A server cluster harmonic modeling method and system based on operating condition clustering

Abstract

This invention belongs to the field of harmonic analysis technology and provides a method and system for harmonic modeling of server clusters based on operating condition clustering. It acquires multi-dimensional operating condition parameters of a single server, preprocesses the acquired data, matches the preprocessed data with a preset equivalent operating condition class of server harmonic characteristics, calls the corresponding dedicated harmonic model, and superimposes them to obtain the harmonic model of the server cluster. If the operating condition parameters do not fall within the parameter range of any preset equivalent operating condition class, it is considered a newly added equivalent operating condition class, and the corresponding dedicated harmonic model is fitted to the dedicated harmonic model library. This invention achieves adaptive calculation of the server cluster harmonic model, reduces the computational complexity of the model, and improves the calculation speed and modeling accuracy.

Description

Technical Field

[0001] This invention belongs to the field of harmonic analysis technology, specifically relating to a harmonic modeling method and system for server clusters based on operating condition clustering. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Currently, artificial intelligence computing centers are continuously expanding in scale, with the number and power density of server clusters constantly increasing, making them typical nonlinear loads in power systems. Servers generate a large amount of harmonic currents during operation, which are injected into the power grid, leading to a decline in power quality. This not only affects the operational stability and lifespan of the servers themselves but may also interfere with the normal operation of other precision equipment within the data center, increase grid losses, and even trigger power system failures. Therefore, constructing accurate harmonic models of server clusters and achieving precise characterization of harmonic characteristics and accurate calculation of harmonic currents are prerequisites and key aspects of power quality management for artificial intelligence computing centers.

[0004] Currently, harmonic modeling of server clusters mainly employs static modeling methods. These methods typically construct fixed harmonic models based on test data from a single rated operating condition of the server. This ignores the diverse operating conditions of servers in actual operation, such as idle, light load, and heavy load, as well as the dynamic changes in harmonic emission characteristics during condition switching. When the server's operating condition deviates from the rated condition, the model accuracy drops significantly. Furthermore, when the cluster structure is adjusted or server operating conditions are switched, manual retesting and remodeling are required, which is cumbersome and inefficient. Simultaneously, current operating condition clustering algorithms usually cluster various operating conditions indiscriminately, without incorporating the neighboring conditions into the clustering criterion calculation. This makes it impossible to dynamically adjust the clustering algorithm based on the discriminative requirements of neighboring operating conditions. Summary of the Invention

[0005] To address the aforementioned problems, this invention proposes a server cluster harmonic modeling method and system based on operational condition clustering. This invention considers the complex operational conditions that servers may encounter, constructs a representation space based on dual-domain feature decoupling and physical information constraints, considers neighborhood conditions, introduces a neighborhood adjustment factor for neighborhood-sensitive adaptive distance metric calculation, and combines an adaptive density peak condition clustering algorithm to divide the server harmonic characteristics into equivalent operational condition classes. This enables adaptive calculation of the server cluster harmonic model, reducing the computational complexity of the model and improving computational speed and modeling accuracy.

[0006] According to some embodiments, the present invention adopts the following technical solution: A harmonic modeling method for server clusters based on operational condition clustering includes the following steps: The system acquires multi-dimensional operating condition parameters of a single server, preprocesses the acquired data, matches the pre-defined server harmonic characteristic equivalent operating condition class based on the pre-processed data, calls the corresponding dedicated harmonic model, and superimposes the data to obtain the harmonic model of the server cluster. If the operating condition parameters do not fall within the parameter range of any pre-defined equivalent operating condition class, it is considered a new equivalent operating condition class. The system then fits the dedicated harmonic model corresponding to the new equivalent operating condition class and adds it to the dedicated harmonic model library. The process of constructing the server harmonic characteristic equivalent operating condition class and the dedicated harmonic model includes: Select various typical servers in the server cluster, simulate all the operating conditions involved in actual operation, obtain real-time operating data of voltage and current of various typical servers under different operating conditions, and perform preprocessing. For each segment of real-time running data from the preprocessed server, frequency domain features and time domain features are extracted respectively. A high-dimensional feature vector space is constructed based on dual-domain optimal transmission and physical information constraints, and neighborhood operating conditions are defined. All high-dimensional feature vectors are used as clustering inputs. Considering the neighborhood conditions, a neighborhood adjustment factor is introduced to calculate the neighborhood-sensitive adaptive distance metric. All conditions are clustered, and conditions with similar harmonic characteristics are grouped into one class to obtain several server harmonic characteristic equivalent condition classes. For each server's equivalent operating condition class of harmonic characteristics, the measured voltage and current sample data of all operating conditions under each class are selected, and the harmonic coupling admittance matrix corresponding to each equivalent operating condition class is fitted, which is used as the exclusive harmonic model for the corresponding equivalent operating condition class.

[0007] As an alternative implementation, the multi-dimensional operating conditions include multiple sub-conditions based on different values ​​of load, GPU utilization, GPU temperature, NPU utilization, NPU temperature, ASIC utilization, ASIC temperature and memory occupancy, comprehensively covering all the operating conditions involved in the actual operation of the server. The acquired multi-dimensional operating condition parameters include real-time operating data of voltage and current under each sub-condition, as well as GPU utilization, GPU temperature, NPU utilization, NPU temperature, ASIC utilization, ASIC temperature and memory occupancy parameters for the corresponding condition, ensuring the synchronization and correspondence between the operating condition parameters and the voltage and current data; Voltage data includes instantaneous phase voltage values, effective phase voltage values, and voltage harmonic components; The current data includes the instantaneous value of the phase current, the effective value of the phase current, and the harmonic components of the current. The acquisition frequency and the data acquisition for each operating condition are set in accordance with international standards.

[0008] As an optional implementation method, the process of preprocessing the acquired data includes: performing preprocessing to remove outliers and noise data from the data, identifying outliers using the 3σ criterion, and performing denoising processing using a wavelet transform algorithm; normalizing the denoised data by using the min-max normalization method to map voltage and current data to the [0, 1] interval to eliminate the influence of dimensions and ensure the accuracy of subsequent feature extraction and cluster analysis; and dividing the preprocessed data into segments, with each set duration of data corresponding to a stable operating condition segment.

[0009] As an alternative implementation, the process of extracting frequency domain features from each segment of preprocessed real-time server operating data includes: performing a Fast Fourier Transform on the preprocessed measured data to extract physical features related to the server's harmonic characteristics: amplitude and phase of the fundamental voltage and current, amplitude and phase of each harmonic voltage and current, total harmonic distortion rate, and harmonic content rate; calculating the mean of all harmonic power spectra in the sample set to obtain a global reference measure. μ ref For the sample i Spectral distribution μ i Solving from the global reference measure μ ref To spectral distribution μ i Optimal transport mapping T i ; Centroids are extracted from the power spectra of all operating conditions to obtain a standard benchmark for the spectral distribution of any sample to be processed. μ i Calculate the global reference measure μ ref With spectral distribution μ i The second-order Wasserstein distance between The calculation formula is: ; In the formula, For the pre-calculated global reference harmonic distribution, For the first i The harmonic power distribution of each operating condition to be identified, where inf is the infimum operator used to lock in the optimal solution among an infinite number of possible mapping functions. x The original frequency coordinates of the energy units in the reference spectrum. For the optimal transfer mapping function, point to x The target coordinates reached after mapping; The nonlinear Wasserstein space is linearized using a Log mapping, and its displacement vector in the tangent space is extracted as an eigenvalue in the frequency domain. F freq .

[0010] As an alternative implementation method, the process of extracting temporal features includes: mining temporal features using a residual temporal convolutional network with a self-attention mechanism; calculating the contribution of regions with significant physical meaning within the waveform using the self-attention mechanism; performing temporal weight calculation; and capturing the intrinsic features of the temporal physical manifold under steady-state conditions. F time ; By forcing the two to align within the same manifold space through physical conservation laws, and using a fusion network with a physical bottleneck layer, the residual between the time-domain average power and the energy of each harmonic in the frequency domain is calculated. The residual is used as a regularization term to dynamically adjust the weights of the two domains. If the feature fusion of a certain operating condition leads to a deterioration in physical consistency, a penalty term is used to correct the feature representation, thereby filtering out false feature points caused by sensor malfunctions or electromagnetic interference. ; ; In the formula This is the fused global high-dimensional feature vector. The vector concatenation operator represents the feature merging of cross-domain information. The fusion weight factor is based on information entropy adaptive adjustment, and its value range is [0, 1]. The effective values ​​of the h-th harmonic voltage and current are... The impedance angle of the h-th harmonic. These are instantaneous samples in the time domain. The physical consistency residual is used to quantify the deviation of the model output from the law of conservation of energy.

[0011] As an alternative implementation, a high-dimensional feature vector space is constructed based on dual-domain optimal transmission and physical information constraints. The process of defining neighborhood operating conditions includes: for each operating condition sample i, a physical configuration vector S is introduced to define its hardware configuration state. ; In the formula For the first i The load conditions of each working condition For the first i GUP utilization rate under various operating conditions For the first i GUP temperature under various operating conditions For the first iNUP utilization rate for each working condition For the first i NUP temperature under various operating conditions For the first i ASIC utilization rate under various operating conditions For the first i ASIC temperature under various operating conditions For the first i Memory occupancy under each working condition; If two working conditions i , j If the physical configuration vectors satisfy the following inequality, then the two are considered to be physical neighbors: ; In the formula The preset step size vector is adjusted according to the expected neighborhood discrimination.

[0012] As an alternative implementation, considering neighborhood conditions, a neighborhood adjustment factor is introduced to calculate a neighborhood-sensitive adaptive distance metric. The clustering process for all conditions includes: using a similarity metric in non-Euclidean geometric space to measure distance; constructing a k-nearest neighbor graph based on all sample points in the fused feature space; and using Dijkstra's algorithm to calculate the shortest path length between any two points in the graph, which is taken as the geodesic distance between the two points. d geo ( i , j Simultaneously, a neighborhood operating condition adjustment factor is introduced. Ω ij To achieve neighborhood-sensitive adaptive distance metric calculation: ; In the formula The compaction factor is used to amplify the characteristic distance between neighboring operating conditions, thereby promoting the subdivision of neighboring operating conditions. The relaxation factor is used to reduce the characteristic distance between neighboring operating conditions and promote the merging of neighboring operating conditions. It is the attenuation factor; For each sample point i Calculate its local density of manifold ρ i Density values ​​reflect the degree to which the corresponding sample points are surrounded by similar operating conditions on the manifold. High-density regions correspond to potential equivalent operating condition classes. i Local density The calculation formula is: ; In the formula, S is the sample set. d cThis is the distance cutoff parameter, used to control the rate of density decay. Its value can be preset or adaptively determined through a data-driven approach. Clustering is performed using an adaptive density peak clustering algorithm based on non-Euclidean geometric manifolds.

[0013] As a further defined implementation, a hard constraint of physical consistency is introduced when performing clustering, for each sample point i Calculate the minimum geodesic distance from it to all sample points with a density higher than its own, denoted as . δ i Construct decision values For each one that passes γ Calculate the physical consistency residuals of the candidate center points selected in the initial screening. Only when When the value is below a preset threshold, the candidate point is ultimately confirmed as a valid working condition category center.

[0014] As a further defined implementation, during clustering, the final clustering results in... N After clustering, the Wasserstein centroid is used as the representation method for each server harmonic characteristic equivalent operating condition class. This is achieved by calculating the average distance from all samples within the cluster to the centroid. W 2. Distance, generating the center vector that best matches the physical eigenvalues ​​in the frequency domain. V center The calculation formula is: ; In the formula, For the characteristic distribution of the kth working condition sample belonging to this cluster, Cluster c For the first c The set of all samples contained in each equivalent working condition class.

[0015] As an alternative implementation, for each server harmonic characteristic equivalent operating condition class, the process of filtering measured voltage and current sample data for all operating conditions under each class and fitting the harmonic coupling admittance matrix corresponding to each equivalent operating condition class includes: for the obtained N For each server harmonic characteristic equivalent operating condition class, the measured voltage and current sample data of all sub-conditions under each equivalent operating condition class are screened one by one, and the sample dataset of that equivalent operating condition class is summarized. Combining the measured voltage and current data under each equivalent operating condition class, the specific harmonic coupling admittance matrix model of each equivalent operating condition class is obtained by fitting using the weighted least squares method. The expression is as follows: ; In the formula, Indicates the first k Second harmonic voltage value; Indicates the firsth The value of the second harmonic current; Indicates the first h The second harmonic current and the first k Coupling relationship of subharmonic voltages; H This represents the highest harmonic order considered in the model. The fitting process is based on all measured samples under this operating condition, with the objective of minimizing the sum of squared residuals between the measured harmonic current and the fitted harmonic current. The weighted least squares method is used to solve for the coupling admittance matrix parameters. The objective function and solution formula are as follows: ; In the formula The higher harmonic weighting factor has a value of [value missing]. The higher the harmonic order, the larger the weighting factor. For the first i The first sample h Measured current value of the subharmonic; For the first i The first sample k Measured voltage value of the subharmonic; N This represents the number of measured samples under this harmonic emission condition; the meanings of the other symbols are the same as described above.

[0016] As a further solution, the construction process of the dedicated harmonic model includes: comparing each server operating condition with the corresponding parameter range of each equivalent operating condition class, counting the number of servers operating in each equivalent operating condition class, and then superimposing the data according to the following formula to obtain the harmonic model of the server cluster: ; In the formula The harmonic coupling admittance matrix of the server cluster obtained by superposition is... For a certain model j The server is running under equivalent operating conditions i The number of servers under, The corresponding model obtained by fitting the measured data j The server is operating under equivalent conditions. i The harmonic coupling admittance matrix at time.

[0017] A server cluster harmonic modeling system based on operating condition clustering includes: The data preprocessing module is configured to select various typical servers in the server cluster, simulate all the relevant operating conditions in actual operation, obtain real-time operating data of voltage and current of various typical servers under different operating conditions, and perform preprocessing. The feature extraction module is configured to extract frequency domain features and time domain features from each segment of real-time running data of the preprocessed server, and construct a high-dimensional feature vector space based on dual-domain optimal transmission and physical information constraints to define neighborhood conditions. The working condition clustering module is configured to take all high-dimensional feature vectors as clustering input, consider the neighborhood working conditions, introduce a neighborhood adjustment factor to calculate the neighborhood-sensitive adaptive distance metric, cluster all working conditions, and group working conditions with similar harmonic characteristics into one class to obtain several server harmonic characteristic equivalent working condition classes. The model fitting module is configured to filter the measured voltage and current sample data of all operating conditions under each server harmonic characteristic equivalent operating condition class, and fit the harmonic coupling admittance matrix corresponding to each equivalent operating condition class, which is then used as the exclusive harmonic model for the corresponding equivalent operating condition class. The harmonic execution module is configured to acquire multi-dimensional operating condition parameters of a single server, preprocess the acquired data, match the preprocessed data with a preset equivalent operating condition class of server harmonic characteristics, call the corresponding dedicated harmonic model, and superimpose to obtain the harmonic model of the server cluster; if the operating condition parameters do not fall within the parameter range of any preset equivalent operating condition class, it is regarded as a new equivalent operating condition class, and the dedicated harmonic model corresponding to the new equivalent operating condition class is fitted and added to the dedicated harmonic model library.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention adds a multi-dimensional server operating condition preset and clustering strategy, which fully covers the combination of GPU / NPU / ASIC utilization, temperature, memory occupancy and load conditions. By clustering, subdivided operating conditions with similar harmonic characteristics are grouped into one category and share a dedicated model. This solves the problem of the surge in the number of models and the complexity of calculation caused by the change of server operating conditions. At the same time, the clustering criteria ensure the consistency of harmonic characteristics of the same type of operating conditions, taking into account the dual needs of simplified calculation and accurate cluster modeling. (2) By introducing the physical configuration vector neighborhood weighting operator, this invention introduces whether the server's working condition is in the neighborhood into the clustering criterion calculation, breaking the limitation of the traditional clustering algorithm that analyzes all working conditions without difference. For neighborhood working conditions with only slight differences in hardware status, the geodesic distance weight can be adaptively adjusted to tighten or loosen the geodesic distance calculation according to the neighborhood discrimination expected. (3) This invention establishes a dedicated harmonic coupling model for various equivalent operating conditions of the server, breaking the limitation that traditional models cannot take into account harmonic coupling. Moreover, only the server operating conditions and the total voltage of the cluster are needed to perform harmonic characteristic analysis and cluster harmonic current calculation. When the server operating conditions are switched, the cluster structure is adjusted, or a new equivalent operating condition is added, the dedicated model of the new category can be automatically fitted to achieve adaptive model expansion and adapt to the dynamic operating requirements of the server cluster.

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0020] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0021] Figure 1 This is a flowchart of a server cluster harmonic modeling method based on operating condition clustering in one embodiment; Figure 2 This is a flowchart of the core algorithm in one embodiment. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0023] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0024] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0025] Where there is no conflict, the embodiments and features described in this application may be combined with each other.

[0026] Example 1 A server cluster harmonic modeling method based on operating condition clustering, such as Figure 1 As shown, it specifically includes the following: S101: Select various typical servers in the server cluster, simulate all the operating conditions involved in actual operation, and collect real-time operating data of voltage and current of various servers under different operating conditions; preprocess the collected raw data.

[0027] The server operates under pre-defined multi-dimensional conditions, which are divided into the following categories: 1. Load conditions: No load (load rate < 20%), light load (load rate 20%-40%), medium load (load rate 40%-60%), heavy load (load rate 60%-80%), peak load (load rate above 80%). 2. GPU operating dimensions: GPU utilization (0%-20%, 20%-40%, 40%-60%, 60%-80%, 80%-100%), GPU temperature (<50℃, 50℃-60℃, 60℃-70℃, 70℃-80℃, above 80℃); 3. NPU operation: NPU utilization rate (0%-20%, 20%-40%, 40%-60%, 60%-80%, 80%-100%), NPU temperature (<40℃, 40℃-50℃, 50℃-60℃, 60℃-70℃, above 70℃); 4. AI Accelerated Chip (ASIC) Operation: ASIC utilization rate (0%-20%, 20%-40%, 40%-60%, 60%-80%, 80%-100%), ASIC temperature (<40℃, 40℃-50℃, 50℃-60℃, 60℃-70℃, above 70℃); 5. Memory operation dimension: memory utilization (0%-20%, 20%-40%, 40%-60%, 60%-80%, 80%-100%).

[0028] In this embodiment, the combination of parameters from various dimensions forms a complete set of multi-dimensional operating conditions for the server. This set contains at most 5 (load) × 5 (GPU utilization) × 5 (GPU temperature) × 5 (NPU utilization) × 5 (NPU temperature) × 5 (ASIC utilization) × 5 (ASIC temperature) × 5 (memory occupancy) = 390,625 sub-conditions, comprehensively covering all operating conditions involved in actual server operation. Typical servers of different brands, models, and configurations are selected from the server cluster. The various sub-conditions mentioned above are simulated through the server operation and maintenance management platform. Real-time operating data of voltage and current for each type of server under each sub-condition are collected. Simultaneously, the GPU utilization, GPU temperature, NPU utilization, NPU temperature, ASIC utilization, ASIC temperature, and memory occupancy parameters for the corresponding conditions are recorded through sensors and the operation and maintenance platform to ensure the synchronization and correspondence between the operating condition parameters and the voltage and current data. The collected voltage data includes instantaneous phase voltage, effective phase voltage, and voltage harmonic components; the collected current data includes instantaneous phase current, effective phase current, and current harmonic components; the collection frequency and data collection under each operating condition are set according to international standards.

[0029] The collected raw data were preprocessed to remove outliers and noise. Outliers were identified using the 3σ criterion and denoised using wavelet transform. The denoised data were then normalized by using the min-max normalization method to map voltage and current data to the [0, 1] interval, eliminating the influence of dimensions and ensuring the accuracy of subsequent feature extraction and cluster analysis. Finally, the preprocessed data was segmented, with each 1 second of data corresponding to a stable operating condition segment, providing a data foundation for subsequent feature extraction.

[0030] S102: For each segment of server running data after preprocessing in step S101, extract frequency domain measures and mine time domain manifolds respectively. Construct a high-dimensional feature space based on dual-domain optimal transmission and physical information constraints, and define neighborhood working conditions.

[0031] Fast Fourier Transform (FFT) is performed on the preprocessed measured data to extract physical features related to the server's harmonic characteristics: amplitude and phase of the fundamental voltage and current, amplitude and phase of each harmonic voltage and current, total harmonic distortion (THD), and harmonic content (HRI). The mean of the power spectra of all harmonics in the sample set is calculated to obtain the global reference measure. μ ref For the sample i Spectral distribution μ i Solve from μ ref arrive μ i Optimal transport mapping T i .

[0032] A standard benchmark is obtained by extracting the centroid of the power spectrum of all historical operating conditions in the sample database. For any sample to be processed... μ i ,calculate μ ref and μ i The second-order Wasserstein distance between The calculation formula is: ; In the formula For the pre-calculated global reference harmonic distribution, For the first i The harmonic power distribution of each operating condition to be identified, where inf is the infimum operator used to lock in the optimal solution among an infinite number of possible mapping functions. x The original frequency coordinates of the energy units in the reference spectrum. For the optimal transfer mapping function, point to xThe target coordinates reached after mapping.

[0033] The nonlinear Wasserstein space is linearized using a Log mapping, and its displacement vector in the tangent space is extracted as an eigenvalue in the frequency domain. F freq .

[0034] A residual temporal convolutional network with a self-attention mechanism is employed to mine temporal features. The self-attention mechanism is used to calculate the contribution of regions with significant physical meaning within the waveform, thereby performing temporal weight calculation and capturing the intrinsic features of the temporal physical manifold under steady-state conditions. F time .

[0035] like Figure 2 As shown, after obtaining the dual-domain features, the two are forced to align in the same manifold space by physical conservation laws. Using a fusion network with a physical bottleneck layer, the residual between the time-domain average power and the harmonic energies in the frequency domain is calculated. The residual is used as a regularization term to dynamically adjust the weights of the two domains. If the feature fusion of a certain working condition leads to a deterioration in physical consistency, the system will correct the feature expression through a penalty term, thereby filtering out false feature points caused by sensor failure or electromagnetic interference.

[0036] ; ; In the formula This is the fused global high-dimensional feature vector. The vector concatenation operator represents the feature merging of cross-domain information. The fusion weight factor is based on information entropy adaptive adjustment, and its value range is [0, 1]. The effective values ​​of the h-th harmonic voltage and current are... The impedance angle of the h-th harmonic. These are instantaneous samples in the time domain. The physical consistency residual is used to quantify the deviation of the model output from the law of conservation of energy.

[0037] For each working condition sample i Introducing physical configuration vectors S Define its hardware configuration status: ; In the formula For the first i The load conditions of each working condition (divided into five intervals numbered 1-5 according to the load rate from low to high, and the following numbering scheme is similar and will not be repeated). For the first iGUP utilization rate under various operating conditions For the first i GUP temperature under various operating conditions For the first i NUP utilization rate for each working condition For the first i NUP temperature under various operating conditions For the first i ASIC utilization rate under various operating conditions For the first i ASIC temperature under various operating conditions For the first i Memory usage under each working condition.

[0038] At the same time, if two operating conditions i , j If the physical configuration vectors satisfy the following inequality, then the two are determined to be physical neighboring operating conditions: ; In the formula The preset step size vector can be adjusted according to the expected neighborhood discrimination. For example, if it is desired that a certain dimension is in the neighborhood condition, then... Set to 1.

[0039] S103: Take all the high-dimensional feature vectors obtained in S102 as clustering input, and consider the neighborhood conditions. Introduce a neighborhood adjustment factor to calculate the neighborhood-sensitive adaptive distance metric. Use an adaptive density peak clustering algorithm based on non-Euclidean geometric manifold to cluster all the subdivided conditions in the multi-dimensional condition set. Group the subdivided conditions with similar harmonic characteristics into one class to obtain several server harmonic equivalent condition classes.

[0040] Distance is measured using a similarity metric in a non-Euclidean geometric space, and is constructed based on all sample points in the fused feature space. k In a nearest neighbor graph, Dijkstra's algorithm is used to calculate the shortest path length between any two points, which is then used as the geodesic distance between the two points. d geo ( i , j Simultaneously, a neighborhood operating condition adjustment factor is introduced. Ω- ij This enables adaptive distance measurement calculation that is sensitive to neighborhood operating conditions.

[0041] ; In the formula The compaction factor is used to amplify the characteristic distance between neighboring operating conditions, thereby promoting the subdivision of neighboring operating conditions. The relaxation factor is used to reduce the characteristic distance between neighboring operating conditions and promote the merging of neighboring operating conditions. As a decay factor, it ensures that the weight adjustment is significant in strict neighborhood states and decays rapidly as neighborhood differences increase.

[0042] Based on this, for each sample point i Calculate its local density of manifold ρ i The density value reflects how tightly a point is surrounded by similar operating conditions on the manifold. High-density regions correspond to potential equivalent operating condition classes. The local density of sample point i... The calculation formula is as follows: ; In the formula, S is the sample set. d c This is the distance cutoff parameter, used to control the rate of density decay. Its value can be preset or adaptively determined through data-driven methods.

[0043] At the same time, a hard constraint of physical consistency is introduced for each sample point. i Calculate the minimum geodesic distance from it to all sample points with a density higher than its own, denoted as . δ i Construct decision values For each one that passes γ The candidate center points selected in the initial screening are traced back to their physical consistency residuals calculated during the S102 fusion stage. Only when Only when the value is below a preset threshold is the candidate point finally confirmed as a valid working condition category center.

[0044] In this embodiment, the final clustering results in N After clustering, the Wasserstein centroid is used as the representation method for each server harmonic characteristic equivalent operating condition class. This is achieved by calculating the average distance from all samples within the cluster to the centroid. W 2. Distance, generating the center vector that best matches the physical eigenvalues ​​in the frequency domain. V center The calculation formula is as follows: ; In the formula For the characteristic distribution of the kth working condition sample belonging to this cluster, Cluster c It is the set of all samples contained in the c-th equivalent working condition class.

[0045] S104: For each server harmonic characteristic equivalent operating condition class obtained in S103, the measured voltage and current sample data of all sub-operating conditions under each class are selected, and the harmonic coupling admittance matrix corresponding to the equivalent operating condition class is obtained by fitting with weighted least squares method, and it is used as the exclusive harmonic model of the equivalent operating condition class.

[0046] Regarding the results obtained in step S103 N For each server harmonic characteristic equivalent operating condition class, the measured voltage and current sample data of all sub-conditions under each equivalent operating condition class are screened and summarized to form a sample dataset for that equivalent operating condition class. Combining the measured voltage and current data under each equivalent operating condition class, a specific harmonic coupling admittance matrix model for each equivalent operating condition class is obtained by fitting using the weighted least squares method. The expression is as follows: ; In the formula, Indicates the first k Second harmonic voltage value; Indicates the first h The value of the second harmonic current; Indicates the first h The second harmonic current and the first k Coupling relationship of subharmonic voltages; H This represents the highest harmonic order considered in the model. The fitting process is based on all measured samples under this operating condition, with the objective of minimizing the sum of squared residuals between the measured harmonic current and the fitted harmonic current. The weighted least squares method is used to solve for the coupling admittance matrix parameters. The objective function and solution formula are as follows: ; In the formula The higher harmonic weighting factor has a value of [value missing]. The higher the harmonic order, the larger the weighting factor. Let be the measured current value of the h-th harmonic of the i-th sample; Let be the measured voltage value of the kth harmonic of the i-th sample; N This represents the number of measured samples under this harmonic emission condition; the meanings of the other symbols are the same as described above.

[0047] S105: During the actual operation of the server cluster, monitor the multi-dimensional operating condition parameters of a single server in real time, match the server harmonic characteristics equivalent operating condition class obtained in S103, call the corresponding dedicated harmonic model, and superimpose to obtain the harmonic model of the server cluster; if the real-time operating condition parameters do not fall into the parameter range of any preset equivalent operating condition class, fit the dedicated harmonic model corresponding to the newly added equivalent operating condition class and add it to the harmonic model library.

[0048] During the actual operation of the server cluster, multi-dimensional operating condition parameters of a single server are collected in real time. These parameters are consistent with the operating condition parameters preset in step S101, including load conditions (idle, light load, medium load, heavy load, peak load), GPU utilization, GPU temperature, NPU utilization, NPU temperature, ASIC utilization, ASIC temperature, and memory occupancy. The server operating conditions are then compared with... NThe corresponding parameter ranges of each equivalent operating condition class are compared, and the number of servers running in each equivalent operating condition class is counted. The harmonic model of the server cluster is obtained by superimposing the data according to the following formula: ; In the formula The harmonic coupling admittance matrix of the server cluster obtained by superposition is... For a certain model j The server is running under equivalent operating conditions i The number of servers under, This is a model obtained by fitting measured data. j The server is operating under equivalent conditions. i The harmonic coupling admittance matrix at time.

[0049] If the real-time operating parameters do not fall within the parameter range of any preset equivalent operating condition class, it is determined to be a new equivalent operating condition class. The historical voltage and current data corresponding to the real-time operating parameters are used as the measured samples of the new equivalent operating condition class. The clustering and model fitting process of steps S102-S105 is repeated to obtain the exclusive harmonic model of the new equivalent operating condition class. It is added to the harmonic model library and the "equivalent operating condition class - model" relationship is updated to realize the adaptive expansion of the model.

[0050] Example 2 A server cluster harmonic modeling system based on operating condition clustering includes: The data preprocessing module is configured to select various typical servers in the server cluster, simulate all the relevant operating conditions in actual operation, obtain real-time operating data of voltage and current of various typical servers under different operating conditions, and perform preprocessing. The feature extraction module is configured to extract frequency domain features and time domain features from each segment of real-time running data of the preprocessed server, and construct a high-dimensional feature vector space based on dual-domain optimal transmission and physical information constraints to define neighborhood conditions. The working condition clustering module is configured to take all high-dimensional feature vectors as clustering input, consider the neighborhood working conditions, introduce a neighborhood adjustment factor to calculate the neighborhood-sensitive adaptive distance metric, cluster all working conditions, and group working conditions with similar harmonic characteristics into one class to obtain several server harmonic characteristic equivalent working condition classes. The model fitting module is configured to filter the measured voltage and current sample data of all operating conditions under each server harmonic characteristic equivalent operating condition class, and fit the harmonic coupling admittance matrix corresponding to each equivalent operating condition class, which is then used as the exclusive harmonic model for the corresponding equivalent operating condition class. The harmonic execution module is configured to acquire multi-dimensional operating condition parameters of a single server, preprocess the acquired data, match the preprocessed data with a preset equivalent operating condition class of server harmonic characteristics, call the corresponding dedicated harmonic model, and superimpose to obtain the harmonic model of the server cluster; if the operating condition parameters do not fall within the parameter range of any preset equivalent operating condition class, it is regarded as a new equivalent operating condition class, and the dedicated harmonic model corresponding to the new equivalent operating condition class is fitted and added to the dedicated harmonic model library.

[0051] It should be noted that the specific implementation methods of the above modules are exactly the same as those in Example 1, and will not be described in detail here.

[0052] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of one or more computer-usable storage media (including, but not limited to, disk storage, etc.) containing computer-usable program code. CD - ROM It takes the form of a computer program product implemented on (such as optical memory, etc.).

[0053] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0054] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0055] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0056] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made by those skilled in the art without creative effort within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A server cluster harmonic modeling method based on operating condition clustering, characterized in that, Includes the following steps: The system acquires multi-dimensional operating condition parameters of a single server, preprocesses the acquired data, matches the pre-defined server harmonic characteristic equivalent operating condition class based on the pre-processed data, calls the corresponding dedicated harmonic model, and superimposes the data to obtain the harmonic model of the server cluster. If the operating condition parameters do not fall within the parameter range of any pre-defined equivalent operating condition class, it is considered a new equivalent operating condition class. The system then fits the dedicated harmonic model corresponding to the new equivalent operating condition class and adds it to the dedicated harmonic model library. The process of constructing the server harmonic characteristic equivalent operating condition class and the dedicated harmonic model includes: Select various typical servers in the server cluster, simulate all the operating conditions involved in actual operation, obtain real-time operating data of voltage and current of various typical servers under different operating conditions, and perform preprocessing. For each segment of real-time running data from the preprocessed server, frequency domain features and time domain features are extracted respectively. A high-dimensional feature vector space is constructed based on dual-domain optimal transmission and physical information constraints, and neighborhood operating conditions are defined. All high-dimensional feature vectors are used as clustering inputs. Considering the neighborhood conditions, a neighborhood adjustment factor is introduced to calculate the neighborhood-sensitive adaptive distance metric. All conditions are clustered, and conditions with similar harmonic characteristics are grouped into one class to obtain several server harmonic characteristic equivalent condition classes. For each server's equivalent operating condition class of harmonic characteristics, the measured voltage and current sample data of all operating conditions under each class are selected, and the harmonic coupling admittance matrix corresponding to each equivalent operating condition class is fitted, which is used as the exclusive harmonic model for the corresponding equivalent operating condition class.

2. The server cluster harmonic modeling method based on operating condition clustering as described in claim 1, characterized in that, The multi-dimensional operating conditions include multiple sub-conditions based on different values ​​of load, GPU utilization, GPU temperature, NPU utilization, NPU temperature, ASIC utilization, ASIC temperature, and memory occupancy, comprehensively covering all the operating conditions involved in the actual operation of the server. The acquired multi-dimensional operating condition parameters include real-time operating data of voltage and current under each sub-condition, as well as GPU utilization, GPU temperature, NPU utilization, NPU temperature, ASIC utilization, ASIC temperature and memory occupancy parameters for the corresponding condition, ensuring the synchronization and correspondence between the operating condition parameters and the voltage and current data; Voltage data includes instantaneous phase voltage values, effective phase voltage values, and voltage harmonic components; The current data includes the instantaneous value of the phase current, the effective value of the phase current, and the harmonic components of the current. The acquisition frequency and the data acquisition for each operating condition are set in accordance with international standards.

3. The server cluster harmonic modeling method based on operating condition clustering as described in claim 1, characterized in that, The process of extracting frequency domain features from each segment of preprocessed real-time server data includes: performing a Fast Fourier Transform on the preprocessed measured data to extract physical features related to the server's harmonic characteristics: amplitude and phase of the fundamental voltage and current, amplitude and phase of each harmonic voltage and current, total harmonic distortion rate, and harmonic content rate; calculating the mean of all harmonic power spectra in the sample set to obtain a global reference measure. μ ref For the sample i Spectral distribution μ i Solving from the global reference measure μ ref To spectral distribution μ i Optimal transport mapping T i ; Centroids are extracted from the power spectra of all operating conditions to obtain a standard benchmark for the spectral distribution of any sample to be processed. μ i Calculate the global reference measure μ ref With spectral distribution μ i The second-order Wasserstein distance between The calculation formula is: ； In the formula, For the pre-calculated global reference harmonic distribution, For the first i The harmonic power distribution of each operating condition to be identified, where inf is the infimum operator used to lock in the optimal solution among an infinite number of possible mapping functions. x The original frequency coordinates of the energy units in the reference spectrum. For the optimal transfer mapping function, point to x The target coordinates reached after mapping; The nonlinear Wasserstein space is linearized using a Log mapping, and its displacement vector in the tangent space is extracted as an eigenvalue in the frequency domain. F freq .

4. The server cluster harmonic modeling method based on operating condition clustering as described in claim 1, characterized in that, The process of extracting temporal features includes: using a residual temporal convolutional network with a self-attention mechanism to mine temporal features; calculating the contribution of regions with significant physical meaning within the waveform through the self-attention mechanism; performing temporal weight calculation; and capturing the intrinsic features of the temporal physical manifold under steady-state conditions. F time ; By forcing the two to align within the same manifold space through physical conservation laws, and using a fusion network with a physical bottleneck layer, the residual between the time-domain average power and the energy of each harmonic in the frequency domain is calculated. The residual is used as a regularization term to dynamically adjust the weights of the two domains. If the feature fusion of a certain operating condition leads to a deterioration in physical consistency, a penalty term is used to correct the feature representation, thereby filtering out false feature points caused by sensor malfunctions or electromagnetic interference. ； ； In the formula This is the fused global high-dimensional feature vector. The vector concatenation operator represents the feature merging of cross-domain information. The fusion weight factor is based on information entropy adaptive adjustment, and its value range is [0, 1]. The effective values ​​of the h-th harmonic voltage and current are... The impedance angle of the h-th harmonic. These are instantaneous samples in the time domain. The physical consistency residual is used to quantify the deviation of the model output from the law of conservation of energy.

5. The server cluster harmonic modeling method based on operating condition clustering as described in claim 1, characterized in that, Based on dual-domain optimal transmission and physical information constraints, a high-dimensional feature vector space is constructed. The process of defining neighborhood operating conditions includes: for each operating condition sample i, a physical configuration vector S is introduced to define its hardware configuration state. In the formula For the first i Load conditions for each working condition For the first i GUP utilization rate under various operating conditions For the first i GUP temperature under various operating conditions For the first i NUP utilization rate for each working condition For the first i NUP temperature under various operating conditions For the first i ASIC utilization rate under various operating conditions For the first i ASIC temperature under various operating conditions For the first i Memory occupancy under each working condition; If two working conditions i , j If the physical configuration vectors satisfy the following inequality, then the two are considered to be physical neighbors: In the formula The preset step size vector is adjusted according to the expected neighborhood discrimination.

6. The server cluster harmonic modeling method based on operating condition clustering as described in claim 1, characterized in that, Considering neighborhood conditions, a neighborhood adjustment factor is introduced to calculate a neighborhood-sensitive adaptive distance metric. The clustering process for all conditions includes: using a similarity metric in non-Euclidean geometric space to measure distance; constructing a k-nearest neighbor graph based on all sample points in the fused feature space; and using Dijkstra's algorithm to calculate the shortest path length between any two points in the graph, which is then used as the geodesic distance between the two points. d geo ( i , j Simultaneously, a neighborhood operating condition adjustment factor is introduced. Ω ij To achieve neighborhood-sensitive adaptive distance metric calculation: In the formula The compaction factor is used to amplify the characteristic distance between neighboring operating conditions, thereby promoting the subdivision of neighboring operating conditions. The relaxation factor is used to reduce the characteristic distance between neighboring operating conditions and promote the merging of neighboring operating conditions. It is the attenuation factor; For each sample point i Calculate its local density of manifold ρ i Density values ​​reflect the degree to which the corresponding sample points are surrounded by similar operating conditions on the manifold. High-density regions correspond to potential equivalent operating condition classes. i Local density The calculation formula is: In the formula, S is the sample set. d c This is the distance cutoff parameter, used to control the rate of density decay. Its value can be preset or adaptively determined through a data-driven approach. Clustering is performed using an adaptive density peak clustering algorithm based on non-Euclidean geometric manifolds.

7. The server cluster harmonic modeling method based on operating condition clustering as described in claim 6, characterized in that, When performing clustering, a hard constraint of physical consistency is introduced for each sample point. i Calculate the minimum geodesic distance from it to all sample points with a density higher than its own, denoted as . δ i Construct decision values For each one that passes γ Calculate the physical consistency residuals of the candidate center points selected in the initial screening. Only when When the value is below a preset threshold, the candidate point is ultimately confirmed as a valid working condition category center.

8. The server cluster harmonic modeling method based on operating condition clustering as described in claim 6, characterized in that, When performing clustering, the final clustering results in N After clustering, the Wasserstein centroid is used as the representation method for each server harmonic characteristic equivalent operating condition class. This is achieved by calculating the average distance from all samples within the cluster to the centroid. W 2. Distance, generating the center vector that best matches the physical eigenvalues ​​in the frequency domain. V center The calculation formula is: In the formula, For the characteristic distribution of the kth working condition sample belonging to this cluster, Cluster c For the first c The set of all samples contained in each equivalent working condition class.

9. The server cluster harmonic modeling method based on operating condition clustering as described in claim 1, characterized in that, For each server harmonic characteristic equivalent operating condition class, the process of selecting measured voltage and current sample data for all operating conditions under each class and fitting the harmonic coupling admittance matrix corresponding to each equivalent operating condition class includes: [The process is repeated here, so the translation ends abruptly.] N For each server harmonic characteristic equivalent operating condition class, the measured voltage and current sample data of all sub-conditions under each equivalent operating condition class are screened one by one, and the sample dataset of that equivalent operating condition class is summarized. Combining the measured voltage and current data under each equivalent operating condition class, the specific harmonic coupling admittance matrix model of each equivalent operating condition class is obtained by fitting using the weighted least squares method. The expression is as follows: In the formula, Indicates the first k Second harmonic voltage value; Indicates the first h The value of the second harmonic current; Indicates the first h The second harmonic current and the first k Coupling relationship of subharmonic voltages; H This represents the highest harmonic order considered in the model. The fitting process is based on all measured samples under the corresponding operating conditions, with the objective of minimizing the sum of squared residuals between the measured harmonic current and the fitted harmonic current. The weighted least squares method is used to solve for the coupling admittance matrix parameters. The objective function and solution formula are as follows: In the formula The higher harmonic weighting factor has a value of [value missing]. The higher the harmonic order, the larger the weighting factor. For the first i The first sample h Measured current value of the subharmonic; For the first i The first sample k Measured voltage value of the subharmonic; N This represents the number of measured samples under this harmonic emission condition; the meanings of the other symbols are the same as described above.

10. A server cluster harmonic modeling system based on operating condition clustering, characterized in that, include: The data preprocessing module is configured to select various typical servers in the server cluster, simulate all the relevant operating conditions in actual operation, obtain real-time operating data of voltage and current of various typical servers under different operating conditions, and perform preprocessing. The feature extraction module is configured to extract frequency domain features and time domain features from each segment of real-time running data of the preprocessed server, and construct a high-dimensional feature vector space based on dual-domain optimal transmission and physical information constraints to define neighborhood conditions. The working condition clustering module is configured to take all high-dimensional feature vectors as clustering input, consider the neighborhood working conditions, introduce a neighborhood adjustment factor to calculate the neighborhood-sensitive adaptive distance metric, cluster all working conditions, and group working conditions with similar harmonic characteristics into one class to obtain several server harmonic characteristic equivalent working condition classes. The model fitting module is configured to filter the measured voltage and current sample data of all operating conditions under each server harmonic characteristic equivalent operating condition class, and fit the harmonic coupling admittance matrix corresponding to each equivalent operating condition class, which is then used as the exclusive harmonic model for the corresponding equivalent operating condition class. The harmonic execution module is configured to acquire multi-dimensional operating condition parameters of a single server, preprocess the acquired data, match the preprocessed data with a preset equivalent operating condition class of server harmonic characteristics, call the corresponding dedicated harmonic model, and superimpose to obtain the harmonic model of the server cluster; if the operating condition parameters do not fall within the parameter range of any preset equivalent operating condition class, it is regarded as a new equivalent operating condition class, and the dedicated harmonic model corresponding to the new equivalent operating condition class is fitted and added to the dedicated harmonic model library.

Images