A server limit power consumption prediction method and system

Abstract

The application discloses a kind of server limit power consumption prediction method and system, it is related to server energy consumption monitoring technical field, including obtaining multi-source data and preprocessing, obtain original time series dataset, construct lag feature matrix based on original time series dataset, analyze thermal coupling derived features and carry out supplementary coding to obtain the time series feature matrix of temperature lag effect, construct multi-branch fusion deep learning model, based on multi-branch fusion deep learning model, model fitting and parameter optimization training are carried out to time series feature matrix, obtain target limit power consumption prediction model, based on the real-time feature data of target server and target cluster, hardware configuration data and computer room cabinet topology data are predicted using target limit power consumption prediction model and are handled using multi-dimensional engineering margin correction to obtain the limit power consumption prediction value of engineering adaptation single machine and cluster two levels.The application realizes limit power consumption high-precision prediction using the above method, solves the problem of prediction deviation caused by temperature lag.

Description

Technical Field

[0001] This invention relates to the field of server energy consumption monitoring and data center operation and maintenance technology, and in particular to a method and system for predicting server power consumption limits. Background Technology

[0002] Server power consumption limits are core technical indicators for data center power distribution capacity planning, heat dissipation system design, and power supply security management, directly determining infrastructure redundancy configuration and operation and maintenance risk prevention and control.

[0003] Current server power consumption prediction technologies generally suffer from the following technical shortcomings: First, they rely solely on static features such as real-time CPU utilization and instantaneous temperature for modeling, completely ignoring the fact that server power consumption changes are not instantaneous responses to load utilization, but are affected by the accumulation of heat and heat conduction delays in core heat-generating components such as CPU, GPU, and NVMe, resulting in a significant and dynamic temperature lag effect. This leads to a large deviation between the predicted extreme power consumption and the actual peak value. Second, most models use a single CPU component for modeling, failing to consider the differences in thermal lag between heterogeneous components such as GPU and high-speed storage, as well as the thermal coupling conduction effect between components. This makes them unsuitable for extreme power consumption prediction scenarios of current heterogeneous accelerated servers. Third, purely data-driven deep learning models lack physical constraints, making them prone to overfitting and predictions that violate thermodynamic principles. They also exhibit extremely poor generalization under extreme conditions such as sudden changes in the data center environment, hardware aging, and heat dissipation failures. Fourth, existing solutions mostly focus on single-machine power consumption estimation, failing to consider the spatial lag effect caused by heat radiation between servers in cluster scenarios. Furthermore, they lack data augmentation and robustness optimization for extreme conditions, making it difficult to meet the actual needs of data center cluster-level power distribution planning and security early warning.

[0004] The aforementioned shortcomings collectively result in low accuracy in predicting extreme power consumption, poor adaptability to extreme scenarios, and insufficient engineering feasibility of existing methods, making it impossible to provide reliable data support for the design and operation of data center infrastructure.

[0005] To address these issues, there is an urgent need for a method and system for predicting the maximum power consumption of servers. Summary of the Invention

[0006] To address the aforementioned issues, this application proposes a method and system for predicting the ultimate power consumption of servers. By explicitly constructing multi-component, multi-time-step lag features and combining them with a physically constrained deep learning model, high-precision prediction of ultimate power consumption is achieved, resolving the prediction bias problem caused by temperature lag. The specific details are as follows: On the one hand, this application proposes a method for predicting the maximum power consumption of a server, the specific contents of which include: S1. Obtain multi-source data and perform timestamp alignment to obtain the original time-series dataset; S2. Construct a hysteresis feature matrix based on the original time series dataset, analyze the thermal coupling-derived features, and perform supplementary encoding to obtain the time series feature matrix of the temperature hysteresis effect; S3. Construct a multi-branch fusion deep learning model, and perform model fitting and parameter optimization training on the time-series feature matrix based on the multi-branch fusion deep learning model to obtain the target limit power consumption prediction model. S4. Based on the real-time feature data, hardware configuration data and data center cabinet topology data of the target server and target cluster, the target limit power consumption prediction model is used to predict and perform multi-dimensional engineering margin correction to obtain the engineering-adapted single-machine and cluster two-level limit power consumption prediction values.

[0007] Preferably, the multi-source data mentioned in S1 is collected based on server hardware registers, monitoring interfaces, high-precision PDUs, data center environment sensors and hardware specification manuals, including five dimensions: hardware static characteristics, component-level load dynamic characteristics, full-dimensional temperature characteristics, heat dissipation and environmental characteristics, and power consumption true value characteristics.

[0008] Preferably, in S2, the specific content of constructing a hysteresis feature matrix based on the original time-series dataset, analyzing thermally coupled derived features, and performing supplementary encoding to obtain the time-series feature matrix of the temperature hysteresis effect includes: Obtain timing data for the CPU, GPU, memory controller, and NVMe SSD from the server hardware registers respectively; With a preset lag time gradient, the historical time series data and load information are used as the basic features of the current moment based on the lag time gradient, which fully covers the inherent thermal lag time interval of each component to obtain the lag feature matrix of multiple components and multiple time steps. Dynamic hysteresis derivative features and inter-component thermal coupling hysteresis features are calculated based on time-series data. The dynamic hysteresis derivative features include temperature change rate features, load-temperature coupling features, and heat dissipation efficiency features. The inter-component thermal coupling hysteresis characteristics include inter-component temperature difference characteristics, cross-component hysteresis characteristics, and chassis thermal accumulation characteristics; The dynamic hysteresis derivative features and the inter-component thermal coupling hysteresis features are added to the hysteresis feature matrix and encoded to obtain the time-series feature matrix of temperature hysteresis effect.

[0009] Preferably, the multi-branch fusion deep learning model is based on the LSTM-Attention framework and introduces an adaptive dynamic attention mechanism, a graph attention network with thermal coupling between components, thermodynamic physical constraints, and a multi-feature fusion mechanism, specifically including: Multi-component temporal feature encoding and dynamic attention module: A two-layer bidirectional LSTM network is used, and a multi-head temporal attention layer is connected after the LSTM. The power consumption prediction target at the current moment is used as the query vector Q, and the full-time LSTM output is used as the key vector K and value vector V. The feature weights of different historical time steps are automatically learned. The input is a temporal feature matrix for a period of time, and the output is the full-time dependency features. The Graph Attention (GAT) module for inter-component thermal coupling constructs an undirected hardware component topology graph, using CPU, GPU, memory, NVMe, and chassis environment as graph nodes and physical distance and heat conduction path between components as edge weights. A two-layer graph attention network is used, with the input being the temperature, load, and hysteresis characteristics of each component, and the output being the thermal coupling characteristics between components. The network automatically learns the weights of the impact of the heat generation of different components on the power consumption of the whole machine. Static hardware feature encoding module: It adopts a 2-layer fully connected network to encode static features such as CPU model, number of GPUs, and power efficiency, and outputs static hardware features; Multi-source feature fusion and basic power consumption output module: It splices and fuses time-dependent features, thermally coupled features, and static hardware features, connects them to a 2-layer fully connected network, and, with the help of ReLU activation function and dropout regularization, outputs the basic full-load power consumption prediction value fitted by the model. Thermodynamic physical information constraint layer: Fourier's law of heat conduction and Newton's law of cooling are combined to construct the differential equation of heat conduction, which is used as the regularization term of the model loss function to obtain the total loss function.

[0010] Preferably, the expression for the total loss function is: ; in, To predict losses, mean square error (MSE) is used to measure the deviation between the predicted and measured power consumption values. Physical canonical loss measures the deviation between the model output and thermodynamic laws. For L2 regularization terms, The physical regularization loss weight coefficients are as follows: The L2 regularization term loss weight coefficient; ; in, Let i be the temperature of the i-th component. Let be the heat transfer coefficient of the i-th component. Let i be the heat dissipation power of the i-th component. Let be the heat dissipation coefficient of the i-th component, be the heat dissipation coefficient of the i-th component, and t be the server runtime variable, representing the sampling time of temperature and power consumption timing data.

[0011] Preferably, the specific content of the target limit power consumption prediction model obtained by performing model fitting and parameter optimization training on the time-series feature matrix based on a multi-branch fusion deep learning model in S3 includes: The time series feature matrix is ​​divided into several samples according to time periods; The samples were divided into training set, validation set and test set in chronological order, and the sample size was increased by using a sliding window. The training samples under extreme conditions are supplemented through a dual approach: physical simulation augmentation and generative data augmentation. The specific content of the physical simulation enhancement includes: based on the server thermal simulation model, simulating extreme scenarios, generating corresponding temperature-power consumption time series data, and supplementing it to the training set; The specific content of the generative enhancement includes: using a temporal generative adversarial network to learn the data distribution of real extreme working conditions, generating extreme scene samples, and thus balancing the working condition distribution of the training set; A phased training strategy was adopted to train the model and perform multi-dimensional model validation to obtain the target limit power consumption prediction model.

[0012] Preferably, the specific content of the phased training strategy is as follows: Phase 1: Freeze the model's prediction head and train it using only physical regularization loss. This allows the model to learn the fundamental laws of thermodynamics and provides initial weights that conform to physical logic for subsequent training. Phase 2: Unfreeze the full model, perform end-to-end training using the total loss function, use AdamW as the optimizer, set the batch size to 32, and set the number of training epochs to 100. The third stage: The model is fine-tuned using extreme operating condition samples. In conjunction with the weighted loss function, the extreme operating condition samples are given a weight of 10 times, forcing the model to learn the temperature lag and power consumption rules of extreme scenarios. The phased training strategy is configured with an early stopping mechanism, which stops training when the validation set loss does not decrease for a preset number of consecutive rounds.

[0013] Preferably, the multi-dimensional engineering margin correction processing in S4 includes single-machine limit power consumption engineering correction and cluster-level spatiotemporal collaborative limit power consumption prediction. Specifically, the single-machine limit power consumption engineering correction involves adding a multi-dimensional engineering margin correction to the base full-load power consumption output by the model to obtain the final single-machine limit power consumption. The expression for the multi-dimensional engineering margin correction is as follows: ; in, This is the predicted base full-load power consumption value output by the model. This is the transient impact factor, with a value of 1.2, covering the power consumption peaks of short-term CPU / GPU turbo boost (PL3 / MaxBoost). This is a hardware aging margin factor, with a value of 1.12, covering the increase in power consumption caused by hardware aging and decreased heat dissipation efficiency during the server's 5-year service life. This is the environmental temperature rise compensation factor, with a value of 1.023 / 10℃, covering the impact of temperature fluctuations in the computer room environment. The full-load conversion efficiency of the power supply is set to 0.9, and power conversion losses are corrected. This represents the maximum power consumption of a single machine.

[0014] Preferably, the specific content of the cluster-level spatiotemporal collaborative limit power consumption prediction includes: For the central cluster scenario of the sample, each server in the rack is used as a node, and the rack location and adjacent distance are used as edge weights to construct a cluster spatial topology graph; Graph Convolutional Network (GCN) is used to learn the impact of thermal radiation from neighboring servers and rack thermal field distribution on the ambient temperature of local servers, thereby capturing the spatial temperature hysteresis effect at the cluster level. Combining the predicted maximum power consumption of a single machine, the spatial thermal coupling effect of the cluster, and the concurrent service load rate, the total maximum power consumption of the cluster is output. The expression for the total maximum power consumption of the cluster is: ; in, This represents the total maximum power consumption of the cluster. Let be the maximum single-machine power consumption value of the i-th server. The cluster thermal coupling correction coefficient for the i-th server is output by the GCN model. This refers to the concurrent service load rate.

[0015] On the other hand, this application proposes a server extreme power consumption prediction system, comprising: Data acquisition unit: Acquires data from multiple sources and performs timestamp alignment to obtain the original time-series dataset; Feature construction unit: Based on the original time series dataset, a hysteresis feature matrix is ​​constructed, and thermal coupling-derived features are analyzed and supplemented with encoding to obtain the time series feature matrix of temperature hysteresis effect; Model building unit: Construct a multi-branch fusion deep learning model, and perform model fitting and parameter optimization training on the time-series feature matrix based on the multi-branch fusion deep learning model to obtain the target limit power consumption prediction model; Prediction and Correction Unit: Based on the real-time feature data, hardware configuration data and data center cabinet topology data of the target server and target cluster, the target limit power consumption prediction model is used to make predictions and perform multi-dimensional engineering margin correction processing to obtain engineering-adapted single-machine and cluster two-level limit power consumption prediction values.

[0016] In summary, the server extreme power consumption prediction method and system of the present invention, compared with traditional technologies, improves the prediction accuracy of extreme conditions by explicitly constructing the basic lag characteristics of multi-component matching time, accurately quantifying the temperature lag effect, and combining a physical constraint deep learning model; it has a two-level adaptation from single machine to cluster, and can be directly used for data center power distribution planning and operation and maintenance scheduling, with strong engineering practicality and generalization.

[0017] The technical method of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0018] Figure 1 This invention provides a method and system for predicting the maximum power consumption of a server. (See diagram below) Figure 2 This is a unit diagram of a server limit power consumption prediction system according to the present invention. Detailed Implementation

[0019] The technical method of the present invention will be further described below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of this application.

[0020] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the scope of this application and its application or use.

[0021] Techniques, systems, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, they should be considered part of the instruction manual.

[0022] In all the examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0023] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0024] Example 1 This embodiment proposes a method for predicting the maximum power consumption of a server, such as... Figure 1 As shown, the specific content includes: S1. Obtain multi-source data and perform timestamp alignment to obtain the original time-series dataset.

[0025] Furthermore, the multi-source data mentioned in S1 is collected based on server hardware registers, monitoring interfaces, high-precision PDUs, data center environment sensors, and hardware specification manuals, including five dimensions: hardware static characteristics, component-level load dynamic characteristics, full-dimensional temperature characteristics, heat dissipation and environmental characteristics, and power consumption true value characteristics.

[0026] This phase breaks through the data limitations of the original single component, providing a complete data foundation for full-dimensional temperature hysteresis modeling. The data collection frequency is uniformly set at 1 time / second, and the collection period covers a full operating cycle of ≥14 days (including idle, peak business, full load, and extreme operating conditions). The collection dimensions are divided into 5 categories: Hardware static characteristics: CPU model / number of cores / PL2 / PL3 turbo boost power consumption, GPU model / quantity / maximum acceleration power consumption, memory capacity / number of modules / type, NVMeSSD quantity / model, power supply rated efficiency, chassis heat dissipation structure parameters, and physical distance between components.

[0027] Component-level load dynamic characteristics: CPU / GPU utilization, CPU core frequency / voltage, GPU computing power / memory utilization, memory bandwidth utilization, NVMe IOPS / bandwidth utilization.

[0028] Comprehensive temperature characteristics: CPU core temperatures, GPU memory / core temperatures, memory controller temperatures, NVMe controller temperatures, server intake / exhaust air temperatures, and ambient temperatures within the server room rack.

[0029] Heat dissipation and environmental characteristics: server fan speed / duty cycle, data center air conditioning set temperature, rack location, altitude, and mains input voltage.

[0030] True power consumption characteristics: overall input power consumption (PDU high-precision acquisition, accuracy ±0.5%), component-level power consumption (CPURAPL interface, GPUDCGM interface acquisition).

[0031] S2. Construct a hysteresis feature matrix based on the original time series dataset, analyze the thermal coupling-derived features, and perform supplementary encoding to obtain the time series feature matrix of temperature hysteresis effect.

[0032] Furthermore, in S2, a hysteresis feature matrix is ​​constructed based on the original time-series dataset. The thermal coupling-derived features are analyzed, and supplementary encoding is performed to obtain the specific content of the time-series feature matrix of the temperature hysteresis effect, including: Obtain timing data for the CPU, GPU, memory controller, and NVMeSSD from the server hardware registers respectively.

[0033] With a pre-set lag time gradient, the historical time series data and load information are used as the basic features of the current moment based on the lag time gradient, which fully covers the inherent thermal lag time interval of each component to obtain the lag feature matrix of multiple components and multiple time steps.

[0034] Dynamic hysteresis derivative features and inter-component thermal coupling hysteresis features are calculated based on time-series data. The dynamic hysteresis derivative features include temperature change rate features, load-temperature coupling features, and heat dissipation efficiency features.

[0035] The inter-component thermal coupling hysteresis characteristics include inter-component temperature difference characteristics, cross-component hysteresis characteristics, and chassis thermal accumulation characteristics.

[0036] The dynamic hysteresis derivative features and the inter-component thermal coupling hysteresis features are added to the hysteresis feature matrix and encoded to obtain the time-series feature matrix of temperature hysteresis effect.

[0037] Construction of dynamic hysteresis derived features: Temperature change rate characteristics: Calculate the temperature change value (°C / s) of each component per unit time, including the 10s moving average change rate and the 30s cumulative change rate, quantify the rate of heat accumulation / dissipation, and capture the dynamic trend of the hysteresis process.

[0038] Load-temperature coupling characteristics: Multiply the utilization rate of the same component and the same lag time by the temperature to generate coupling characteristics such as CPU utilization_lag 30s × CPU temperature_lag 30s, quantifying the synergistic lag effect of load and temperature on power consumption.

[0039] Heat dissipation efficiency characteristics: The temperature difference is calculated based on "component temperature - inlet air temperature" and combined with fan speed to generate heat dissipation efficiency characteristics, quantifying the impact of heat dissipation capacity on temperature lag time (the lower the heat dissipation efficiency, the longer the lag time).

[0040] Construction of inter-component thermal coupling hysteresis characteristics: Temperature difference characteristics between components: Calculate the real-time temperature difference between CPU and GPU, GPU and NVMe, core components and air intake, and capture the intensity of heat conduction between components.

[0041] Cross-component hysteresis features: Construct cross-features of GPU temperature hysteresis 60s and CPU temperature hysteresis 30s to quantify the hysteresis impact of a component’s historical heat generation on the subsequent temperature and power consumption of other components.

[0042] Chassis thermal accumulation characteristics: Based on the temperature difference between the exhaust and intake air, combined with historical load data, the thermal accumulation in the computer chassis is captured to identify the thermal hysteresis effect at the whole system level.

[0043] Adaptive lag coding preprocessing: Relative position coding is added to all temporal features to provide explicit information on the "temporal sequence" of the model, supporting the subsequent adaptive learning of the model to dynamic lag time and breaking through the rigid limitation of fixed lag time. At the same time, Min-Max normalization is performed on all features to map them to the [0,1] interval to eliminate dimensional differences, and missing values ​​caused by lag operations are filled in using "back-filling + forward-filling".

[0044] S3. Construct a multi-branch fusion deep learning model, and perform model fitting and parameter optimization training on the time-series feature matrix based on the multi-branch fusion deep learning model to obtain the target limit power consumption prediction model.

[0045] Furthermore, the multi-branch fusion deep learning model is based on the LSTM-Attention framework and introduces an adaptive dynamic attention mechanism, a graph attention network with thermal coupling between components, thermodynamic physical constraints, and a multi-feature fusion mechanism, specifically including: Multi-component temporal feature encoding and dynamic attention module (capturing long-term temporal dependencies of component utilization, temperature, and power consumption, adaptively learning the optimal lag time, and overcoming the limitations of fixed lag windows): A two-layer bidirectional LSTM network (128 hidden layer dimensions, dropout=0.2) is used as the input, which is a temporal feature window of 60~90s in length (containing all dynamic features and lag features). After the LSTM, a multi-head temporal attention layer (8 heads) is connected. The power consumption prediction target at the current moment is used as the query vector Q, and the full-time LSTM output is used as the key vector K and value vector V. The feature weights of different historical time steps are automatically learned. The input is a temporal feature matrix of a certain period of time, and the output is the full-time dependency features.

[0046] For example, under compute-intensive loads, the model automatically focuses on GPU temperature characteristics from 30 to 45 seconds ago. Under I / O-intensive loads, it automatically focuses on NVMe temperature characteristics from 10 to 20 seconds ago, achieving adaptive dynamic modeling of lag time.

[0047] The Component Thermal Coupling Graph Attention (GAT) module (captures the hysteresis effect of thermal conduction coupling between different components, solving the modeling gap of the collaborative impact of multiple components in heterogeneous servers): uses CPU, GPU, memory, NVMe, and chassis environment as graph nodes, and physical distance between components and thermal conduction path as edge weights to construct an undirected hardware component topology graph.

[0048] A two-layer graph attention network (GAT, number of attention heads = 4) is adopted. The input is the temperature, load and hysteresis features of each component, and the output is the thermal coupling features between components. The network automatically learns the weight of the impact of the heat generation of different components on the power consumption of the whole machine (e.g., when the GPU is under full load, the weight of its temperature hysteresis feature is significantly higher than that of the CPU).

[0049] Static hardware feature encoding module (extracts the basic impact of hardware specifications on power consumption and adapts to servers with different configurations): It adopts a 2-layer fully connected network (input dimension: 32-16) to encode static features such as CPU model, number of GPUs, and power efficiency, and outputs static hardware features.

[0050] Multi-source feature fusion and basic power consumption output module (deeply fuses time-dependent features, thermally coupled features, and static hardware features to output the basic full-load power consumption prediction value of the whole machine): It splices and fuses time-dependent features, thermally coupled features, and static hardware features, connects to a 2-layer fully connected network (dimension: 128-32-1), and, with the help of ReLU activation function and dropout regularization, outputs the basic full-load power consumption prediction value fitted by the model.

[0051] Thermodynamic physical information constraint layer (integrating the physical laws of heat conduction into the model, solving the problems of poor generalization and violation of physical laws in pure data-driven models, and making the modeling of temperature hysteresis effect conform to the thermodynamic essence): combining Fourier's law of heat conduction with Newton's law of cooling to construct the differential equation of heat conduction, which is used as the regularization term of the model loss function to obtain the total loss function.

[0052] Furthermore, the expression for the total loss function is: .

[0053] in, To predict losses, mean square error (MSE) is used to measure the deviation between the predicted and measured power consumption values. Physical canonical loss measures the deviation between the model output and thermodynamic laws. For L2 regularization terms, The physical regularization loss weight coefficients are as follows: The loss weight coefficient for the L2 regularization term.

[0054] .

[0055] in, Let i be the temperature of the i-th component. Let be the heat transfer coefficient of the i-th component. Let i be the heat dissipation power of the i-th component. Let be the heat dissipation coefficient of the i-th component, be the heat dissipation coefficient of the i-th component, and t be the server runtime variable, representing the sampling time of temperature and power consumption timing data.

[0056] Furthermore, in S3, the specific content of the target limit power consumption prediction model obtained by performing model fitting and parameter optimization training on the time-series feature matrix based on a multi-branch fusion deep learning model includes: The time series feature matrix is ​​divided into several samples according to time periods.

[0057] The samples were divided into training, validation, and test sets in chronological order, and the sample size was increased using a sliding window.

[0058] The training samples under extreme conditions are supplemented through a dual approach: physical simulation augmentation and generative data augmentation. The specific content of the physical simulation enhancement includes: simulating extreme scenarios based on the server thermal simulation model, generating corresponding temperature-power consumption time series data, and supplementing it to the training set.

[0059] The specific content of the generative enhancement includes: using a temporal generative adversarial network to learn the data distribution of real extreme working conditions, generating extreme scenario samples, and thus balancing the working condition distribution of the training set.

[0060] A phased training strategy was adopted to train the model and perform multi-dimensional model validation to obtain the target limit power consumption prediction model.

[0061] Furthermore, the specific details of the phased training strategy are as follows: Phase 1 (Physical Pre-training): Freeze the model's prediction head and train it using only physical regularization loss. This allows the model to learn the fundamental laws of thermodynamics and provides initial weights that conform to physical logic for subsequent training.

[0062] Phase 2 (Full Parameter Training): Unfreeze the entire model and perform end-to-end training using the total loss function. The optimizer used is AdamW (learning rate 0.001, weight decay 1e-4), the batch size is set to 32, and the number of training epochs is set to 100.

[0063] The third stage (fine-tuning and optimization): The model is fine-tuned using extreme operating condition samples. With the weighted loss function, the extreme operating condition samples are given a weight of 10 times, forcing the model to learn the temperature lag and power consumption rules of extreme scenarios.

[0064] The phased training strategy is configured with an early stopping mechanism. The early stopping mechanism stops training when the validation set loss does not decrease for a preset number of consecutive rounds, and saves the optimal model weights with the lowest validation set loss to avoid overfitting.

[0065] Multi-dimensional model validation: Breaking through the limitations of simply validating conventional accuracy, a full-scenario validation system is constructed. Routine accuracy verification: Calculate the mean absolute percentage error (MAPE) and root mean square error (RMSE). The requirement is that MAPE ≤ 2% under normal operating conditions.

[0066] Extreme scenario verification: Trigger extreme conditions such as full load of CPU+GPU+memory+NVMe, fan half-speed failure, and 45℃ high temperature environment, and compare the predicted value with the measured value of high precision power analyzer. The requirement is that MAPE ≤ 4% in extreme scenarios.

[0067] Physical consistency verification: Verify the deviation between the physical parameters such as thermal resistance and heat dissipation coefficient learned by the model and the values ​​in the hardware manual. The deviation should be ≤5% to ensure that the model prediction conforms to physical laws.

[0068] Generalization validation: Test the model using data from other server models that were not used in the training, requiring cross-model MAPE ≤ 5%, to validate the model's generalization ability.

[0069] S4. Based on the real-time feature data, hardware configuration data and data center cabinet topology data of the target server and target cluster, the target limit power consumption prediction model is used to predict and perform multi-dimensional engineering margin correction to obtain the engineering-adapted single-machine and cluster two-level limit power consumption prediction values.

[0070] Furthermore, the multi-dimensional engineering margin correction process in S4 includes single-machine limit power consumption engineering correction and cluster-level spatiotemporal collaborative limit power consumption prediction.

[0071] Specifically, the single-machine limit power consumption engineering correction involves adding a multi-dimensional engineering margin correction to the base full-load power consumption output by the model to obtain the final single-machine limit power consumption. The expression for the multi-dimensional engineering margin correction is as follows: .

[0072] in, This is the predicted base full-load power consumption value output by the model. This is the transient impact factor, with a value of 1.2, covering the power consumption peaks of short-term CPU / GPU turbo boost (PL3 / MaxBoost). This is a hardware aging margin factor, with a value of 1.12, covering the increase in power consumption caused by hardware aging and decreased heat dissipation efficiency during the server's 5-year service life. This is the environmental temperature rise compensation factor, with a value of 1.023 / 10℃, covering the impact of temperature fluctuations in the computer room environment. The full-load conversion efficiency of the power supply is set to 0.9, and power conversion losses are corrected. This represents the maximum power consumption of a single machine.

[0073] Furthermore, the specific content of the cluster-level spatiotemporal collaborative limit power consumption prediction includes: For the central cluster scenario of the sample, each server in the rack is used as a node, and the rack location and adjacent distance are used as edge weights to construct a cluster spatial topology graph.

[0074] A graph convolutional network (GCN) is used to learn the impact of thermal radiation from neighboring servers and the distribution of heat fields in the rack on the ambient temperature of local servers, thereby capturing the spatial temperature lag effect at the cluster level.

[0075] Combining the predicted maximum power consumption of a single machine, the spatial thermal coupling effect of the cluster, and the concurrent service load rate, the total maximum power consumption of the cluster is output. The expression for the total maximum power consumption of the cluster is: .

[0076] in, This represents the total maximum power consumption of the cluster. Let be the maximum single-machine power consumption value of the i-th server. The cluster thermal coupling correction coefficient for the i-th server is output by the GCN model. This refers to the concurrent service load rate.

[0077] Example 2 Step S1: Synchronous acquisition of multi-source data.

[0078] Data was collected from the target single-machine server, including: static hardware specifications such as Intel Xeon CPU model, number of GPUs, NVMe drive bays, and power efficiency; component-level load data such as CPU utilization, GPU computing power utilization, and NVMe IOPS; comprehensive temperature data such as CPU core temperature, GPU core temperature, and NVMe controller temperature; thermal environment data such as intake air temperature and fan speed; and measured power consumption values ​​were collected using a high-precision PDU. A uniform 1Hz sampling frequency was used, and millisecond-level timestamp alignment was achieved. 14 days of full-condition data were collected to obtain the raw time-series dataset.

[0079] Step S2: Construction and preprocessing of hysteresis features.

[0080] Hysteresis feature construction is performed on the raw timing data: CPU core temperature timing data is shifted forward by 10 time steps to generate CPU core temperature hysteresis 10s feature; GPU core temperature timing data is shifted forward by 30 time steps to generate GPU core temperature hysteresis 30s feature; NVMe temperature timing data is shifted forward by 60 time steps to generate NVMe temperature hysteresis 60s feature; CPU utilization timing data is shifted forward by 30 time steps to generate CPU utilization hysteresis 30s feature, fully covering the inherent thermal hysteresis range of each component.

[0081] Further processing of temperature change rate, CPU utilization-temperature coupling features, and CPU-GPU temperature difference features; performing relative position encoding and Min-Max normalization to the [0,1] interval on all features; using forward and backward padding to fill in missing values ​​caused by hysteresis shift to obtain a standardized time series feature matrix.

[0082] Step S3: Sample construction and data augmentation.

[0083] A 90s sliding window and a 1s step size were used to divide the time-series samples, with the measured power consumption at the end of the window as the label. The samples were divided into a 75% training set, a 15% validation set, and a 10% test set according to time sequence. Extreme scenario samples were generated by simulating fan failure and a 45℃ high-temperature environment through thermal simulation and combined with time-series GAN to supplement the training set and balance the sample distribution.

[0084] Step S4: Model training and optimization.

[0085] A multi-branch deep learning model is constructed: a bidirectional LSTM layer captures temporal dependencies, a multi-head temporal attention layer adaptively focuses on key lagging features, a GAT layer learns the thermal coupling relationships of multiple components, and a fully connected layer fuses static hardware features. A heat conduction differential equation is introduced as a physical regularization term to construct the total loss function. A three-stage strategy of "physical pre-training - full parameter training - extreme sample fine-tuning" is adopted, and the model is trained for 100 epochs using the AdamW optimizer, with early stopping to save the optimal model.

[0086] Step S5: Limit power consumption inference and correction.

[0087] The model is input with real-time standardized features and outputs the basic full-load power consumption. The model is then superimposed with 1.2 times transient impact factor, 1.12 times aging margin, and 1.023 times temperature rise compensation. The result is divided by 0.9 power efficiency to obtain the predicted value of the single-machine limit power consumption. The MAPE is ≤2% under normal operating conditions.

[0088] Example 3 Based on Example 2, rack topology and adjacent server location data are collected. The cluster spatial thermal coupling hysteresis effect is modeled through GCN network. Combined with a load simulacivity rate of 0.9, the total limit power consumption of the cluster is calculated. Attention weight analysis results are also output, and the GPU core temperature hysteresis of 30s is identified as the core influencing feature of the limit power consumption. Load scheduling suggestions are also provided.

[0089] Example 4 This embodiment proposes a server extreme power consumption prediction system, such as... Figure 2 As shown, it includes: Data acquisition unit: Acquires data from multiple sources and performs timestamp alignment to obtain the original time-series dataset.

[0090] Feature construction unit: Based on the original time series dataset, a hysteresis feature matrix is ​​constructed, and thermal coupling-derived features are analyzed and supplemented with encoding to obtain the time series feature matrix of temperature hysteresis effect.

[0091] Model building unit: Construct a multi-branch fusion deep learning model, and perform model fitting and parameter optimization training on the time-series feature matrix based on the multi-branch fusion deep learning model to obtain the target limit power consumption prediction model.

[0092] Prediction and Correction Unit: Based on the real-time feature data, hardware configuration data and data center cabinet topology data of the target server and target cluster, the target limit power consumption prediction model is used to make predictions and perform multi-dimensional engineering margin correction processing to obtain engineering-adapted single-machine and cluster two-level limit power consumption prediction values.

[0093] Finally, it should be noted that the above embodiments are only used to illustrate the technical methods of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical methods of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical methods to deviate from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for predicting the ultimate power consumption of a server, characterized in that, Includes the following steps: S1. Obtain multi-source data and perform timestamp alignment to obtain the original time-series dataset; S2. Construct a hysteresis feature matrix based on the original time series dataset, analyze the thermal coupling-derived features, and perform supplementary encoding to obtain the time series feature matrix of the temperature hysteresis effect; S3. Construct a multi-branch fusion deep learning model, and perform model fitting and parameter optimization training on the time-series feature matrix based on the multi-branch fusion deep learning model to obtain the target limit power consumption prediction model. S4. Based on the real-time feature data, hardware configuration data and data center cabinet topology data of the target server and target cluster, the target limit power consumption prediction model is used to predict and perform multi-dimensional engineering margin correction to obtain the engineering-adapted single-machine and cluster two-level limit power consumption prediction values.

2. The server limit power consumption prediction method according to claim 1, characterized in that, The multi-source data mentioned in S1 is collected based on server hardware registers, monitoring interfaces, high-precision PDUs, data center environmental sensors, and hardware specification manuals, including five dimensions: hardware static characteristics, component-level load dynamic characteristics, full-dimensional temperature characteristics, heat dissipation and environmental characteristics, and power consumption true value characteristics.

3. The server limit power consumption prediction method according to claim 2, characterized in that, The specific content of the time-series feature matrix of temperature hysteresis effect obtained by constructing a hysteresis feature matrix based on the original time-series dataset in S2, analyzing thermally coupled derived features, and performing supplementary encoding includes: Obtain timing data for the CPU, GPU, memory controller, and NVMe SSD from the server hardware registers respectively; With a preset lag time gradient, the historical time series data and load information are used as the basic features of the current moment based on the lag time gradient, which fully covers the inherent thermal lag time interval of each component to obtain the lag feature matrix of multiple components and multiple time steps. Dynamic hysteresis derivative features and inter-component thermal coupling hysteresis features are calculated based on time-series data. The dynamic hysteresis derivative features include temperature change rate features, load-temperature coupling features, and heat dissipation efficiency features. The inter-component thermal coupling hysteresis characteristics include inter-component temperature difference characteristics, cross-component hysteresis characteristics, and chassis thermal accumulation characteristics; The dynamic hysteresis derivative features and the inter-component thermal coupling hysteresis features are added to the hysteresis feature matrix and encoded to obtain the time-series feature matrix of temperature hysteresis effect.

4. The server limit power consumption prediction method according to claim 3, characterized in that, The multi-branch fusion deep learning model is based on the LSTM-Attention framework and introduces an adaptive dynamic attention mechanism, a graph attention network with thermal coupling between components, thermodynamic physical constraints, and a multi-feature fusion mechanism, specifically including: Multi-component temporal feature encoding and dynamic attention module: A two-layer bidirectional LSTM network is used, and a multi-head temporal attention layer is connected after the LSTM. The power consumption prediction target at the current moment is used as the query vector Q, and the full-time LSTM output is used as the key vector K and value vector V. The feature weights of different historical time steps are automatically learned. The input is the temporal feature matrix, and the output is the full-time dependency feature. The Graph Attention (GAT) module for inter-component thermal coupling constructs an undirected hardware component topology graph, using CPU, GPU, memory, NVMe, and chassis environment as graph nodes and physical distance and heat conduction path between components as edge weights. A two-layer graph attention network is used, with the input being the temperature, load, and hysteresis characteristics of each component, and the output being the thermal coupling characteristics between components. The network automatically learns the weights of the impact of the heat generation of different components on the power consumption of the whole machine. Static hardware feature encoding module: It adopts a 2-layer fully connected network to encode static features such as CPU model, number of GPUs, and power efficiency, and outputs static hardware features; Multi-source feature fusion and basic power consumption output module: It splices and fuses time-dependent features, thermally coupled features, and static hardware features, connects them to a 2-layer fully connected network, and, with the help of ReLU activation function and dropout regularization, outputs the basic full-load power consumption prediction value fitted by the model. Thermodynamic physical information constraint layer: Fourier's law of heat conduction and Newton's law of cooling are combined to construct the differential equation of heat conduction, which is used as the regularization term of the model loss function to obtain the total loss function.

5. The server limit power consumption prediction method according to claim 4, characterized in that, The expression for the total loss function is: ； in, To predict losses, mean square error (MSE) is used to measure the deviation between the predicted and measured power consumption values. Physical canonical loss measures the deviation between the model output and thermodynamic laws. For L2 regularization terms, The physical regularization loss weight coefficients are used. The L2 regularization term loss weight coefficient; ； in, Let i be the temperature of the i-th component. Let be the heat transfer coefficient of the i-th component. Let i be the heat dissipation power of the i-th component. Let be the heat dissipation coefficient of the i-th component, be the heat dissipation coefficient of the i-th component, and t be the server runtime variable, representing the sampling time of temperature and power consumption timing data.

6. The server limit power consumption prediction method according to claim 5, characterized in that, S3 uses a multi-branch fusion deep learning model to perform model fitting and parameter optimization training on the time-series feature matrix, resulting in the target limit power consumption prediction model. The specific content of this model includes: The time series feature matrix is ​​divided into several samples according to time periods; The samples were divided into training set, validation set and test set in chronological order, and the sample size was increased by using a sliding window. The training samples under extreme conditions are supplemented through a dual approach: physical simulation enhancement and generative data augmentation. The specific content of the physical simulation enhancement includes: based on the server thermal simulation model, simulating extreme scenarios, generating corresponding temperature-power consumption time series data, and supplementing it to the training set; The specific content of the generative enhancement includes: using a temporal generative adversarial network to learn the data distribution of real extreme working conditions, generating extreme scene samples, and thus balancing the working condition distribution of the training set; A phased training strategy was adopted to train the model and perform multi-dimensional model validation to obtain the target limit power consumption prediction model.

7. The server limit power consumption prediction method according to claim 1, characterized in that, The specific details of the phased training strategy are as follows: Phase 1: Freeze the model's prediction head and train it using only physical regularization loss. This allows the model to learn the fundamental laws of thermodynamics first, providing initial weights that conform to physical logic for subsequent training. Phase 2: Unfreeze the full model, perform end-to-end training using the total loss function, use AdamW as the optimizer, set the batch size to 32, and set the number of training epochs to 100. The third stage: The model is fine-tuned using extreme operating condition samples. In conjunction with the weighted loss function, the extreme operating condition samples are given a weight of 10 times, forcing the model to learn the temperature lag and power consumption rules of extreme scenarios. The phased training strategy is configured with an early stopping mechanism, which stops training when the validation set loss does not decrease for a preset number of consecutive rounds.

8. The server limit power consumption prediction method according to claim 1, characterized in that, S4's multi-dimensional engineering margin correction process includes single-machine limit power consumption engineering correction and cluster-level spatiotemporal collaborative limit power consumption prediction. Specifically, the single-machine limit power consumption engineering correction involves adding a multi-dimensional engineering margin correction to the base full-load power consumption output by the model to obtain the final single-machine limit power consumption. The expression for the multi-dimensional engineering margin correction is as follows: ； in, This is the predicted base full-load power consumption value output by the model. Transient impact factor This is a hardware aging margin factor. As a factor to compensate for environmental temperature rise, For full-load power conversion efficiency, This represents the maximum power consumption of a single machine.

9. The server limit power consumption prediction method according to claim 1, characterized in that, The specific content of the cluster-level spatiotemporal collaborative limit power consumption prediction includes: For the central cluster scenario of the sample, each server in the rack is used as a node, and the rack location and adjacent distance are used as edge weights to construct a cluster spatial topology graph; Graph Convolutional Network (GCN) is used to learn the impact of thermal radiation from neighboring servers and rack thermal field distribution on the ambient temperature of local servers, thereby capturing the spatial temperature hysteresis effect at the cluster level. Combining the predicted maximum power consumption of a single machine, the spatial thermal coupling effect of the cluster, and the concurrent service load rate, the total maximum power consumption of the cluster is output. The expression for the total maximum power consumption of the cluster is: ； in, This represents the total maximum power consumption of the cluster. Let be the maximum single-machine power consumption value of the i-th server. The cluster thermal coupling correction coefficient for the i-th server is output by the GCN model. This refers to the concurrent service load rate.

10. A server extreme power consumption prediction system, characterized in that, include: Data acquisition unit: Acquires data from multiple sources and performs timestamp alignment to obtain the original time-series dataset; Feature construction unit: Based on the original time series dataset, a hysteresis feature matrix is ​​constructed, and thermal coupling-derived features are analyzed and supplemented with encoding to obtain the time series feature matrix of temperature hysteresis effect; Model building unit: Construct a multi-branch fusion deep learning model, and perform model fitting and parameter optimization training on the time-series feature matrix based on the multi-branch fusion deep learning model to obtain the target limit power consumption prediction model; Prediction and Correction Unit: Based on the real-time feature data, hardware configuration data and data center cabinet topology data of the target server and target cluster, the target limit power consumption prediction model is used to make predictions and perform multi-dimensional engineering margin correction processing to obtain engineering-adapted single-machine and cluster two-level limit power consumption prediction values.

Images