CUE real-time monitoring method and system

Abstract

The invention relates to the technical field of carbon effect real-time monitoring, in particular to a CUE real-time monitoring method and system, and the method comprises the following steps: obtaining a real-time voltage and current sequence and a power grid carbon emission factor through a distributed collection terminal, carrying out the sliding window time alignment operation on the sequence through a timestamp mark, so as to generate aligned energy consumption time sequence data, and carrying out the time alignment operation on the aligned energy consumption time sequence data; decomposing the data into different energy components according to a carbon source tracking logic, executing weighted summation in combination with an emission coefficient to calculate a real-time carbon use efficiency index, and inputting the index, service load characteristics and meteorological environment data into a long short-term memory network model to generate a carbon effect trend prediction curve; a high carbon emission risk period is identified and a computing power resource scheduling instruction including business load migration and frequency adjustment parameters is constructed. According to the invention, by eliminating time sequence dislocation data caused by transmission delay, the conversion from passive accounting to active prediction scheduling is realized, and the real-time performance and granularity of carbon efficiency management are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of real-time carbon efficiency monitoring technology, and in particular to a real-time CUE monitoring method and system. Background Technology

[0002] The field of real-time carbon efficiency monitoring technology mainly involves comprehensive data collection and quantitative analysis of energy consumption structure, carbon emission intensity, and computing power output efficiency of data centers during operation. Traditional CUE real-time monitoring methods typically rely solely on monthly or quarterly average grid emission factors for ex-post calculations, using manual recording or simple periodic polling to obtain total energy consumption data. This approach fails to differentiate the dynamic ratio of grid power, diesel generators, and renewable energy in real time, and ignores the nonlinear correlation between IT load fluctuations and instantaneous carbon emission intensity. Consequently, carbon emission data acquisition exhibits significant lag, making it difficult to capture carbon efficiency changes during energy structure transitions. This leads to a disconnect between data center operation strategies and actual carbon reduction targets, hindering millisecond-level energy efficiency optimization decisions and reducing the precision of carbon management.

[0003] Existing technologies mostly rely on long-term published average grid emission factors for ex-post calculations. When obtaining total energy consumption data through manual recording or periodic polling, it is impossible to accurately distinguish the real-time dynamic ratio of grid electricity, diesel, and renewable energy. This extensive data collection mode ignores the nonlinear relationship between IT load fluctuations and instantaneous carbon emission intensity, resulting in a significant lag in carbon emission data acquisition. It is difficult to accurately capture subtle changes in carbon efficiency during energy structure switching, causing a serious disconnect between data center operation strategies and actual carbon emission reduction targets. Due to the lack of high-frequency data support, it is impossible to support millisecond-level energy efficiency optimization decisions, thus significantly reducing the precision of carbon management. Summary of the Invention

[0004] To achieve the above objectives, the present invention adopts the following technical solution: a CUE real-time monitoring method, comprising the following steps: S1: Real-time voltage sequence, current sequence and grid carbon emission factor are acquired through a distributed acquisition terminal that supports multi-protocol conversion. The real-time voltage sequence and current sequence are time aligned using a sliding window based on timestamps to remove time-series misaligned data caused by transmission delay and generate aligned energy consumption time-series data. S2: Based on the preset carbon source tracking logic, the aligned energy consumption time series data is decomposed into mains power consumption component, fuel consumption component and renewable energy power generation component. The mains power consumption component, fuel consumption component and renewable energy power generation component are weighted and summed using the grid carbon emission factor and fuel emission coefficient to calculate the real-time carbon use efficiency index. S3: Obtain the business load feature vector and meteorological environment data, input the real-time carbon utilization efficiency index, the business load feature vector and the meteorological environment data into the long short-term memory network model to perform time-series regression prediction, and generate a carbon efficiency trend prediction curve; S4: Compare the carbon efficiency trend prediction curve with the preset carbon efficiency safety threshold, identify high carbon emission risk periods exceeding the carbon efficiency safety threshold, determine the business load migration strategy and processor frequency adjustment parameters for the period, and construct computing resource scheduling instructions.

[0005] As a further aspect of the present invention, the process of generating the aligned energy consumption time series data specifically includes: S11: Controls distributed acquisition terminals deployed at the edge, and connects in parallel to the Modbus bus and SNMP network interface through the built-in adaptive protocol parsing engine. It reads the raw voltage and raw current values ​​in the data center power supply link in real time by setting the sampling frequency, and captures the dynamic carbon emission factor data released by the current regional power grid in real time through the external API interface. S12: Mark the original voltage and current values ​​read in each frame with a nanosecond-level timestamp based on the network time protocol, construct a sequence of original data frames with time index, and inject the sequence into a first-in-first-out circular buffer of a preset length to form a data stream window to be processed. S13: Execute a sliding window-based time alignment algorithm within the data stream window. Using the master clock signal as a reference, perform phase deviation calculation and linear interpolation correction on the specific timestamps of each channel's data. Eliminate timing outliers caused by network congestion or device response delays. Merge and reassemble the corrected multi-channel data to generate the aligned energy consumption timing data that is strictly synchronized in the time dimension.

[0006] As a further aspect of the present invention, the process of calculating the real-time carbon utilization efficiency index specifically includes: S21: Call the preset energy feature fingerprint library to perform spectrum analysis on the waveform features in the aligned energy consumption time series data, identify the access status signals of the mains power input, diesel generator set and photovoltaic inverter in the power supply circuit, and decompose the total energy consumption data into the corresponding mains power consumption component, the fuel consumption component and the renewable energy power generation component based on the status signal. S22: Retrieve the fuel emission coefficient that matches the current time period and fuel type from the local database, extract the power grid carbon emission factor obtained in S1, construct a weighted calculation matrix to characterize the carbon emission intensity per unit energy consumption, and determine the real-time carbon conversion weight value corresponding to each energy component. S23: Substitute the mains power consumption component, the fuel consumption component, and the renewable energy power generation component into the pre-built carbon efficiency calculation model, perform a weighted accumulation operation in combination with the real-time carbon conversion weight value, compare and normalize it with the total power consumption of the current IT equipment, and calculate the real-time carbon use efficiency index that reflects the current instantaneous carbon emission level.

[0007] As a further aspect of the present invention, the process of generating the carbon efficiency trend prediction curve specifically includes: S31: Extract CPU utilization, memory usage and I / O throughput data from the data center IT operation and maintenance management system to construct the business load feature vector, collect temperature, humidity and light intensity data inside and outside the computer room as the meteorological environment data, perform maximum and minimum normalization processing on the above data and the real-time carbon usage efficiency index, and construct a multi-dimensional input tensor. S32: The multidimensional input tensor is sliced ​​according to the time step and then input into the input layer of the long short-term memory network model. The forget gate mechanism is used to filter historical redundant information. The cell state is updated through the input gate and the output gate to capture the temporal dependency between multiple variables and extract the nonlinear high-dimensional feature vector in the hidden layer. S33: Map the nonlinear high-dimensional feature vector to a fully connected layer for regression analysis, decode and output a continuous carbon efficiency numerical sequence within a preset time window, use a smoothing filtering algorithm to eliminate prediction noise, and generate the carbon efficiency trend prediction curve that can reflect the future energy efficiency change trend.

[0008] As a further aspect of the present invention, the process of constructing the computing resource scheduling instruction specifically includes: S41: Iterate through each predicted time point value in the carbon efficiency trend prediction curve, compare it with the preset carbon efficiency safety threshold point by point, filter out all time intervals in which the predicted values ​​exceed the carbon efficiency safety threshold, and mark these consecutive time intervals as the high carbon emission risk period. S42: Based on the duration and peak intensity of the high carbon emission risk period, retrieve a matching control scheme from the preset energy-saving strategy knowledge base, determine the list of non-critical business loads to be migrated and the physical address of the target server, and calculate the processor frequency adjustment parameters that meet the heat dissipation constraints and performance requirements. S43: Encapsulate the business load list, the target server physical address, and the processor frequency adjustment parameters into a control message conforming to the data center infrastructure management protocol format, add execution priority and effective timestamp, and construct the computing resource scheduling instruction used to guide the device's actions.

[0009] As a further aspect of the present invention, the process of generating the aligned energy consumption time series data specifically includes: Obtain the start and end timestamps of the current window within the circular buffer, and calculate the variance of the sampling density distribution of each channel's data stream within the current window; The order of the Lagrange interpolation polynomial is dynamically adjusted based on the variance of the sampling density distribution. Nonlinear interpolation completion is performed on time points with missing data. The Kalman filter algorithm is used to smooth and denoise the completed real-time voltage and current sequences to eliminate high-frequency random interference components. The phase angle deviation between the real-time voltage sequence and the current sequence is detected, and the phase drift caused by the nonlinear characteristics of the transformer is corrected by using the fast Fourier transform algorithm. The corrected voltage amplitude and current amplitude are stored in a one-to-one correspondence according to the time step to generate aligned energy consumption time series data.

[0010] As a further aspect of the present invention, the calculation process of the real-time carbon utilization efficiency index specifically includes: Obtain the mains power consumption component, the fuel consumption component, the renewable energy generation component, the grid carbon emission factor, and the fuel emission coefficient; Based on the pre-defined carbon emission intensity quantification logic, the real-time carbon use efficiency index is calculated using the following formula: ; in, The real-time carbon utilization efficiency index represents time t. The mains power consumption component at time t, The fuel consumption component at time t is represented by [the fuel consumption component]. The renewable energy generation component represents the value at time t. The power grid carbon emission factor at time t, Represents the fuel emission coefficient, The inherent emission factor representing renewable energy. This represents the total effective computing power consumption of IT equipment at time t.

[0011] As a further aspect of the present invention, the process of generating the carbon efficiency trend prediction curve specifically includes: Obtain historical carbon efficiency data and historical load environment data from the training set, calculate the residual value between the model prediction output and the true label using the mean squared error loss function, and construct a comprehensive objective function by combining the L2 regularization term. An adaptive moment estimation optimization algorithm is used to backpropagate and update the weight matrix and bias vector in the long short-term memory network model until the comprehensive objective function converges to the preset error tolerance range, thus obtaining the trained prediction model. The real-time carbon utilization efficiency index, the business load feature vector, and the meteorological environment data at the current moment are input into the trained prediction model, forward inference calculation is performed, and a sequence covering the carbon efficiency change value every minute in the next hour is output, which is then visualized and rendered as the carbon efficiency trend prediction curve.

[0012] As a further aspect of the present invention, the process of constructing the computing resource scheduling instruction specifically includes: Obtain the estimated carbon emission exceedance during the high carbon emission risk period, calculate the target value of the total power consumption of IT equipment to be reduced based on the exceedance, and select a set of candidate virtual machines to be hibernated or migrated in order of business criticality from low to high. Using a dynamic voltage and frequency adjustment algorithm, combined with the current chip temperature sensor readings of the server, the minimum operating voltage and corresponding clock frequency of each core are calculated under the premise of meeting the total power consumption target value, and the processor frequency adjustment parameters are generated. The migration path information of the candidate virtual machine set and the processor frequency adjustment parameters are encoded and packaged. The integrity and security of instruction transmission are ensured by adding checksum and encrypted signature fields, and the computing resource scheduling instruction that can be directly parsed and executed by the underlying controller is generated.

[0013] A CUE real-time monitoring system, the system being used to implement the above-described CUE real-time monitoring method, the system comprising: The distributed data acquisition and alignment module is used to acquire real-time voltage sequences, current sequences, and grid carbon emission factors through a distributed acquisition terminal that supports multi-protocol conversion. It uses timestamps to perform time alignment operations on the real-time voltage sequences and current sequences based on a sliding window, eliminates time-series misaligned data caused by transmission delays, and generates aligned energy consumption time-series data. The carbon source tracking and efficiency calculation module is used to decompose the aligned energy consumption time series data into mains power consumption component, fuel consumption component and renewable energy power generation component according to the preset carbon source tracking logic, and to perform a weighted summation operation on the mains power consumption component, the fuel consumption component and the renewable energy power generation component using the grid carbon emission factor and fuel emission coefficient to calculate the real-time carbon use efficiency index. The multidimensional time-series regression prediction module is used to obtain the business load feature vector and meteorological environment data, and input the real-time carbon utilization efficiency index, the business load feature vector and the meteorological environment data into the long short-term memory network model to perform time-series regression prediction and generate a carbon efficiency trend prediction curve. The active scheduling module for computing resources is used to compare the carbon efficiency trend prediction curve with the preset carbon efficiency safety threshold, identify high carbon emission risk periods that exceed the carbon efficiency safety threshold, determine the business load migration strategy and processor frequency adjustment parameters for the period, and construct the computing resource scheduling instruction.

[0014] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, time alignment is performed by multi-protocol distributed data acquisition and using timestamps and sliding windows. Aligned energy consumption time series data is generated after eliminating time-series misaligned data caused by transmission delays. Based on carbon source tracking logic, total energy consumption is decomposed into different energy components and real-time indicators are calculated in combination with emission coefficients. A long short-term memory network model is introduced to perform time series regression prediction on business load characteristics and meteorological data. After identifying high carbon emission risk periods, scheduling instructions containing business load migration and frequency adjustment parameters are constructed. This realizes the transformation from passive accounting to active predictive scheduling and effectively improves the real-time performance and granularity of carbon efficiency management. Attached Figure Description

[0015] Figure 1 This is a flowchart of the CUE real-time monitoring method of the present invention; Figure 2 This is a flowchart illustrating the energy consumption data alignment and generation process of this invention. Figure 3 This is a flowchart of the real-time carbon utilization efficiency index calculation process of the present invention; Figure 4 This is a flowchart illustrating the generation process of the carbon efficiency trend prediction curve in this invention. Figure 5 This is a flowchart illustrating the construction process of the computing resource scheduling instructions of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the software-based technical solution is described in detail below with reference to system architecture diagrams and embodiments. It should be understood that the specific embodiments described herein are only for explaining the technical solutions of this invention and do not constitute a limitation on the scope of protection.

[0017] In the description of this invention, the system architecture relationships or data processing flows indicated by terms such as "layer," "module," "interface," "data flow," "client," and "server" are all defined based on the architecture diagram or flowchart corresponding to the embodiments. This way of describing is only used to clearly illustrate the logical relationships between the elements in the technical solution, and not to limit the physical deployment form. The term "multiple" includes two or more technical units, including but not limited to multiple data nodes, processing threads, service instances, or functional components and other scalable elements. The specific number is determined according to the actual business scenario and needs to be specifically specified.

[0018] Please see Figure 1 and Figure 2 This invention provides a technical solution: a CUE real-time monitoring method, comprising the following steps: S1: Real-time voltage sequence, current sequence and grid carbon emission factor are acquired through a distributed acquisition terminal that supports multi-protocol conversion. The real-time voltage sequence and current sequence are time aligned using a sliding window based on timestamps to remove time-series misaligned data caused by transmission delay and generate aligned energy consumption time-series data. The process of generating aligned energy consumption time series data specifically includes: S11: Controls distributed acquisition terminals deployed at the edge, and connects in parallel to the Modbus bus and SNMP network interface through the built-in adaptive protocol parsing engine. It reads the raw voltage and raw current values ​​in the data center power supply link in real time by setting the sampling frequency, and captures the dynamic carbon emission factor data released by the current regional power grid in real time through the external API interface. S12: Mark the original voltage and current values ​​read in each frame with a nanosecond-level timestamp based on the network time protocol, construct a sequence of original data frames with time index, and inject the sequence into a first-in-first-out circular buffer of a preset length to form a data stream window to be processed. S13: Execute a sliding window-based time alignment algorithm within the data stream window. Using the master clock signal as a reference, perform phase deviation calculation and linear interpolation correction on the specific timestamps of each channel's data. Eliminate timing outliers caused by network congestion or device response delays. Merge and reassemble the corrected multi-channel data to generate aligned energy consumption timing data that is strictly synchronized in the time dimension. The process of generating aligned energy consumption time series data specifically includes: Obtain the start and end timestamps of the current window within the circular buffer, and calculate the variance of the sampling density distribution of each channel's data stream within the current window; The order of the Lagrange interpolation polynomial is dynamically adjusted based on the variance of the sampling density distribution. Nonlinear interpolation completion is performed on time points with missing data. The Kalman filter algorithm is used to smooth and denoise the completed real-time voltage and current sequences to eliminate high-frequency random interference components. The phase angle deviation between the real-time voltage sequence and the current sequence is detected. The phase drift caused by the nonlinear characteristics of the transformer is corrected by using the fast Fourier transform algorithm. The corrected voltage amplitude and current amplitude are stored in a one-to-one correspondence according to the time step to generate aligned energy consumption time series data.

[0019] In this step, the distributed acquisition terminals deployed at the edge are first activated to establish a parallel data acquisition channel for the heterogeneous power supply equipment in the data center. For traditional distribution cabinets and rack-mount cabinets that do not have direct network communication capabilities, an RS-485 serial interface is used to connect to the Modbus-RTU bus, and the baud rate is set to... Data bits are The stop bit is 0. No parity bit, with For the polling cycle, the function code is sent to each electricity metering instrument. The query instruction (read holding register) reads the address offset. to The registers contain the raw voltage and current values. Simultaneously, for intelligent UPS and PDU devices supporting Ethernet communication, a secure connection is established via the SNMPv3 protocol, configured with SHA-256 authentication and AES-128 encryption algorithms, and the voltage and current MIB object values ​​under the OID node are retrieved in batches using the GetBulkRequest primitive. While acquiring physical link data, a GET request is initiated via the HTTPS protocol to the open API interface of the regional power grid dispatch center to obtain the current dynamic carbon emission factor data.

[0020] The aforementioned dynamic carbon emission factor data refers to data reflecting the mixed carbon emission intensity of different power sources, including thermal power, hydropower, wind power, and nuclear power, within the current power grid dispatch range. This data fluctuates in real time with the power grid dispatch strategy and is typically expressed as follows: Units.

[0021] Subsequently, a timestamp marking operation based on a high-precision clock source is performed. An NTP server deployed within the local area network is used as the primary time source to synchronize the system clock of the local acquisition terminal to the Stratum2 level, ensuring that the time synchronization error is controlled within a specified range. Within this range, as each frame of raw voltage and current values ​​is read into memory, a high-precision timing function at the operating system kernel level is immediately invoked to obtain the current nanosecond-level timestamp. This timestamp is then appended as metadata to the data payload header, constructing a sequence of raw data frames with a time index. These data frames are then sequentially injected into a memory with a capacity of [missing information]. A dynamically updated window of data stream to be processed is formed in the first-in-first-out circular buffer of the frame.

[0022] Within the data stream window, perform time alignment processing based on a sliding window. Set the time span of the sliding window to... Step size is Using the master clock signal as a reference, traverse the data frames from different channels within the window. Calculate the variance of the sampling density distribution of each channel's data stream within the current window. If the variance of the sampling density distribution of a certain channel within the current window... Greater than the preset stability threshold For time points with missing data within the window, select the adjacent time points before and after that time point. Each valid data frame is nonlinearly completed using a third-order Lagrange interpolation polynomial; if the variance... Less than or equal to If the result is not satisfactory, then linear interpolation is used for rapid completion.

[0023] After interpolation completion, Kalman filtering is performed on the real-time voltage and current sequences to eliminate measurement noise. A state vector is then constructed. ,in In voltage state, The current state is defined. The state transition matrix is ​​set. The second-order identity matrix, the observation matrix Also a second-order identity matrix. Measurement noise covariance matrix. The value is set according to the sensor accuracy manual. Diagonal elements are set to Process noise covariance matrix Set as The completed sequence is input into a Kalman filter, and through a "prediction-update" loop, a smoothed sequence of voltage and current amplitudes is output.

[0024] Finally, phase drift of the AC signal is corrected. A Fast Fourier Transform is performed on the voltage and current sequences within the sliding window to calculate the fundamental voltage phase. Phase with the fundamental current The difference between If the deviation between this difference and the theoretical value exceeds the preset phase tolerance... Then, the current sequence is corrected by time shift in the time domain. The corrected voltage and current amplitudes are then combined with the previously obtained grid carbon emission factor and aligned and merged at a uniform time step to generate aligned energy consumption time series data that are strictly synchronized in the time dimension.

[0025] Table 1 lists the key parameter configurations for multi-protocol acquisition and alignment processing in this embodiment.

[0026] Table 1 Data Acquisition and Alignment Parameter Configuration Table ; As shown in Table 1, the problem of data timing misalignment between heterogeneous devices was solved through precise configuration of the above parameters and algorithm execution. Experimental data shows that after the above alignment process, the average time synchronization error of multi-source data was reduced from the original... Reduce to .

[0027] Please see Figure 1 and Figure 3 S2: Based on the preset carbon source tracking logic, the aligned energy consumption time series data is decomposed into mains power consumption component, fuel consumption component and renewable energy power generation component. The mains power consumption component, fuel consumption component and renewable energy power generation component are weighted and summed using the grid carbon emission factor and fuel emission coefficient to calculate the real-time carbon use efficiency index. The process of calculating real-time carbon use efficiency indicators specifically includes: S21: Call the preset energy feature fingerprint library to perform spectrum analysis on the waveform features in the aligned energy consumption time series data, identify the access status signals of the mains power input, diesel generator set and photovoltaic inverter in the power supply circuit, and decompose the total energy consumption data into the corresponding mains power consumption component, fuel consumption component and renewable energy power generation component based on the status signal. S22: Retrieve fuel emission coefficients that match the current time period and fuel type from the local database, extract the grid carbon emission factor obtained in S1, construct a weighted calculation matrix to characterize the carbon emission intensity per unit energy consumption, and determine the real-time carbon conversion weight value corresponding to each energy component. S23: Substitute the mains power consumption component, fuel consumption component and renewable energy power generation component into the pre-built carbon efficiency calculation model, perform weighted accumulation calculation in combination with the real-time carbon conversion weight value, compare and normalize with the total power consumption of the current IT equipment, and calculate the real-time carbon use efficiency index that reflects the current instantaneous carbon emission level. The calculation process for real-time carbon use efficiency indicators specifically includes: Obtain the components of mains power consumption, fuel consumption, renewable energy generation, grid carbon emission factor, and fuel emission coefficient; Based on the pre-defined carbon emission intensity quantification logic, the real-time carbon use efficiency index is calculated using the following formula: ; in, This represents the real-time carbon utilization efficiency index at time t. The mains power consumption component at time t. This represents the fuel consumption component at time t. The renewable energy generation component at time t, The carbon emission factor of the power grid at time t. Represents the fuel emission factor. The inherent emission factor representing renewable energy. This represents the total effective computing power consumption of IT equipment at time t.

[0028] In this step, the aligned energy consumption time-series data generated in step S1 is first decomposed from source to end based on energy feature fingerprinting technology. High-frequency harmonic analysis is then performed on the voltage and current waveforms by calling a pre-defined energy feature fingerprint database. Wavelet transform is used to extract the time-frequency feature vector of the current energy consumption data, and this vector is then matched with the template in the fingerprint database using cosine similarity. When the similarity between the feature vector and the diesel generator set fingerprint exceeds a certain threshold... At that time, it is determined that the current partial load is borne by the diesel generator, and the corresponding fuel consumption component is identified based on the circuit topology; similarly, by identifying the unique switching frequency ripple characteristics of the photovoltaic inverter, the renewable energy power generation component is identified.

[0029] Subsequently, a weighted calculation matrix for carbon emission intensity is constructed. Fuel emission coefficients are retrieved from the local database. In this embodiment, the emission coefficient for No. 0 diesel is set to be... Simultaneously, the grid carbon emission factor acquired in real time in step S1 is extracted. For renewable energy sources, an inherent emission factor is set. for .

[0030] Next, we will perform the specific calculation of the real-time carbon use efficiency index. (Get the current time...) Data for each component: Mains power consumption component Fuel consumption Renewable energy generation component and the total effective computing power consumption of IT equipment The real-time carbon use efficiency index is calculated using the following formula: ; in, represent Real-time carbon efficiency index at any given time, in units of ; represent The amount of mains power consumed at any given time, in units of ; represent Fuel consumption at any given time, in units of ; represent Renewable energy generation at any given time, in units of ; represent The carbon emission factor of the power grid at any given time, in units of ; Represents the fuel emission factor, in units of ; The inherent emission factor representing renewable energy, in units of ; represent Total effective computing power consumption of IT equipment at any given time, in units of .

[0031] To verify the accuracy of the computational logic, a practical example is introduced for illustration. Assume that at time... The system monitored the following data: mains power consumption. for fuel consumption for Renewable energy generation component for Total effective computing power consumption of IT equipment for The carbon emission factor of the power grid obtained at this time for Fuel emission coefficient Take a fixed value Inherent emission coefficient of renewable energy Take a fixed value .

[0032] Substituting the above parameters into the formula for calculation: the result of the numerator (carbon emission rate) is: The final indicator calculation results are as follows: .

[0033] The calculation result This indicates that at the current moment, the data center outputs... The effective computing power corresponds to the emissions into the environment. Carbon dioxide.

[0034] Please see Figure 1 and Figure 4 S3: Obtain the business load feature vector and meteorological environment data, input the real-time carbon utilization efficiency index, business load feature vector and meteorological environment data into the long short-term memory network model to perform time series regression prediction, and generate carbon efficiency trend prediction curve; The process of generating a carbon efficiency trend prediction curve specifically includes: S31: Extract CPU utilization, memory usage and I / O throughput data from the data center IT operation and maintenance management system to construct a business load feature vector, collect temperature, humidity and light intensity data inside and outside the computer room as meteorological environment data, perform maximum and minimum normalization processing on the above data and real-time carbon usage efficiency indicators, and construct a multi-dimensional input tensor. S32: The multidimensional input tensor is sliced ​​according to the time step and then input into the input layer of the long short-term memory network model. The forget gate mechanism is used to filter historical redundant information. The cell state is updated through the input gate and the output gate to capture the temporal dependence between multiple variables and extract the nonlinear high-dimensional feature vector in the hidden layer. S33: Map the nonlinear high-dimensional feature vector to the fully connected layer for regression analysis, decode and output the continuous carbon efficiency numerical sequence within the future preset time window, use the smoothing filtering algorithm to eliminate prediction noise, and generate a carbon efficiency trend prediction curve that can reflect the future energy efficiency change trend. The process of generating the carbon efficiency trend prediction curve specifically includes: Obtain historical carbon efficiency data and historical load environment data from the training set, calculate the residual value between the model prediction output and the true label using the mean squared error loss function, and construct a comprehensive objective function by combining the L2 regularization term. An adaptive moment estimation optimization algorithm is used to backpropagate and update the weight matrix and bias vector in the long short-term memory network model until the comprehensive objective function converges to the preset error tolerance range, thus obtaining the trained prediction model. The real-time carbon utilization efficiency index, business load feature vector, and meteorological environment data are input into the trained prediction model. Forward inference calculation is performed, and the output is a sequence covering the carbon efficiency change value every minute in the next hour. This sequence is then visualized and rendered as a carbon efficiency trend prediction curve.

[0035] In this step, a multidimensional input tensor is first constructed. Business load characteristic data is extracted from the Data Center IT Operations and Maintenance Management System (DCIM) via API interface, specifically including: average CPU utilization, memory usage, and disk I / O throughput across the entire network. Simultaneously, meteorological environmental data is collected through an environmental sensor network deployed inside and outside the data center, including cold aisle temperature, outdoor ambient temperature, relative humidity, and light intensity. Past data is then selected. Historical data in minutes is used as the input sequence, collected once per minute. For the data of different units mentioned above, max-min normalization is performed. For example, for CPU utilization data... Using the formula Map it to Interval. The normalized data is compared with the historical data calculated in step S2. Sequence concatenation to construct a shape of The three-dimensional input tensor.

[0036] Next, a Long Short-Term Memory (LSTM) network model is constructed and trained. This model contains two LSTM hidden layers; the first layer contains… One neuron, with return_sequences=True set to preserve time step outputs; the second layer contains Each neuron has a return_sequences=False value to extract the hidden state at the final time step. A fully connected layer is then connected after the LSTM layer, with an output dimension of... This corresponds to the predicted carbon efficiency value per minute for the next hour.

[0037] During the training phase, mean squared error (MSE) is selected as the loss function. The Adam optimizer is used for parameter updates. The loss function is expressed as: in, Represents the total loss value. Represents the number of samples. Represents the true carbon efficiency value, This represents the predicted value output by the model. Represents the L2 regularization coefficient (set to in this embodiment). ), Represents the model weight parameters. This is the weight decay term.

[0038] The current time calculated by S2 The model is trained using real-time collected load and environmental data for forward inference. The LSTM unit utilizes a forgetting gate mechanism to determine which historical information to discard and updates the cell state using the input gate, capturing the long-term dependency between load fluctuations and carbon efficiency changes. The model ultimately outputs the future... A prediction sequence of minutes. To eliminate high-frequency jitter in the prediction results, a window size of [missing value] is used. The moving average filtering algorithm is used to smooth the predicted sequence. The smoothed sequence is then mapped onto the time axis to generate a carbon efficiency trend prediction curve.

[0039] Experimental data show that the root mean square error (RMSE) of the model is 0.018 on the test set, which is about 25% higher than the traditional ARIMA model, and can effectively capture the nonlinear changes in carbon efficiency caused by fluctuations in photovoltaic power output.

[0040] Please see Figure 1 and Figure 5 S4: Compare the carbon efficiency trend prediction curve with the preset carbon efficiency safety threshold, identify the high carbon emission risk period that exceeds the carbon efficiency safety threshold, determine the business load migration strategy and processor frequency adjustment parameters for the period, and construct computing resource scheduling instructions. The process of constructing computing resource scheduling instructions specifically includes: S41: Iterate through the carbon efficiency trend prediction curve for each predicted time point value, compare it with the preset carbon efficiency safety threshold point by point, filter out all time intervals in which the predicted values ​​exceed the carbon efficiency safety threshold, and mark these consecutive time intervals as high carbon emission risk periods. S42: Based on the duration and peak intensity of the high carbon emission risk period, retrieve the matching control scheme from the pre-set energy-saving strategy knowledge base, determine the list of non-critical business loads to be migrated and the physical address of the target server, and calculate the processor frequency adjustment parameters that meet the heat dissipation constraints and performance requirements. S43: Encapsulate the business load list, target server physical address, and processor frequency adjustment parameters into a control message conforming to the data center infrastructure management protocol format, add execution priority and effective timestamp, and build and generate computing resource scheduling instructions to guide device actions; The process of constructing computing resource scheduling instructions specifically includes: Obtain the estimated carbon emission exceedance during periods of high carbon emission risk, calculate the target value of total power consumption of IT equipment to be reduced based on the exceedance, and select a set of candidate virtual machines to be hibernated or migrated in order of business criticality from low to high. By using a dynamic voltage and frequency adjustment algorithm and combining the current chip temperature sensor readings of the server, the minimum operating voltage and corresponding clock frequency of each core are calculated under the premise of meeting the total power consumption target value, and the processor frequency adjustment parameters are generated. The migration path information of the candidate virtual machine set and the processor frequency adjustment parameters are encoded and packaged. The integrity and security of instruction transmission are ensured by adding checksum and encrypted signature fields, and computing resource scheduling instructions that can be directly parsed and executed by the underlying controller are generated.

[0041] In the execution of this step, the carbon efficiency safety threshold is first set. This threshold is based on the historical average CUE value calculated in step S2. ) rise Setting, i.e. Traverse the future generated in step S3. Minute carbon efficiency trend prediction curve, point-by-point detection of predicted values If continuous Predictions for more than a minute satisfy the following conditions: If so, then the time interval is marked as a period of high carbon emission risk.

[0042] For the identified high-carbon emission risk periods, calculate the target value for the total power consumption reduction of IT equipment. Assume that the model predicts a peak CUE during the high-carbon emission risk periods. The expected total load for this period for In order to keep the CUE metric within a safe threshold. Based on the principles of energy conservation and total carbon emission constraints, the total load needs to be reduced to the target value. The calculation process is as follows: Therefore, the target power consumption to be reduced is: .

[0043] Based on the calculation To achieve power reduction targets, a business workload migration strategy is implemented. Based on a pre-built energy-saving strategy knowledge base, currently running virtual machines are sorted according to the criticality level of the business. Non-critical virtual machines (Level 1 and Level 2) are selected as a candidate migration set, and their target physical addresses for migration to a remote low-carbon data center are determined.

[0044] Simultaneously, for services that remain locally, the processor frequency adjustment parameter (DVFS) is calculated. This is achieved using a dynamic voltage-frequency adjustment algorithm, combined with current chip temperature sensor readings from the server. It is assumed that the server CPU is currently running at... (P0 state), core voltage To reduce power consumption while meeting the minimum performance SLA requirements, the CPU core frequency was adjusted by consulting the P-State table. (P2 state), corresponding to a reduction in operating voltage to According to the power consumption formula The adjusted single-core power consumption is approximately 52% of the original power consumption, thereby helping to achieve the overall power consumption reduction target.

[0045] Finally, a computing resource scheduling instruction is constructed. This instruction is encapsulated in JSON format and includes a digital signature to ensure security. Table 2 shows the key parameters of an example scheduling instruction generated in this step.

[0046] Table 2 Examples of Computing Resource Scheduling Command Parameters ; As shown in Table 2, the generated scheduling instructions specify the exact power reduction values ​​and device action parameters within a specific time window. After the instructions are sent to the underlying controller, the system will automatically perform migration and frequency reduction operations, causing the actual carbon efficiency indicators to return to the safe threshold. the following.

[0047] CUE stands for Carbon Use Efficiency, a carbon efficiency metric used in the data center field to quantify the relationship between greenhouse gas emissions (carbon dioxide equivalent) generated by data centers during operation and the energy consumption of IT equipment.

[0048] A CUE real-time monitoring system, the CUE real-time monitoring system being used to execute the above-described CUE real-time monitoring method, the system comprising: The distributed data acquisition and alignment module is used to acquire real-time voltage sequences, current sequences, and grid carbon emission factors through distributed acquisition terminals that support multi-protocol conversion. It uses timestamps to perform time alignment operations on real-time voltage sequences and current sequences based on sliding windows, eliminates time-series misaligned data caused by transmission delays, and generates aligned energy consumption time-series data. The carbon source tracking and efficiency calculation module is used to decompose the aligned energy consumption time series data into grid power consumption component, fuel consumption component and renewable energy power generation component according to the preset carbon source tracking logic. It uses the grid carbon emission factor and fuel emission coefficient to perform a weighted summation operation on the grid power consumption component, fuel consumption component and renewable energy power generation component to calculate the real-time carbon use efficiency index. The multidimensional time-series regression prediction module is used to obtain the business load feature vector and meteorological environment data. It inputs the real-time carbon utilization efficiency index, business load feature vector and meteorological environment data into the long short-term memory network model to perform time-series regression prediction and generate carbon efficiency trend prediction curve. The active scheduling module for computing resources is used to compare the carbon efficiency trend prediction curve with the preset carbon efficiency safety threshold, identify high carbon emission risk periods that exceed the carbon efficiency safety threshold, determine the business load migration strategy and processor frequency adjustment parameters for the period, and construct computing resource scheduling instructions.

[0049] The above embodiments illustrate preferred embodiments of the present invention. Any equivalent adjustments to the technical solution based on software engineering methods are within the scope of protection, including but not limited to: implementing algorithm logic using different programming languages, refactoring functional modules into services, adjusting data interaction protocols, and optimizing resource scheduling strategies. Any implementation scheme derived from reasonable modifications to the data processing flow, service call chain, or system architecture layer without departing from the core technology of the present invention should be considered within the protection scope defined by the technical solution of the present invention.

Claims

1. A method for real-time monitoring of CUE, characterized in that, Includes the following steps: S1: Real-time voltage sequence, current sequence and grid carbon emission factor are acquired through a distributed acquisition terminal that supports multi-protocol conversion. The real-time voltage sequence and current sequence are time aligned using a sliding window based on timestamps to remove time-series misaligned data caused by transmission delay and generate aligned energy consumption time-series data. S2: Based on the preset carbon source tracking logic, the aligned energy consumption time series data is decomposed into mains power consumption component, fuel consumption component and renewable energy power generation component. The mains power consumption component, fuel consumption component and renewable energy power generation component are weighted and summed using the grid carbon emission factor and fuel emission coefficient to calculate the real-time carbon use efficiency index. S3: Obtain the business load feature vector and meteorological environment data, input the real-time carbon utilization efficiency index, the business load feature vector and the meteorological environment data into the long short-term memory network model to perform time-series regression prediction, and generate a carbon efficiency trend prediction curve; S4: Compare the carbon efficiency trend prediction curve with the preset carbon efficiency safety threshold, identify high carbon emission risk periods exceeding the carbon efficiency safety threshold, determine the business load migration strategy and processor frequency adjustment parameters for the period, and construct computing resource scheduling instructions.

2. The CUE real-time monitoring method according to claim 1, characterized in that, The process of generating the aligned energy consumption time series data specifically includes: S11: Controls distributed acquisition terminals deployed at the edge, and connects in parallel to the Modbus bus and SNMP network interface through the built-in adaptive protocol parsing engine. It reads the raw voltage and raw current values ​​in the data center power supply link in real time by setting the sampling frequency, and captures the dynamic carbon emission factor data released by the current regional power grid in real time through the external API interface. S12: Mark the original voltage and current values ​​read in each frame with a nanosecond-level timestamp based on the network time protocol, construct a sequence of original data frames with time index, and inject the sequence into a first-in-first-out circular buffer of a preset length to form a data stream window to be processed. S13: Execute a sliding window-based time alignment algorithm within the data stream window. Using the master clock signal as a reference, perform phase deviation calculation and linear interpolation correction on the specific timestamps of each channel's data. Eliminate timing outliers caused by network congestion or device response delays. Merge and reassemble the corrected multi-channel data to generate the aligned energy consumption timing data that is strictly synchronized in the time dimension.

3. The CUE real-time monitoring method according to claim 1, characterized in that, The process of calculating the real-time carbon use efficiency index specifically includes: S21: Call the preset energy feature fingerprint library to perform spectrum analysis on the waveform features in the aligned energy consumption time series data, identify the access status signals of the mains power input, diesel generator set and photovoltaic inverter in the power supply circuit, and decompose the total energy consumption data into the corresponding mains power consumption component, the fuel consumption component and the renewable energy power generation component based on the status signal. S22: Retrieve the fuel emission coefficient that matches the current time period and fuel type from the local database, extract the power grid carbon emission factor obtained in S1, construct a weighted calculation matrix to characterize the carbon emission intensity per unit energy consumption, and determine the real-time carbon conversion weight value corresponding to each energy component. S23: Substitute the mains power consumption component, the fuel consumption component, and the renewable energy power generation component into the pre-built carbon efficiency calculation model, perform a weighted accumulation operation in combination with the real-time carbon conversion weight value, compare and normalize it with the total power consumption of the current IT equipment, and calculate the real-time carbon use efficiency index that reflects the current instantaneous carbon emission level.

4. The CUE real-time monitoring method according to claim 1, characterized in that, The process of generating the carbon efficiency trend prediction curve specifically includes: S31: Extract CPU utilization, memory usage and I / O throughput data from the data center IT operation and maintenance management system to construct the business load feature vector, collect temperature, humidity and light intensity data inside and outside the computer room as the meteorological environment data, perform maximum and minimum normalization processing on the above data and the real-time carbon usage efficiency index, and construct a multi-dimensional input tensor. S32: The multidimensional input tensor is sliced ​​according to the time step and then input into the input layer of the long short-term memory network model. The forget gate mechanism is used to filter historical redundant information. The cell state is updated through the input gate and the output gate to capture the temporal dependency between multiple variables and extract the nonlinear high-dimensional feature vector in the hidden layer. S33: Map the nonlinear high-dimensional feature vector to a fully connected layer for regression analysis, decode and output a continuous carbon efficiency numerical sequence within a preset time window, use a smoothing filtering algorithm to eliminate prediction noise, and generate the carbon efficiency trend prediction curve that can reflect the future energy efficiency change trend.

5. The CUE real-time monitoring method according to claim 1, characterized in that, The process of constructing the computing resource scheduling instruction specifically includes: S41: Iterate through each predicted time point value in the carbon efficiency trend prediction curve, compare it with the preset carbon efficiency safety threshold point by point, filter out all time intervals in which the predicted values ​​exceed the carbon efficiency safety threshold, and mark these consecutive time intervals as the high carbon emission risk period. S42: Based on the duration and peak intensity of the high carbon emission risk period, retrieve a matching control scheme from the preset energy-saving strategy knowledge base, determine the list of non-critical business loads to be migrated and the physical address of the target server, and calculate the processor frequency adjustment parameters that meet the heat dissipation constraints and performance requirements. S43: Encapsulate the business load list, the target server physical address, and the processor frequency adjustment parameters into a control message conforming to the data center infrastructure management protocol format, add execution priority and effective timestamp, and construct the computing resource scheduling instruction used to guide the device's actions.

6. The CUE real-time monitoring method according to claim 2, characterized in that, The process of generating the aligned energy consumption time series data specifically includes: Obtain the start and end timestamps of the current window within the circular buffer, and calculate the variance of the sampling density distribution of each channel's data stream within the current window; The order of the Lagrange interpolation polynomial is dynamically adjusted based on the variance of the sampling density distribution. Nonlinear interpolation completion is performed on time points with missing data. The Kalman filter algorithm is used to smooth and denoise the completed real-time voltage and current sequences to eliminate high-frequency random interference components. The phase angle deviation between the real-time voltage sequence and the current sequence is detected, and the phase drift caused by the nonlinear characteristics of the transformer is corrected by using the fast Fourier transform algorithm. The corrected voltage amplitude and current amplitude are stored in a one-to-one correspondence according to the time step to generate aligned energy consumption time series data.

7. The CUE real-time monitoring method according to claim 3, characterized in that, The calculation process for the real-time carbon use efficiency index specifically includes: Obtain the mains power consumption component, the fuel consumption component, the renewable energy generation component, the grid carbon emission factor, and the fuel emission coefficient; Based on the pre-defined carbon emission intensity quantification logic, the real-time carbon use efficiency index is calculated using the following formula: ； in, The real-time carbon utilization efficiency index represents time t. The mains power consumption component at time t, The fuel consumption component at time t is represented by [the fuel consumption component]. The renewable energy generation component represents the value at time t. The power grid carbon emission factor at time t, Represents the fuel emission coefficient, The inherent emission factor representing renewable energy. This represents the total effective computing power consumption of IT equipment at time t.

8. The CUE real-time monitoring method according to claim 4, characterized in that, The process of generating the carbon efficiency trend prediction curve specifically includes: Obtain historical carbon efficiency data and historical load environment data from the training set, calculate the residual value between the model prediction output and the true label using the mean squared error loss function, and construct a comprehensive objective function by combining the L2 regularization term. An adaptive moment estimation optimization algorithm is used to backpropagate and update the weight matrix and bias vector in the long short-term memory network model until the comprehensive objective function converges to the preset error tolerance range, thus obtaining the trained prediction model. The real-time carbon utilization efficiency index, the business load feature vector, and the meteorological environment data at the current moment are input into the trained prediction model, forward inference calculation is performed, and a sequence covering the carbon efficiency change value every minute in the next hour is output, which is then visualized and rendered as the carbon efficiency trend prediction curve.

9. The CUE real-time monitoring method according to claim 5, characterized in that, The process of constructing the computing resource scheduling instructions specifically includes: Obtain the estimated carbon emission exceedance during the high carbon emission risk period, calculate the target value of the total power consumption of IT equipment to be reduced based on the exceedance, and select a set of candidate virtual machines to be hibernated or migrated in order of business criticality from low to high. Using a dynamic voltage and frequency adjustment algorithm, combined with the current chip temperature sensor readings of the server, the minimum operating voltage and corresponding clock frequency of each core are calculated under the premise of meeting the total power consumption target value, and the processor frequency adjustment parameters are generated. The migration path information of the candidate virtual machine set and the processor frequency adjustment parameters are encoded and packaged. The integrity and security of instruction transmission are ensured by adding checksum and encrypted signature fields, and the computing resource scheduling instruction that can be directly parsed and executed by the underlying controller is generated.

10. A CUE real-time monitoring system, characterized in that, The system is used to implement the CUE real-time monitoring method according to any one of claims 1-9, the system comprising: The distributed data acquisition and alignment module is used to acquire real-time voltage sequences, current sequences, and grid carbon emission factors through a distributed acquisition terminal that supports multi-protocol conversion. It uses timestamps to perform time alignment operations on the real-time voltage sequences and current sequences based on a sliding window, eliminates time-series misaligned data caused by transmission delays, and generates aligned energy consumption time-series data. The carbon source tracking and efficiency calculation module is used to decompose the aligned energy consumption time series data into mains power consumption component, fuel consumption component and renewable energy power generation component according to the preset carbon source tracking logic, and to perform a weighted summation operation on the mains power consumption component, the fuel consumption component and the renewable energy power generation component using the grid carbon emission factor and fuel emission coefficient to calculate the real-time carbon use efficiency index. The multidimensional time-series regression prediction module is used to obtain the business load feature vector and meteorological environment data, and input the real-time carbon utilization efficiency index, the business load feature vector and the meteorological environment data into the long short-term memory network model to perform time-series regression prediction and generate a carbon efficiency trend prediction curve. The active scheduling module for computing resources is used to compare the carbon efficiency trend prediction curve with the preset carbon efficiency safety threshold, identify high carbon emission risk periods that exceed the carbon efficiency safety threshold, determine the business load migration strategy and processor frequency adjustment parameters for the period, and construct the computing resource scheduling instruction.

Images