optimization, which may not be able to effectively balance the two objectives of improving power output efficiency and reducing ripple residue. The iteratively optimized network internal weights, obtained through multiple rounds of optimization and adjustment, comprehensively consider the influencing factors of power output efficiency and ripple residue, enabling the network to better adapt to the needs of practical applications.

[0194] During the update process, to ensure the accurate replacement of the original weights with new ones, operations can be performed on the network's memory or storage to write the new weight data to the appropriate locations. Simultaneously, to guarantee network stability, the dimensions and format of the new weights can be checked to ensure they match the network structure. After the replacement is complete, a ripple adversarial generative network with optimized weights is obtained. Under the new weight parameters, this network can more effectively extract key features of the input information and generate compensation signals that better meet the dual-objective requirements. In terms of improving power supply output efficiency, it can better utilize input energy and reduce energy loss; in terms of reducing ripple residue, it can more accurately suppress ripple signals and improve the stability of the power supply output.

[0195] Step S500: Based on the weight-optimized ripple adversarial generation network, a ripple suppression adaptive adjustment link is constructed by linking the real-time monitoring link of the power supply output ripple. Feedback calibration is continuously initiated through ripple residual monitoring information, and the phase, amplitude, and timing configuration information of the load-adaptive real-time compensation waveform are iteratively adjusted.

[0196] For example, step S500 may specifically include the following steps S510-S550:

[0197] Step S510: Link the weighted ripple adversarial generation network with the power output ripple real-time monitoring link to establish a bidirectional data transmission channel. The data transmission channel is used to transmit ripple residual monitoring information and compensation waveform information.

[0198] When establishing a data transmission channel, the stability and real-time performance of data transmission must be considered. Wired or wireless communication methods can be used, selected based on the actual application scenario and requirements. For example, in situations requiring high data transmission stability, wired Ethernet or fiber optic communication can be used; while in scenarios requiring flexible deployment, wireless Wi-Fi or Bluetooth communication can be used. One end of the data transmission channel connects to a real-time power output ripple monitoring link, which is responsible for collecting residual ripple monitoring information at the power output in real time. The collected information is transmitted through the data transmission channel to a weighted, optimized ripple adversarial generation network. The network analyzes and processes this information to generate corresponding compensation waveform information.

[0199] The compensation waveform information is then transmitted back to the power output via the data transmission channel to suppress ripple. This bidirectional data transmission enables real-time interaction between ripple residual monitoring and compensation signal generation, providing data support for feedback calibration and iterative adjustment.

[0200] Step S520: Obtain residual ripple monitoring information from the power output ripple real-time monitoring link through the data transmission channel, and periodically receive the latest residual ripple data from the monitoring link to obtain residual ripple monitoring information, which includes the current residual ripple status.

[0201] Obtaining residual ripple monitoring information from the power output ripple real-time monitoring link via the established data transmission channel is a crucial step in feedback calibration. The power output ripple real-time monitoring link continuously monitors the ripple condition at the power output and transmits the monitored data through the data transmission channel.

[0202] To monitor the current ripple residual status in a timely manner, it is necessary to periodically receive the latest ripple residual data from the monitoring link. The interval between these periodic receptions can be set according to actual needs, such as receiving data once per second, every ten seconds, or every minute. Each received ripple residual data set contains key information such as the amplitude, frequency, and phase of the current power supply output ripple. This information reflects the ripple suppression effect of the current compensation waveform. If the amplitude of the ripple residual is still large, it indicates that the current compensation waveform may not be effective enough; if the frequency or phase of the ripple changes, the parameters of the compensation waveform also need to be adjusted accordingly.

[0203] Step S530: Input the residual ripple monitoring information into the feedback layer of the weighted ripple adversarial generator network, start feedback calibration, analyze the residual ripple information and generate calibration instructions, and obtain calibration instructions, which contain adjustment information for phase, amplitude and timing.

[0204] For example, step S530 may specifically include the following steps S531 to S535:

[0205] Step S531: Input the ripple residue monitoring information into the information parsing sublayer of the feedback layer, perform information parsing, extract the phase, amplitude and timing information of the ripple residue, and obtain the parsed ripple residue information, which is more detailed.

[0206] Ripple residual monitoring information typically contains a large amount of data, but this data may be raw and unprocessed, and key phase, amplitude, and timing information needs to be extracted from it.

[0207] During the analysis process, the information analysis sublayer first processes the amplitude data in the ripple residue monitoring information. By analyzing the magnitude and trend of the amplitude, the current amplitude of the ripple residue is determined. For phase information, signal processing techniques, such as phase detection algorithms, are used to extract the phase value of the ripple from the ripple residue monitoring information.

[0208] When extracting time-series information, the occurrence patterns and durations of ripples at different times are analyzed based on the timestamps and data changes in the ripple residue monitoring information. In this way, time-series information such as the start time, end time, and duration of ripple residue are obtained.

[0209] Step S532: Input the parsed residual ripple information into the calibration judgment sublayer of the feedback layer to determine the necessity of calibration. Compare the parsed residual ripple information with the preset threshold to determine whether calibration is required and obtain the calibration necessity result.

[0210] The preset thresholds are set according to the performance requirements of the power system and the actual application scenario. For the amplitude of residual ripple, a maximum allowable value is set; for phase deviation, an allowable range is also set; and for timing, there are corresponding time error thresholds.

[0211] During the comparison, the calibration judgment sublayer compares the amplitude, phase, and timing information of the residual ripple with the corresponding thresholds. If the amplitude of the residual ripple exceeds the maximum allowable value, or the phase deviation exceeds the allowable range, or the timing error exceeds the set threshold, calibration is considered necessary.

[0212] If the analyzed residual ripple information is within a preset threshold range, calibration is deemed unnecessary. This calibration necessity assessment avoids unnecessary calibration operations, improving system efficiency and stability. The final calibration necessity result will determine whether to generate a calibration command to adjust the compensation waveform.

[0213] Step S533: If the calibration necessity result indicates that calibration is required, input the resolved ripple residual information into the instruction generation sublayer of the feedback layer to generate a calibration instruction. The calibration instruction includes the specific adjustment direction and adjustment amount of phase, amplitude and timing.

[0214] When the calibration necessity result indicates that calibration is required, the parsed ripple residue information is input into the instruction generation sublayer of the feedback layer. The task of the instruction generation sublayer is to generate specific calibration instructions based on the parsed ripple residue information.

[0215] When generating calibration instructions, for phase adjustment, the direction and amount of phase adjustment are determined based on the difference between the phase of the residual ripple and the phase of the current compensation waveform. If the phase of the residual ripple differs significantly from the phase of the compensation waveform, the phase of the compensation waveform needs to be rotated accordingly to make their phases more opposite, thus enhancing the cancellation effect. The magnitude of the adjustment is determined by the degree of phase difference; the greater the difference, the greater the adjustment.

[0216] For amplitude adjustment, analyze the relationship between the residual ripple amplitude and the current compensation waveform amplitude. If the residual ripple amplitude is large, it indicates that the current compensation waveform amplitude is insufficient, and the amplitude of the compensation waveform needs to be increased; conversely, if the residual ripple amplitude is small, the amplitude of the compensation waveform can be appropriately decreased. The adjustment amount is determined based on the magnitude of the amplitude difference.

[0217] Regarding timing adjustments, the start time and duration of the compensation waveform are determined based on the time characteristics of ripple residue and the timing of the current compensation waveform. If the ripple residue is significant within a certain time period, the start time of the compensation waveform can be advanced or its duration extended. The final generated calibration command explicitly includes the specific adjustment direction and amount for phase, amplitude, and timing.

[0218] Step S534: If the calibration necessity result is that calibration is not required, generate a no-calibration instruction. A no-calibration instruction means that the compensation waveform does not need to be adjusted.

[0219] When the calibration necessity result indicates that calibration is not required, the instruction generation sublayer of the feedback layer will generate a no-calibration instruction. This instruction indicates that the current load-adaptive real-time compensation waveform is effective in suppressing ripple and does not require adjustment of its phase, amplitude, and timing configuration information. Generating a no-calibration instruction helps avoid unnecessary adjustment operations, reducing the system's computational burden and energy consumption. It also ensures system stability, preventing instability of the compensation waveform due to frequent adjustments. After receiving the no-calibration instruction, the system will maintain the current compensation waveform parameters and continue monitoring the ripple at the power supply output. Only when subsequent residual ripple monitoring information indicates that calibration is required will the calibration process be restarted, generating the corresponding calibration instruction.

[0220] Step S535: The calibration command or no calibration command is taken as the result of the feedback calibration process and output to the configuration adjustment module of the ripple suppression adaptive adjustment link. The configuration adjustment module will adjust or maintain the load-adaptive real-time compensation waveform according to the command.

[0221] Whether a calibration command is received or not, the result of the feedback calibration process will be output to the configuration adjustment module of the ripple suppression adaptive adjustment link. The main responsibility of the configuration adjustment module is to perform corresponding operations on the load-adaptive real-time compensation waveform based on the received command. If a calibration command is received, the configuration adjustment module will precisely adjust the load-adaptive real-time compensation waveform according to the specific adjustment direction and amount of phase, amplitude, and timing contained in the command. When adjusting the phase, the phase of the compensation waveform is rotated using a corresponding signal processing algorithm; when adjusting the amplitude, the amplitude value of the compensation waveform is changed; and when adjusting the timing, the start time and duration of the compensation waveform are modified.

[0222] If a calibration-free command is received, the configuration adjustment module will maintain the current load-adaptive real-time compensation waveform parameters without making any adjustments. This ensures stable system operation when no adjustments are needed, improving system efficiency and reliability.

[0223] By configuring and adjusting the module, the load-adaptive real-time compensation waveform can be adjusted in a timely manner based on the feedback calibration results, so that it can better adapt to changes in power output and dynamic fluctuations in load, and continuously improve the ripple suppression effect.

[0224] Step S540: Input the calibration command into the configuration adjustment module of the ripple suppression adaptive adjustment link, and adjust the phase, amplitude and timing configuration information of the load-adaptive real-time compensation waveform according to the calibration command to obtain the adjusted load-adaptive real-time compensation waveform.

[0225] For example, step S540 may specifically include the following steps S541 to S545:

[0226] Step S541: Input the calibration command into the command parsing submodule of the configuration adjustment module, perform command parsing, extract the phase adjustment information, amplitude adjustment information and timing configuration adjustment information in the command, and obtain the phase adjustment information, amplitude adjustment information and timing configuration adjustment information. The adjustment information includes the specific adjustment amount.

[0227] The calibration instructions contain adjustment information for the phase, amplitude, and timing of the load-adaptive real-time compensation waveform. However, this information is usually presented in a comprehensive form and requires parsing to extract the specific adjustment information.

[0228] During the parsing process, the instruction parsing submodule carefully analyzes each part of the calibration instruction. For phase adjustment information, it extracts the adjustment direction (clockwise or counterclockwise rotation) and the specific adjustment amount (the angle of rotation). For amplitude adjustment information, it determines whether the amplitude is increased or decreased, and the specific increase or decrease value.

[0229] Regarding timing configuration adjustments, the instruction parsing submodule extracts the adjustment direction (advanced or delayed) and adjustment amount (specific time interval) of the compensation waveform start time, as well as the adjustment direction (extended or shortened) and adjustment amount of the duration.

[0230] Step S542: Input the phase adjustment information into the phase adjustment submodule of the configuration adjustment module, adjust the phase of the load-adaptive real-time compensation waveform, change the phase of the compensation waveform according to the adjustment amount in the phase adjustment information, and obtain the phase-adjusted compensation waveform. The phase-adjusted compensation waveform is opposite to the phase of the ripple.

[0231] After the phase adjustment information is input into the phase adjustment submodule of the configuration adjustment module, this submodule will precisely adjust the phase of the load-adaptive real-time compensation waveform according to the adjustment amount in the phase adjustment information. The phase adjustment submodule typically uses signal processing algorithms to achieve phase adjustment. First, it performs phase analysis on the current compensation waveform to determine its current phase value. Then, according to the adjustment direction and adjustment amount in the phase adjustment information, it changes the phase of the compensation waveform.

[0232] If the adjustment direction is clockwise, the phase adjustment submodule will subtract the adjustment amount from the phase value of the compensation waveform; if it is counterclockwise, it will add the adjustment amount to the phase value. In this way, the phase of the compensation waveform is adjusted to the appropriate position.

[0233] The goal of the adjustment is to make the phase-adjusted compensation waveform out of phase with the ripple. Because when the phases of the compensation waveform and the ripple are out of phase, they can cancel each other out during superposition, effectively reducing the amplitude of the ripple. After the operation of the phase adjustment submodule, the phase-adjusted compensation waveform is obtained, laying the foundation for amplitude and timing adjustments.

[0234] Step S543: Input the amplitude adjustment information into the amplitude adjustment submodule of the configuration adjustment module, adjust the amplitude of the compensation waveform after phase adjustment, change the amplitude of the compensation waveform according to the adjustment amount in the amplitude adjustment information, and obtain the compensation waveform after phase amplitude adjustment. The amplitude of the compensation waveform after phase amplitude adjustment matches the amplitude of the ripple.

[0235] The amplitude adjustment submodule can adjust the amplitude in several ways. For compensation waveforms in analog signal form, amplifier circuits or attenuator circuits can be used to change the signal amplitude. Amplifier circuits can increase the signal amplitude, while attenuator circuits can decrease it. For compensation waveforms in digital signal form, the amplitude adjustment submodule can use digital signal processing algorithms to adjust the amplitude. For example, the value of each sample point of the compensation waveform can be multiplied by a coefficient; if the coefficient is greater than 1, the amplitude is increased; if the coefficient is less than 1, the amplitude is decreased.

[0236] The goal of the adjustment is to match the amplitude of the phase-amplitude adjusted compensation waveform with the amplitude of the ripple. When the amplitude of the compensation waveform is comparable to the amplitude of the ripple, the ripple can be better canceled, improving the ripple suppression effect. After the operation of the amplitude adjustment submodule, the phase-amplitude adjusted compensation waveform is obtained, and the parameters of the compensation waveform are further optimized.

[0237] Step S544: Input the timing configuration adjustment information into the timing adjustment submodule of the configuration adjustment module, adjust the timing configuration information of the compensation waveform after phase and amplitude adjustment, change the start time and duration of the compensation waveform according to the adjustment amount in the timing configuration adjustment information, and obtain the timing-adjusted compensation waveform. The timing of the timing-adjusted compensation waveform matches the load fluctuation.

[0238] The timing adjustment submodule mainly focuses on the start time and duration of the compensation waveform. For the adjustment of the start time, if the timing configuration adjustment information requires an earlier start time, the submodule will adjust the triggering mechanism of the signal to make the compensation waveform start output earlier; if it requires a delayed start time, the triggering time will be delayed accordingly.

[0239] Regarding duration adjustment, if a longer duration is needed, the timing adjustment submodule increases the output duration of the compensation waveform; if a shorter duration is needed, the output duration decreases. The purpose of the adjustment is to match the timing of the timing-adjusted compensation waveform with load fluctuations. Load fluctuations cause ripple to appear and change, and only when the timing of the compensation waveform is synchronized with load fluctuations can ripple be effectively suppressed at the appropriate time. Through the operation of the timing adjustment submodule, a timing-adjusted compensation waveform is obtained, making it more adaptable to dynamic load changes in time.

[0240] Step S545: The timing-adjusted compensation waveform is used as the adjusted load-adaptive real-time compensation waveform and output to the power output ripple real-time monitoring link. The adjusted waveform is sent to the monitoring link through the data transmission channel to complete this adjustment.

[0241] After phase, amplitude, and timing adjustments, the timing-adjusted compensation waveform is used as the adjusted load-adaptive real-time compensation waveform. This waveform is then output to the power supply output ripple real-time monitoring link. The adjusted waveform is sent to the monitoring link via the previously established data transmission channel. The data transmission channel ensures accurate and stable transmission of waveform data to the monitoring link. Sending the adjusted waveform to the monitoring link completes the adjustment process for the load-adaptive real-time compensation waveform. The monitoring link continues to monitor the ripple at the power supply output in real time, acquiring new ripple residual monitoring information. Based on the new monitoring information, the feedback calibration mechanism is restarted to perform a new round of adjustments to the compensation waveform, forming a continuously optimized closed-loop system that continuously improves ripple suppression and ensures the stability of the power supply output.

[0242] Step S550: Feed back the adjusted load-adaptive real-time compensation waveform to the power output ripple real-time monitoring link. Send the adjusted waveform to the monitoring link through the data transmission channel to complete one feedback calibration and enter the next iteration adjustment. The next iteration adjustment will be based on the new ripple residual monitoring information.

[0243] After receiving the adjusted waveform, the power output ripple real-time monitoring link continues to monitor the ripple at the power output in real time. It will collect new residual ripple monitoring information, which reflects the actual ripple suppression effect of the adjusted compensation waveform.

[0244] Based on the new residual ripple monitoring information, the system will enter the next iterative adjustment. In this new round of adjustments, the previous feedback calibration process is repeated: analyzing the new residual ripple monitoring information, determining if calibration is needed, generating calibration commands, and adjusting the phase, amplitude, and timing of the compensation waveform. Through this continuous feedback calibration and iterative adjustment mechanism, the load-adaptive real-time compensation waveform can continuously adapt to changes in power output and dynamic fluctuations in the load, continuously improving the ripple suppression effect, ultimately achieving high efficiency and stability in power output, and providing a more reliable power supply for devices such as computers.

[0245] Please refer to Figure 2, which is a schematic diagram of a computer system according to an embodiment of the present invention. This computer system includes at least a processor 101, a communication interface 102, and a memory 103. The processor 101, communication interface 102, and memory 103 can be connected via a bus or other means. The processor 101 (or Central Processing Unit, CPU) is the computing and control core of the computer system, capable of parsing various instructions and processing various data within the computer system. The communication interface 102 may optionally include standard wired interfaces or wireless interfaces (such as Wi-Fi, mobile communication interfaces, etc.), and can be used to send and receive data under the control of the processor 101; the communication interface 102 can also be used for data transmission and interaction within the computer system. The memory 103 is a storage device in the computer system used to store programs and data. It is understood that the memory 103 here can include the computer system's built-in memory, or it can include extended memory supported by the computer system. The memory 103 provides storage space, which stores the computer system's operating system; this invention does not limit this storage space.

[0246] In one embodiment, the processor 101 executes the deep learning-based computer power supply ripple interference suppression optimization method provided in the above embodiments of the present invention by running a computer program in the memory 103.

Claims

1. A deep learning-based optimization method for suppressing computer power supply ripple interference, characterized in that, The method includes: performing time-domain and frequency-domain joint acquisition and analysis on the ripple signal at the computer power supply output terminal; synchronously associating it with power supply operating status information to enhance spectral features, generating a time-domain and frequency-domain joint ripple spectral feature map and power supply operating status association information; inputting the time-domain and frequency-domain joint ripple spectral feature map and power supply operating status association information into a ripple adversarial generation network; through the collaborative interaction of the feature encoding branch and the compensation signal generation branch, combined with a cross-branch attention fusion mechanism, performing phase inverse modulation and amplitude adaptive calibration to generate a phase inverse calibration virtual compensation ripple signal; inputting the virtual compensation ripple signal and real-time acquired load dynamic fluctuation timing information into a load-ripple dynamic correlation model; adjusting the compensation signal timing configuration information by matching the correlation between ripple characteristics and load fluctuations to generate a load-adaptive real-time compensation waveform; acquiring power supply output energy efficiency timing feedback information and ripple residual monitoring information; inputting these two information and the load-adaptive real-time compensation waveform into the ripple adversarial generation network; and optimizing the network based on a dual-objective optimization criterion. The weights within the network are iteratively optimized in a hierarchical manner to obtain a weight-optimized ripple adversarial generative network. Specifically, the collaborative input information is input into the target decomposition sublayer of the weight optimization layer, and two optimization targets are obtained based on the dual-objective optimization criterion, corresponding to the energy efficiency optimization target and the ripple residue optimization target, respectively. The energy efficiency optimization target is input into the first optimization sublayer of the weight optimization layer, and the weights related to energy efficiency within the network are iteratively optimized. The weights related to energy efficiency information in the feature encoding branch are adjusted to obtain the first part of the optimized weights. The ripple residue optimization target is input into the second optimization sublayer of the weight optimization layer, and the weights related to ripple residue within the network are iteratively optimized. The weights related to ripple residue in the compensation signal generation branch are adjusted to obtain the second part of the optimized weights. Based on the weight-optimized ripple adversarial generative network, a ripple suppression adaptive adjustment link is constructed in conjunction with the real-time monitoring link of the power supply output ripple. Feedback calibration is continuously initiated through ripple residue monitoring information, and the phase, amplitude, and timing configuration information of the load-adaptive real-time compensation waveform are iteratively adjusted.

2. The method as described in claim 1, characterized in that, The process of inputting the time-domain-frequency domain joint ripple spectrum feature map and power supply operating state association information into a ripple adversarial generation network (PGN), and generating a virtual compensated ripple signal with phase inversion modulation and amplitude adaptive calibration through the collaborative interaction of the feature encoding branch and the compensation signal generation branch, combined with a cross-branch attention fusion mechanism, includes: inputting the time-domain-frequency domain joint ripple spectrum feature map into the input layer of the feature encoding branch; performing channel splitting on the time-domain-frequency domain joint ripple spectrum feature map, uniformly splitting the channels according to the proportion of time-domain and frequency-domain dimensions in the feature map, obtaining two parallel feature sub-maps; inputting the power supply operating state association information into the state embedding layer of the feature encoding branch; performing sequence alignment on the power supply operating state association information, adjusting the sequence length of the power supply operating state association information to be consistent with the sequence length of the feature sub-maps, obtaining a state embedding sequence matching the dimension of the feature sub-maps; and inputting the feature sub-maps of the time-domain dimension information, the feature sub-maps of the frequency-domain dimension information, and the state embedding sequence into the fusion layer of the feature encoding branch, performing step-by-step... The elements are concatenated by concatenating the corresponding elements of the three inputs in the order of time domain, frequency domain, and state to obtain pre-encoding fusion information. The number of channels in the pre-encoding fusion information is the sum of the number of channels of the three inputs, and the length of each element is the sum of the lengths of the corresponding elements of the three inputs. The pre-encoding fusion information is then input into the convolutional processing layer of the feature encoding branch for multi-step convolution operations. First, a preliminary convolution is performed using a first convolutional kernel to extract local information, and then a secondary convolution is performed using a second convolutional kernel to extract global correlation information, resulting in an encoded feature representation. The number of channels in the encoded feature representation is equal to the sum of the number of the first and second convolutional kernels, and each channel corresponds to one extracted feature. The encoded feature representation is then input into the output layer of the feature encoding branch for dimensionality compression. Pointwise convolution is used to compress the number of channels in the encoded feature representation to be consistent with the receiving dimension of the compensation signal generation branch, resulting in the output information of the feature encoding branch. The output information of the feature encoding branch is then transmitted to the receiving layer of the compensation signal generation branch through a data transmission line, initiating the collaborative interaction between the feature encoding branch and the compensation signal generation branch.

3. The method as described in claim 2, characterized in that, The step of transmitting the output information of the feature encoding branch to the receiving layer of the compensation signal generation branch via a data transmission line to initiate the collaborative interaction between the feature encoding branch and the compensation signal generation branch includes: inputting the output information of the feature encoding branch into the receiving layer of the compensation signal generation branch; performing a dimensionality transformation on the output information of the feature encoding branch; adjusting the dimension of the output information of the feature encoding branch to be completely consistent with the internal processing dimension of the compensation signal generation branch through a linear transformation matrix, thereby obtaining received information consistent with the internal processing dimension of the compensation signal generation branch; inputting the received information into the cross-branch attention fusion layer of the compensation signal generation branch; calling the cross-branch attention fusion mechanism to associate and match the output information of the feature encoding branch with the current internal state information of the compensation signal generation branch; firstly extracting the key association points between the two; then calculating the matching degree of each association point to obtain the cross-branch attention weight distribution. Each element corresponds to a matching degree of an associated point; based on the cross-branch attention weight distribution, the received information and the current internal state information of the compensation signal generation branch are weighted and integrated. The received information element and the internal state information element of each associated point are multiplied according to their corresponding weights and then summed to obtain the fused state information; the fused state information is input into the phase modulation layer of the compensation signal generation branch for phase inversion modulation. The phase is inverted according to the phase-related information in the fused state information to obtain the phase-modulated intermediate signal. The phase of the phase-modulated intermediate signal is opposite to the phase of the original ripple signal; the phase-modulated intermediate signal is input into the amplitude calibration layer of the compensation signal generation branch for amplitude adaptive calibration. The amplitude information of the intermediate signal is compared and adjusted with a preset amplitude range to obtain a phase-inverted calibrated virtual compensation ripple signal. The amplitude of the virtual compensation ripple signal matches the amplitude of the original ripple signal.

4. The method as described in claim 3, characterized in that, The process involves inputting the received information into the cross-branch attention fusion layer of the compensation signal generation branch, and invoking the cross-branch attention fusion mechanism to correlate and match the output information of the feature encoding branch with the current internal state information of the compensation signal generation branch. This includes first extracting key correlation points between the two, then calculating the matching degree of each correlation point to obtain the cross-branch attention weight distribution. This includes: extracting a key information sequence from the output information of the feature encoding branch, extracting the maximum response value within each window as a key information element using a sliding window, and obtaining a key information sequence whose length is consistent with the sequence length of the feature encoding branch output information; and extracting a query information sequence from the current internal state information of the compensation signal generation branch, extracting the maximum response value within each window as a query information element using the same sliding window, and obtaining a query information sequence whose length is consistent with the internal state information. The key information sequence and the query information sequence are input into the computation layer of the cross-branch attention fusion mechanism. A pairwise similarity calculation is performed on the key information sequence and the query information sequence, calculating the cosine similarity between each key information element and each query information element to obtain an initial similarity matrix. The initial similarity matrix is ​​normalized in the row dimension by dividing the similarity value of each row by the sum of the elements in that row, resulting in a normalized similarity matrix where the sum of elements in each row is 1. The normalized similarity matrix is ​​smoothed in the column dimension by using a weighted average of adjacent columns to smooth the matrix elements, resulting in a smoothed similarity matrix. The smoothed similarity matrix is ​​used as the cross-branch attention weight distribution and output to the subsequent processing layer of the compensation signal generation branch. The subsequent processing layer performs weighted integration processing based on this weight distribution.

5. The method as described in claim 1, characterized in that, The process of inputting the virtual compensation ripple signal and the real-time acquired load dynamic fluctuation timing information into the load-ripple dynamic correlation model, and adjusting the timing configuration information of the compensation signal by matching the correlation between ripple characteristics and load fluctuations to generate a load-adaptive real-time compensation waveform includes: inputting the virtual compensation ripple signal into the signal preprocessing layer of the load-ripple dynamic correlation model, performing timing segmentation on the virtual compensation ripple signal, dividing the virtual compensation ripple signal into multiple continuous signal segments at fixed time intervals, each signal segment having the same duration; inputting the real-time acquired load dynamic fluctuation timing information into the timing preprocessing layer of the load-ripple dynamic correlation model, adjusting the sampling frequency of the load dynamic fluctuation timing information to be consistent with the sampling frequency of the signal segments, obtaining load fluctuation segments with the same sampling frequency as the signal segments, each load fluctuation segment corresponding to one signal segment; and inputting the virtual compensation ripple signal into the signal preprocessing layer of the load-ripple dynamic correlation model, performing timing segmentation on the virtual compensation ripple signal, dividing the virtual compensation ripple signal into multiple continuous signal segments at fixed time intervals, each signal segment having the same duration; and inputting the real-time acquired load dynamic fluctuation timing information into the timing preprocessing layer of the load-ripple dynamic correlation model, adjusting the sampling frequency of the load dynamic fluctuation timing information to be consistent with the sampling frequency of the signal segments, obtaining load fluctuation segments with the same sampling frequency as the signal segments, each load fluctuation segment corresponding to one signal segment; and inputting the real-time acquired load dynamic fluctuation timing information into the signal preprocessing layer of the load-ripple dynamic correlation model, performing timing segmentation on the virtual compensation ripple signal, and adjusting the sampling frequency of the load dynamic fluctuation timing information to be consistent with the sampling frequency of ... obtaining load fluctuation segments with the same sampling frequency as The signal segment and its corresponding load fluctuation segment are input into the correlation matching layer of the load-ripple dynamic correlation model to perform correlation matching between ripple characteristics and load fluctuations. The characteristics of both are extracted and the correlation degree is calculated to obtain the correlation matching result for each signal segment. The correlation matching result reflects the degree of correlation between ripple characteristics and load fluctuations. The correlation matching results corresponding to all signal segments are input into the configuration adjustment layer of the load-ripple dynamic correlation model. The timing configuration information of the compensation signal for each signal segment is adjusted according to the correlation matching result. The start time and duration of each signal segment are adjusted to obtain the signal segment after configuration adjustment. All the signal segments after configuration adjustment are input into the waveform reconstruction layer of the load-ripple dynamic correlation model for timing splicing. All the signal segments after configuration adjustment are spliced ​​sequentially according to the original timing sequence to obtain the load-adaptive real-time compensation waveform. The timing of the load-adaptive real-time compensation waveform matches the timing of the load fluctuations.

6. The method as described in claim 5, characterized in that, The process of inputting the signal segment and the corresponding load fluctuation segment into the association matching layer of the load-ripple dynamic correlation model to perform correlation matching between ripple characteristics and load fluctuations, extracting the characteristics of both and calculating the correlation degree to obtain the association matching result for each signal segment includes: extracting ripple characteristic information from the signal segment by analyzing the waveform shape and change trend of the signal segment, wherein the ripple characteristic information includes the waveform type and change law of the ripple; extracting load fluctuation characteristic information from the corresponding load fluctuation segment by analyzing the fluctuation shape and change trend of the load fluctuation segment, wherein the load fluctuation characteristic information includes the waveform type and change law of the load fluctuation; and inputting the ripple characteristic information and the load fluctuation characteristic information into the feature alignment of the association matching layer. In the sub-layer, dimension alignment is performed to make the dimensions of the two consistent, resulting in aligned ripple characteristic information and aligned load fluctuation characteristic information. The aligned ripple characteristic information and aligned load fluctuation characteristic information are then input into the association calculation sub-layer of the association matching layer to perform time-by-time correlation degree calculation, calculating the waveform similarity of the two at each time point to obtain the correlation degree value at each time point. The correlation degree value at each time point is then input into the sequence integration sub-layer of the association matching layer to perform time-series integration, calculating the average correlation degree value of all time points within each signal segment to obtain the correlation degree sequence corresponding to each signal segment. The correlation degree sequence is then used as the association matching result corresponding to each signal segment and output to the configuration adjustment layer of the load-ripple dynamic association model. The configuration adjustment layer will adjust the timing configuration information of the compensation signal based on the correlation degree sequence.

7. The method as described in claim 6, characterized in that, The step of inputting the correlation matching results corresponding to all signal segments into the configuration adjustment layer of the load-ripple dynamic correlation model, and adjusting the compensation signal timing configuration information of the corresponding signal segment according to each correlation matching result, includes: inputting the correlation degree sequence corresponding to each signal segment into the weight allocation sublayer of the configuration adjustment layer, allocating adjustment weights according to the numerical distribution of the correlation degree sequence, with higher correlation degree values ​​receiving greater weights, to obtain the adjustment weights corresponding to each signal segment; inputting the adjustment weights corresponding to each signal segment into the configuration adjustment sublayer of the configuration adjustment layer, and adjusting the compensation signal timing configuration information of the corresponding signal segment in combination with the adjustment weights, adjusting the signal output time at each moment, to obtain the configured and adjusted signal segments; inputting all configured and adjusted signal segments into the timing sorting sublayer of the configuration adjustment layer, sorting them according to the original timing order to ensure that the order of the signal segments is consistent with the order of the original virtual compensation ripple signal, to obtain the sorted signal segment sequence; inputting the sorted signal segment sequence into the waveform generation sublayer of the configuration adjustment layer, generating continuous waveforms, and filling the gaps between signal segments through linear interpolation to obtain continuous load-adaptive real-time compensation waveforms.

8. The method as described in claim 1, characterized in that, The process of acquiring power output energy efficiency timing feedback information and ripple residual monitoring information, and then inputting these two information along with the load-adaptive real-time compensation waveform into a ripple adversarial generation network (RAG), involves performing hierarchical iterative optimization of the network's internal weights based on a bi-objective optimization criterion to obtain a weight-optimized ripple adversarial generation network. This includes: acquiring power output energy efficiency timing feedback information by collecting energy efficiency information through a built-in energy efficiency monitoring device in the power supply, wherein the energy efficiency timing feedback information includes energy efficiency data at different times; synchronizing the power output energy efficiency timing feedback information by aligning the time of the energy efficiency timing feedback information with the time of the load-adaptive real-time compensation waveform using timestamp alignment, thereby obtaining energy efficiency feedback information with timing consistent with the load-adaptive real-time compensation waveform, wherein the timestamps of the energy efficiency feedback information correspond one-to-one with the timestamps of the compensation waveform; and acquiring ripple residual monitoring information by collecting ripple residual information at the power output through a ripple monitoring device, wherein the ripple residual monitoring information includes data at different times. The residual ripple data is filtered to obtain smoother residual ripple monitoring information. The energy efficiency feedback information, the filtered residual ripple monitoring information, and the load-adaptive real-time compensation waveform are input into the optimization input layer of the ripple adversarial generation network. After normalization, they are aligned according to timestamps and concatenated along the feature dimension to obtain collaborative input information, which includes information on energy efficiency, residual ripple, and compensation waveform. This collaborative input information is then input into the weight optimization layer of the ripple adversarial generation network. A bi-objective optimization criterion is applied to perform layered iterative optimization of the network weights. First, the weights of the feature encoding branch are optimized, then the weights of the compensation signal generation branch are optimized, resulting in iteratively optimized network weights that are more suitable for the current power state. Finally, the iteratively optimized network weights are updated to replace the original network weights, resulting in a weight-optimized ripple adversarial generation network.

9. The method as described in claim 8, characterized in that, The step of inputting the collaborative input information into the weight optimization layer of the ripple adversarial generator network, invoking the bi-objective optimization criterion, and performing hierarchical iterative optimization of the network internal weights, first optimizing the weights of the feature encoding branch, then optimizing the weights of the compensation signal generation branch, to obtain the iteratively optimized network internal weights, includes: inputting the first part of the optimized weights and the second part of the optimized weights into the weight integration sublayer of the weight optimization layer for weight integration, adding the two parts of weights according to a set ratio to obtain the integrated network internal weights; and outputting the integrated network internal weights as the iteratively optimized network internal weights to the weight update layer of the ripple adversarial generator network, wherein the weight update layer updates the network based on the weights.

10. A computer system, characterized in that, include: A memory storing a computer program; a processor for loading the computer program to implement the deep learning-based computer power supply ripple interference suppression optimization method as described in any one of claims 1-9.

Images