Energy efficiency optimization control method for heterogeneous computing power cluster

Abstract

The invention discloses an energy efficiency optimization control method for a heterogeneous computing power cluster, and particularly relates to the technical field of energy efficiency optimization control. According to the method, a unified time reference with a high-precision energy consumption anchor point as a reference is constructed, and residual error construction is carried out on power consumption data of a multi-source sensor; and a Bayesian online change point detection mechanism based on risk scoring is introduced to identify sensor drift. The system further adopts a multi-sensor Kalman filtering algorithm to carry out joint estimation on a real power consumption state and a channel bias state, and the fusion precision is improved through bias backtracking correction and observation noise self-adaptive adjustment. In order to improve the adaptive capacity to cross-source interference and structural change, the system constructs a dynamic graph structure in a sliding window, topology-time joint modeling and self-supervised correction are performed in combination with a graph convolutional neural network, and fine-grained adjustment and uncertainty quantification of observation values of each channel are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of energy efficiency optimization control, and more particularly, to an energy efficiency optimization control method for heterogeneous computing clusters. BACKGROUND

[0002] With the wide deployment of intelligent video monitoring, edge computing, deep learning inference and other high-density computing tasks, computing clusters have become the core computing support platform in smart parks, industrial sites and critical infrastructure. These clusters are usually composed of multiple heterogeneous computing units such as CPU, GPU and FPGA, and their running states are affected by various factors, including environmental temperature and humidity, task load mutation, power supply fluctuation, etc. How to carry out fine energy consumption management and control while ensuring performance has become an important direction for energy efficiency optimization of computing systems.

[0003] Existing energy consumption optimization schemes mainly rely on static configuration strategies or coarse-grained power monitoring methods, and lack the ability to dynamically identify and compensate for non-ideal factors such as fine-grained sensor drift, time synchronization error and data quality fluctuation. For example, some schemes cannot distinguish between the timing characteristics of "sensor zero offset" and "actual load change", which can easily lead to distorted energy consumption estimates. Some methods do not introduce dynamic decision-making mechanisms, which can not exclude abnormal channels in time and affect the fusion accuracy. In addition, the redundancy and graph structure relationship across sensors are not effectively utilized, resulting in low processing efficiency of high-dimensional heterogeneous information.

[0004] Therefore, there is an urgent need for an energy efficiency optimization control method that can fuse multi-source heterogeneous power consumption data, have spatio-temporal correction capability and support robust decision-making in high dynamic environments to improve the system's estimation accuracy and control ability of the actual power consumption state. SUMMARY

[0005] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present application provide an energy efficiency optimization control method for heterogeneous computing clusters to solve the problems raised in the background art.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical solutions: The energy efficiency optimization control method for heterogeneous computing clusters comprises the following steps: First, set the energy consumption anchor point as the unified time reference, synchronously collect the measurement values of the multi-channel power consumption sensors and environmental sensors, and convert the measurement values of the multi-channel power consumption sensors and environmental sensors into residual sequences with the energy consumption anchor point as the reference; the energy consumption anchor point is an artificially selected absolute reference point with the most accurate measurement and the least interference; Then, a static zero-point displacement index, a noise degradation index and an anchor correlation attenuation index are calculated for each residual sequence, and a risk score is obtained by linear combination according to a preset weight; the risk score is compared with a higher threshold value and a lower threshold value, when the risk score exceeds the higher threshold value, a Bayesian online change point detection model is activated to output a drift probability and a bias estimate; when the risk score is between the lower threshold value and the higher threshold value, the Bayesian online change point detection model is activated to output the drift probability and the bias estimate as long as one of the following two conditions occurs: the risk score is continuously integrated to the threshold value D or the residual sequence jumps out of the steady-state three times of the standard deviation at one time. Finally, the drift probability and the bias estimate are input into a multi-sensor Kalman filter together with the corresponding sensor observation values to jointly estimate the real power consumption state and the channel bias state, and to perform bias backtracking correction and observation noise covariance adaptive adjustment when an anomaly is detected; subsequently, a dynamic graph structure is constructed by taking the sensor channels as nodes and the mutual information and the time delay correlation coefficient as edge weights, and a self-supervised graph convolutional neural network is input to obtain a correction vector and an uncertainty index for each observation value, and the correction vector and the uncertainty index for each observation value are fed back to the Kalman filter to update the fusion weight.

[0007] In a preferred embodiment, the static zero-point displacement index is obtained by: within a fixed rolling sliding window, calculating the window average of the residual sequence; then calling the historical data of the cold start stage, extracting the residual mean and residual standard deviation during the steady state as a reference benchmark, and normalizing and comparing the window average with the reference benchmark to obtain the static zero-point displacement index.

[0008] In a preferred embodiment, the noise degradation index is obtained by: within a fixed rolling sliding window, calculating the sample variance of the residual sequence; then calling the standard deviation of the steady-state residual in the cold start stage as a historical noise benchmark, and normalizing and comparing the sample variance with the historical noise benchmark to obtain the noise degradation index.

[0009] In a preferred embodiment, the anchor correlation attenuation index is obtained by: within a fixed rolling sliding window, performing mean removal operation on the residual sequence and the reference power sequence of the energy anchor to form a centralized sequence; on this basis, the Pearson correlation coefficient between the two centralized sequences is calculated in real time; then the Pearson correlation coefficient is distance mapped with the full correlation state to obtain the anchor correlation attenuation index; if an anti-correlation trend occurs, the anchor correlation attenuation index is forcibly set to the maximum value.

[0010] In a preferred embodiment, the risk score of each residual sequence is filtered by a time integrator before performing Bayesian online change point detection; when the risk score is below a lower threshold, it is directly ignored; when it is above an upper threshold, the Bayesian online change point detection model is immediately activated; when it is between the two thresholds, the time integration method is used, and if the integration result reaches the threshold D or the residual jumps out of the three times of the standard deviation of the steady state, the Bayesian online change point detection model is immediately activated.

[0011] In a preferred embodiment, the calculation process of the threshold D is as follows: first, maintain the risk score sequence in a sliding time window, and calculate the median P50 and the high quantile P90 in real time; take P50 as the reference risk level R_ref, and calculate the high side fluctuation amplitude ΔR = P90-P50; take R_ref and ΔR as inputs to calculate the allowed risk integration duration T_ref, and the formula is as follows: ; wherein is an empirical coefficient, is the reciprocal of the sampling frequency; the threshold D is defined as: .

[0012] In a preferred embodiment, the Bayesian online change point detection model regards the non-drift steady state as a random process subject to normal-inverse gamma prior, and adopts a constant hazard function , , which represents the prior knowledge of the system on the average interval of drift , to calculate the paragraph continuation and change point probability, and then output the drift probability and bias estimation.

[0013] In a preferred embodiment, the state vector of the multi-sensor Kalman filter includes the real power consumption state and the zero point bias state of each channel.

[0014] In a preferred embodiment, the real power consumption state is described by a first-order autoregressive model, and the zero point bias state of each channel is described by a random walk model.

[0015] In a preferred embodiment, whether the residual sequence jumps out of the three times of the standard deviation of the steady state at one time refers to that the deviation of the residual value of a sensor relative to the steady state average value determined in the cold start stage of the sensor exceeds three times of the steady state standard deviation.

[0016] Technical effects and advantages of the present application: The application first establishes a unified time reference with a higher metering accuracy "energy anchor point" in the power consumption monitoring link of the heterogeneous computing cluster, and distributes millisecond-level time stamps to all power consumption and environment sensors by means of PTP-to-NTP transparent clock, so as to ensure the alignment of multi-source data at the collection side, and fundamentally eliminate the power consumption superposition error caused by time drift in the traditional scheme. After the data enters the processing pipeline, the system uses the three-dimensional indicators of "static zero point displacement, noise degradation and anchor point correlation attenuation" to quantitatively evaluate the risk of each residual sequence in real time, and activates the Bayesian online change point detection model only for high-risk or continuous abnormal channels through the "double threshold-integration" scheduler, so as to complete the drift capture at the millisecond level while avoiding the redundant calculation burden of low-risk channels. When the risk score is in the warning interval, the integrator will generate a dynamic threshold D according to the background fluctuation amplitude and sampling frequency, which can suppress false positives and maintain high sensitivity. For the scene where the number of residuals reaches hundreds of levels, the introduced "risk-driven screening" logic can avoid blind instantiation of the detection model, and the overall detection efficiency and fusion stability are thus significantly improved.

[0017] The drift probability and bias estimate output in the detection stage are injected into the multi-sensor Kalman filter, the filter jointly estimates the real power consumption state and the zero point bias of each channel, and automatically adjusts the observation noise and backtracking correction strategy under high load fluctuation or extreme environment, ensuring the continuity, unbiasedness and high accuracy of the power consumption curve. Subsequently, the system constructs a dynamic graph with mutual information and time delay correlation coefficient as edge weight, and outputs a zero point correction vector and normalized uncertainty through a self-supervised graph convolution network, and then feeds back to the filter to realize weight self-adaptation, achieving the purpose of cross-channel redundancy complementation and multi-scale noise suppression. Through the above-mentioned cooperative mechanism, the application can significantly reduce the computing resource consumption on the monitoring side without changing the existing hardware, improve the power consumption estimation accuracy and robustness, and further provide more reliable real-time input for energy consumption scheduling algorithm, which has high portability, low incremental cost and excellent industrial promotion value. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to facilitate the understanding of those skilled in the art, the application will be further described below in conjunction with the drawings; Figure 1 The flowchart of the energy efficiency optimization control method for the heterogeneous computing cluster of the application; Figure 2 The residual sequence risk grading and Bayesian online change point detection model activation mechanism flowchart. DETAILED DESCRIPTION

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1:

[0021] This invention provides an energy efficiency optimization and control method for heterogeneous computing power clusters, such as... Figure 1 As shown, it includes the following steps: Step 1: High-precision anchor points and multi-source synchronous acquisition First, select a busbar or high-load branch that is representative of the power consumption of the entire busbar and has the lowest electrical noise, and install an online power metering device with a high accuracy level. Define it as a globally unique "energy consumption anchor point". The "energy consumption anchor point" here is understood as an "absolute reference point" that is artificially selected in the entire computer room power distribution network, which is the most accurate in metering and the least affected by interference. The output of the "energy consumption anchor point" is used as a reference power sequence.

[0022] The anchor device itself guarantees that the measurement error is at an extremely low level (e.g., within 0.2%) with a power calibration traceability certificate, and distributes the PPS (short for "Pulse-Per-Second") signal to the cluster management network through a unique PTP-to-NTP (PTP-to-NTP refers to a time synchronization bridging / conversion mechanism) transparent clock module. This ensures that all local power consumption sensors (onboard BMC power meter, CPURAPL counter, GPUNVML sampling) and environmental probes (rack temperature and humidity, air pressure, cooling water inlet temperature, etc.) have millisecond-level consistent timestamps when sampling.

[0023] At each sampling instant, the sensor encapsulates the raw reading along with a unified timestamp derived from the anchor point into an indexed data frame, which is then pushed to the streaming engine via the acquisition bus. (A streaming engine is a computing framework or system specifically designed for processing real-time data streams; unlike the traditional "batch processing" mode, it does not wait for all data to arrive before performing unified calculations, but instead performs calculations and processing immediately upon data arrival.) The streaming engine first writes the energy consumption anchor point readings into a dedicated column family, then instantly calculates the difference between each sensor reading and the energy consumption anchor point reading to obtain a residual sequence, which is written to a fixed-length, ring-shaped shared memory buffer using a zero-copy method. The buffer employs a monotonically increasing timestamp index and a "write-then-cut" overwrite strategy to store a multi-hour-level historical window within a limited memory footprint, providing high-frequency random read access capabilities for subsequent algorithms in the system.

[0024] Throughout the process, the data frame carries metadata such as sensor ID, physical location, sampling frequency, and firmware calibration version, enabling subsequent modules to accurately track data lineage and hardware status.

[0025] Step 2: Bayesian online change point detection to identify drift Without altering the validated "residual-first detection" backbone, to avoid blindly instantiating all Bayesian online change point detection models when the number of residual channels reaches hundreds, thus wasting computational resources and reducing detection accuracy, this invention introduces a "risk-driven screening" logic. This logic activates the Bayesian online change point detection model (BOCPD) only for high-risk residual channels, thereby improving overall detection efficiency and enhancing the fusion stability of the Kalman-graph convolutional network. Figure 2 As shown: Within a fixed rolling sliding window (e.g., 60 seconds), a three-dimensional dimensionless index is calculated for each residual sequence; the static zero-point displacement index is the ratio of the window mean deviation to the steady-state standard deviation, used to assess static zero-point displacement; the noise degradation index is the ratio of the window variance to the steady-state variance, used to assess noise degradation; and the anchor correlation decay index is used to assess cross-source consistency breakdown. The process of obtaining the static zero-point displacement index is as follows: First, calculate the average value of the residual sequence within a fixed rolling sliding window, using the following formula: ;in Representative Channel The mean residual within the current sliding time window reflects the average offset of the sensor reading relative to the global energy consumption anchor point; To fix the number of frames for the scrolling sliding window; The frame number representing the start time of the current window indicates the starting point of the fixed scrolling sliding window on the timeline, and the entire time period covered by the window is... ; Indicates channel At any moment The residual value.

[0026] Subsequently, the system accesses channels from historical data during the cold start phase (e.g., the first 24 hours). steady-state residual mean With steady-state standard deviation As a reference benchmark, the static zero-point displacement index is calculated using the following formula: ; This represents the static zero-point displacement index. The channel in the historical data during the cold start phase... steady-state residual mean With steady-state standard deviation The calculation formula is as follows: ; ; Total number of historical data samples used to estimate the steady-state mean and steady-state standard deviation for the cold-start phase.

[0027] The sample variance is calculated for the residual sequence within the fixed rolling sliding window, as follows: The noise degradation index is calculated as follows: ; wherein represents the noise degradation index.

[0028] To evaluate the degree of dynamic synchronization between the residual behavior of each observation channel and the energy anchor point, the system performs real-time calculation of the linear correlation between the residual sequence and the reference power sequence. First, within the current fixed rolling sliding window, the mean removal operation is performed on the residual sequence and the reference power sequence of the channel , respectively, to construct the centralized sequence: ; and represent the mean of the residual sequence and the reference power sequence within the current window, respectively. The reference power sequence refers to the global power consumption reference data stream collected by the energy anchor point device, which has high accuracy, high timeliness, and continuity. It reflects the real power change trend of a representative bus or main branch in the entire power distribution system, and is the basis for the key modules of residual construction, drift detection, synchronization alignment, and dynamic fusion.

[0029] Subsequently, the Pearson correlation coefficient is calculated based on the above-mentioned centralized sequence, which is used to measure the linear synchronization degree between the channel residual and the anchor point fluctuation. The calculation formula is as follows: To map this correlation to a quantitative consistency attenuation factor, the anchor correlation attenuation index is defined as: ; wherein represents the anchor correlation attenuation index. If an anti-correlation occurs (i.e. ), the system considers that there is a directional rupture between the channel and the reference anchor. To avoid introducing false trend signals, the value of is forcibly set, indicating complete loss of consistency, which will greatly suppress the weight of the channel in the fusion algorithm or trigger an abnormal marker.

[0030] The above three-dimensional dimensionless index forms a risk score by linear combination; the specific formula is as follows: ; wherein represents the risk score of the residual sequence with serial number ; α, β, γ are non-negative weighting coefficients reflecting the importance of each index in the total risk assessment. These weights can be set by artificial experience or obtained through offline learning of historical validation data.

[0031] The risk score is sent to a time integrator. The time integrator first sets two thresholds: a lower threshold (e.g. 0.8) corresponding to the highest safety margin acceptable to the system; a higher threshold (e.g. 1.5) representing a level of danger that requires immediate intervention.

[0032] When the risk score of the residual sequence is below the lower threshold, the residual sequence is not inputted as input, saving computing power to the maximum extent; When the risk score of the residual sequence is higher than the threshold, it is immediately identified as high risk, and the residual sequence is immediately inputted as input into the Bayesian online change point detection model; When the risk score of the residual sequence is between the lower threshold and the higher threshold, the detection model is not immediately started, but a time integrator is enabled to continuously integrate the risk score of the residual sequence, and whether it is continuously in an abnormal edge state is observed; At the same time, the system continuously monitors whether the current residual of the residual sequence jumps out of the three times standard deviation at one time. As long as one of the following two conditions occurs: the risk score is continuously integrated to the threshold D or the residual sequence jumps out of the three times standard deviation at one time, the Bayesian online change point detection model is immediately assigned to the residual sequence.

[0033] The calculation process of the threshold D is as follows: First, maintain a risk score sequence in a sliding time window, and calculate its statistical characteristics such as median P50 and high quantile P90 in real time; take P50 as the reference risk level R_ref, and calculate the high side fluctuation amplitude ΔR = P90-P50; take R_ref and ΔR as inputs to calculate the allowed risk integration duration T_ref, for example: ; wherein is an empirical coefficient (such as 1.5), is the reciprocal of the sampling frequency; and the threshold D is defined as: . It means that when the risk is in a higher background value and fluctuates greatly, the threshold D will be appropriately relaxed; otherwise, when the system is in a stable or low risk state, the threshold D will tend to be tightened, and the adjustment sensitivity will be improved.

[0034] In the current sliding window, the system takes the risk score sequence R(t) as a sample set, sorts the set from small to large, and takes the value at the 50th percentile position as P50, i.e. the "median", which reflects the typical normal level of the risk score in the window; similarly, the value at the 90th percentile position is taken as P90, which represents the "high quantile" corresponding to the acceptable upper extreme value under normal fluctuation background.

[0035] The once-in-three-standard-deviation jump of the residual sequence refers to that, in the current sliding buffer of several minutes, a residual value is more than three times (or more than three times) of the standard deviation from the steady-state mean value obtained by the sensor in the cold start stage. If the residual obeys the approximate normal distribution, the probability of falling outside the mean plus or minus three times the standard deviation is less than 0.003, which can be almost determined that it is not random noise but an abnormal event. Therefore, once the real-time residual appears such a "jump", the system immediately regards it as a highly significant signal of sudden zero-point drift or hardware failure, and immediately enables a more refined drift detection and compensation mechanism.

[0036] A set of Bayesian online change point detection models are configured to the screened residual sequence to realize real-time identification of sensor zero-point drift. The Bayesian online change point detection model models the "non-drift steady state" as a random process obeying the normal-inverse gamma prior, and obtains the mean and variance of the steady-state period as prior parameters by offline calibration of the historical residual data collected in the cold start stage (such as 24 hours). The confidence of the prior distribution is controlled by the estimated sample size and the degree of freedom, and the larger the parameter value, the more confident the system is in the assumption that "it is still in the steady state".

[0037] During online operation of the model, multiple "paragraph duration" candidate values are maintained at each sampling period, and sufficient statistics (including aggregated quantities of mean and variance) corresponding to each paragraph length are recursively updated. When a new observation arrives, the system calculates its predicted likelihood value under the Student distribution according to the difference between the observation and the paragraph statistics, and combines a constant hazard function (such as , where represents the prior cognition of the system to the average interval of drift, and the typical value is 10 minutes) to jointly evaluate the prior probabilities of the two mutually exclusive hypotheses of "the current paragraph continues to extend" and "the current time is a change point".

[0038] Through logarithmic domain probability operation and normalization processing, the system can output a posterior probability distribution of the current paragraph duration at each time point, and derive an instantaneous probability estimate of whether the current time is a drift point. This probability value serves as a direct trigger signal for the system to perform bias backtracking and observation noise adjustment.

[0039] At the same time, the corresponding residual expectation value is calculated based on the current statistics, and is taken as the bias estimation value of the residual sequence at the current time. Finally, the Bayesian online change point detection model outputs the following two core detection quantities: instantaneous drift probability, which represents the posterior probability of whether the current time is a change point, and the higher the value, the more likely the channel has a zero-point drift; bias estimation value, which represents the offset of the channel residual relative to the steady-state mean value, and is used to support the dynamic update of the "bias component" in the state vector in the Kalman filter.

[0040] Step 3: Multi-sensor Kalman filter fusion and bias backtracking correction In this phase, the system inputs the current power consumption observations of each sensor channel and the drift probability and bias estimates obtained in Step 2 into a multi-sensor Kalman filter. The state vector of this filter consists of two parts: one is the current true power consumption state of the system, which describes the instantaneous active load of the power distribution system; the other is the zero-point bias state of each sensor channel, which tracks the slow shift trend of the residual source. Among them, the true power consumption state is modeled using a first-order autoregressive model to describe the inertial change of the power signal; the bias state is modeled using a random walk process to describe the slow evolution characteristics of the zero-point drift.

[0041] Whenever a sensor channel is marked as abnormal by the Bayesian online change point detection model, the system automatically increases its observation noise covariance, so that the filter reduces the trust degree of the observation value of this channel when fusing; at the same time, the bias estimate output by the detection model is written into the state prediction value of the current frame, so that the filter can quickly converge to the new steady-state benchmark. In the prediction-update cycle, an adaptive Kalman gain mechanism is used, and its weight coefficient is exponentially decayed and adjusted according to the "household electricity load index" (household electricity load index is a normalized dynamic indicator used to measure the current load running pressure of the target power distribution area), to ensure that in high load fluctuation or high temperature environment, the anchor channel will not be dominated by short-term disturbance to the overall fusion result.

[0042] After the filtering phase, the system immediately performs a short window backward smoothing operation on the observation data in the last few seconds, focusing on correcting the channel just determined to be abnormal, so that its historical bias curve remains continuous and unbiased on the time axis, avoiding estimation faults caused by detection delay. After smoothing, the system writes the true power consumption estimate, channel bias estimate and covariance information contained in the latest state vector back to the shared memory buffer, providing the initial state for the next cycle of filtering loop, and also serving as the input basis for the subsequent graph convolution network correction module and energy efficiency decision module.

[0043] Step 4: Self-supervised graph convolution network learning topology-time correlation and dynamic correction After completing Kalman filtering and historical smoothing in each cycle, the system encodes the latest observation information and residual signals of each channel into a graph structure, and constructs a dynamic sensor graph within a sliding time window. The nodes of this graph represent sensor channels, and each node is attached to a feature vector, including: the original observation sequence of the last frame (e.g. 3.6 seconds), the residual after Kalman filtering, the current bias estimate and its covariance, and the "relative anchor residual slope" and "drift probability", two derived indicators reflecting dynamic abnormal trends.

[0044] The construction of edges in the graph is based on the linkage relationship between nodes. The system calculates the short-time mutual information and time delay correlation coefficient between each pair of sensor nodes, and linearly normalizes the two as edge weight indicators. The higher the mutual information and the smaller the time delay, the greater the edge weight. As the time window moves forward, the graph structure is dynamically updated: new data nodes and edge relationships are added incrementally, old data windows are eliminated; edge weight coefficients are recursively updated by exponential smoothing, isolated nodes and invalid edges are automatically deleted, ensuring that the graph is always sparse and time-efficient.

[0045] A two-layer graph convolutional encoding-decoding network is deployed on the above dynamic graph. The system uses a "random masking-reconstruction" self-supervised method for training. The graph neural network needs to restore the complete vector according to the remaining part, minimizing the restoration error, thereby guiding the model to automatically learn the temporal redundancy structure and spatial causal relationship between sensors.

[0046] In the reasoning phase, the graph neural network outputs two types of results for each node: one is a correction vector used for fine-grained zero-point and scale correction of the current sensor observation value; the second is a normalized uncertainty index generated by reconstruction residual, used to evaluate the predictability of the channel under the current topology. The system injects the correction vector into the Kalman filter observation model of the next period to correct the original observation value in real time; at the same time, the uncertainty index is equivalent to the observation noise increment, which is written into the diagonal line of the covariance matrix, so as to dynamically adjust the fusion weight of each channel in the prediction-update cycle.

[0047] Through the above topological-temporal joint modeling and self-supervised correction mechanism, the system can continuously mine the redundancy structure and temporal regularity between sensors without external labeling, and quickly provide accurate correction and uncertainty evaluation information after external environmental disturbance or local sensor drift, thereby enhancing the reconstruction ability of the Kalman filter to the true power consumption, and making the system still maintain high precision and high stability of energy efficiency estimation in the case of cross-source uncertainty improvement.

[0048] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized in electronic hardware, or a combination of computer software and electronic hardware. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0049] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other manners. For example, the embodiments of the device described above are merely schematic, and the division of the units is merely logical function division. There can be other division manners in actual implementation, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections between the units can be indirect couplings or communication connections through some interfaces, devices or units, and can be in electrical, mechanical or other forms.

[0050] The units described as separated components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e., can be located in one place, or can be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purposes of the embodiments of the present application.

[0051] In addition, each functional unit in the embodiments of the present application can be integrated in a processing unit, or each unit can exist physically as a separate unit, or two or more units can be integrated in one unit.

[0052] The above describes only specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An energy efficiency optimization control method for heterogeneous computing power clusters, characterized in that, Includes the following steps: First, using the set energy consumption anchor point as a unified time reference, the measurement values ​​of multiple power consumption sensors and environmental sensors are collected synchronously, and the measurement values ​​of multiple power consumption sensors and environmental sensors are converted into a residual sequence based on the energy consumption anchor point; the energy consumption anchor point is an absolute reference point that is manually selected with the most accurate measurement and the least interference. Next, for each residual sequence, the static zero-point displacement index, noise degradation index, and anchor point correlation attenuation index are calculated and linearly combined according to preset weights to obtain the risk score; The risk score is compared with a higher threshold and a lower threshold. When the risk score exceeds the higher threshold, the Bayesian online change point detection model is activated to output the drift probability and bias estimate. When the risk score is between the lower and higher thresholds, the Bayesian online change point detection model is activated to output the drift probability and bias estimate if one of the following two conditions occurs: the risk score is continuously integrated to the threshold D or the residual sequence jumps out of the steady state three standard deviations at once. Finally, the drift probability, bias estimate, and corresponding sensor observations are input into a multi-sensor Kalman filter to jointly estimate the true power consumption state and channel bias state. When anomalies are detected, bias backtracking correction and adaptive adjustment of observation noise covariance are performed. Subsequently, a dynamic graph structure is constructed with sensor channels as nodes and mutual information and time delay correlation coefficient as edge weights. This structure is input into a self-supervised graph convolutional neural network to obtain the correction vector and uncertainty index for each observation. The correction vector and uncertainty index for each observation are then fed back to the Kalman filter to update the fusion weights.

2. The energy efficiency optimization control method for heterogeneous computing power clusters according to claim 1, characterized in that: The steps for obtaining the static zero-point displacement index are as follows: within a fixed rolling sliding window, calculate the window average value of the residual sequence; then call the historical data of the cold start phase, extract the residual mean and residual standard deviation during the steady state period as a reference benchmark, and normalize and compare the window average value with the reference benchmark to obtain the static zero-point displacement index.

3. The energy efficiency optimization control method for heterogeneous computing power clusters according to claim 1, characterized in that: The steps for obtaining the noise degradation index are as follows: Within a fixed rolling sliding window, calculate the sample variance of the residual sequence; then use the standard deviation of the steady-state residual during the cold start phase as the historical noise benchmark, and normalize and compare the sample variance with the historical noise benchmark to obtain the noise degradation index.

4. The energy efficiency optimization control method for heterogeneous computing power clusters according to claim 1, characterized in that: The steps for obtaining the anchor point correlation decay index are as follows: Within a fixed rolling sliding window, the mean removal operation is performed on the residual sequence and the reference power sequence of the energy consumption anchor point to form a centered sequence; on this basis, the Pearson correlation coefficient between the two centered sequences is calculated in real time; then, the Pearson correlation coefficient is mapped to the distance of the full correlation state to obtain the anchor point correlation decay index; if an inverse correlation trend appears, the anchor point correlation decay index is forcibly set to the maximum value.

5. The energy efficiency optimization control method for heterogeneous computing power clusters according to claim 1, characterized in that: Before performing Bayesian online change point detection, the risk scores of each residual sequence are first screened using a time integrator. When the risk score is below a lower threshold, it is ignored directly. When it is above a higher threshold, the Bayesian online change point detection model is activated immediately. When it is between the two thresholds, the time integration method is used. If the integration result reaches the threshold D or the residual jumps out of three standard deviations from the steady state at one time, the Bayesian online change point detection model is activated immediately.

6. The energy efficiency optimization control method for heterogeneous computing power clusters according to claim 5, characterized in that: The calculation process for threshold D is as follows: First, maintain a risk score sequence within a sliding time window and calculate the median P50 and the high quantile P90 in real time; use P50 as the reference risk level R_ref and calculate the high-side fluctuation amplitude ΔR = P90 - P50; use R_ref and ΔR as inputs to calculate the allowable risk integral duration T_ref, as shown in the following formula: ;in This is an empirical coefficient. The threshold D is the reciprocal of the sampling frequency; the threshold D is defined as: .

7. The energy efficiency optimization control method for heterogeneous computing power clusters according to claim 1, characterized in that: The Bayesian online change point detection model treats the drift-free steady state as a stochastic process following a normal-inverse gamma prior and employs a constant hazard function. , This represents the system's prior knowledge of the drift mean interval. It calculates the probability of paragraph continuation and change point, and then outputs the drift probability and bias estimate.

8. The energy efficiency optimization control method for heterogeneous computing power clusters according to claim 1, characterized in that: The state vector of a multi-sensor Kalman filter simultaneously contains the actual power consumption state and the zero-point bias state of each channel.

9. The energy efficiency optimization control method for heterogeneous computing power clusters according to claim 8, characterized in that: The actual power consumption state is described by a first-order autoregressive model, and the zero-point bias state of each channel is described by a random walk model.

10. The energy efficiency optimization control method for heterogeneous computing power clusters according to claim 1, characterized in that: Whether the residual sequence jumps out of three times the steady-state standard deviation at once refers to whether the residual value of a certain sensor deviates from the steady-state average value determined by the sensor during the cold start phase by more than three times the steady-state standard deviation.

Images