A data center PUE real-time monitoring method and system

Abstract

The present application relates to the technical field of measuring electric power, in particular to a data center PUE real-time monitoring method and system, comprising the following steps: obtaining current voltage, power, environment and service load data, setting a window to calculate total energy consumption and IT equipment energy consumption, dividing the two to obtain PUE value, storing total energy consumption, PUE and service load in time sequence to build historical sequence, and inputting LSTM network to predict future energy consumption curve.In the present application, current voltage values are obtained in real time, power module input and output power and core component running power are obtained, total energy consumption of the data center and IT equipment energy consumption calculation mapping are established, the problem of one-sided PUE result caused by the limitation of traditional measurement point is overcome, power utilization efficiency is obtained, energy consumption values are associated with service load data to build historical energy consumption sequence, time series data regression prediction ability is used to output future energy consumption expected curve, the defect that the prior art can only statistically analyze after the event but cannot predict the trend is solved, and decision support is provided for active energy efficiency optimization.

Description

Technical Field

[0001] This invention relates to the field of electrical power measurement technology, and in particular to a method and system for real-time monitoring of power usage effectiveness (PUE) in data centers. Background Technology

[0002] The field of power measurement technology involves the quantification and analysis of energy consumption in power systems. Core aspects of this field include the accurate determination of voltage, current, active power, reactive power, and power factor. These measurements typically rely on current transformers to acquire signals and are calculated using dedicated power metering chips or processing units to support energy efficiency management, cost accounting, and grid load monitoring. Traditional real-time PUE monitoring methods for data centers refer to a calculation process used to assess the energy efficiency of data center usage. This method typically involves installing a main power meter at the main power inlet of the data center and IT power meters on the PDU power distribution units in the IT equipment racks. Based on the total energy consumption readings collected by these meters at specific time intervals and the energy consumption readings of the IT equipment, the PUE ratio is derived through manual calculation or simple data aggregation.

[0003] Traditional PUE monitoring methods rely on installing energy meters at the main incoming line and IT equipment. This approach only collects total energy consumption readings and IT equipment energy consumption readings, resulting in a single data dimension. It cannot reflect the detailed operating status of each component within the data center, nor does it correlate with environmental factors or changes in business load. This leads to lagging and one-sided PUE calculation results, making it difficult for analysts to gain in-depth insights into the specific drivers behind energy consumption fluctuations. For example, it is impossible to distinguish whether increased IT load or decreased air conditioning efficiency is causing the energy consumption increase. Therefore, it lacks the ability to effectively predict future energy consumption trends, which is not conducive to developing forward-looking energy-saving strategies. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a real-time monitoring method and system for data center PUE.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a method for real-time monitoring of PUE in a data center, comprising the following steps:

[0006] S1: Obtain the real-time current and voltage values ​​of the intelligent power distribution unit, the input and output power of the power module, the operating power of the core components, environmental parameters, and business load data. The environmental parameters consist of air temperature, humidity, and flow rate values, and the business load data consists of the number of requests and the total number of concurrent connections, all of which are associated with timestamps.

[0007] S2: Set an aggregation window, map the real-time current and voltage values, the power module input and output power, and the core component operating power based on the timestamp, sum the real-time current and voltage values ​​and the power module input and output power in the aggregation window to obtain the total energy consumption value of the data center, and sum the core component operating power to obtain the IT equipment energy consumption value.

[0008] S3: Obtain the power efficiency value by dividing the total energy consumption of the data center by the energy consumption of the IT equipment;

[0009] S4: Store the total energy consumption value of the data center, the optimal power efficiency value, and the business load data in time sequence to construct a historical energy consumption sequence. Input the historical energy consumption sequence into the long short-term memory network for regression prediction to obtain the expected future energy consumption curve.

[0010] As a further aspect of the present invention, the environmental parameters include air temperature, humidity, and flow rate values; the business load data includes the number of requests and the total number of concurrent connections; the total energy consumption value of the data center includes real-time current and voltage values ​​and power module input and output power; the IT equipment energy consumption value specifically refers to the operating power of core components; the power utilization efficiency value includes the total energy consumption value of the data center and the energy consumption value of IT equipment; and the historical energy consumption sequence includes the total energy consumption value of the data center, the power utilization efficiency value, and business load data.

[0011] As a further aspect of the present invention, step S1 specifically comprises:

[0012] S11: Periodically poll the intelligent power distribution unit through the Modbus-TCP protocol interface to obtain real-time current and voltage values ​​associated with timestamps, and simultaneously capture the SNMPMIB library data of the power module to obtain the input and output power of the power module.

[0013] S12: Deploy distributed sensor nodes to collect air temperature, humidity, and flow rate values ​​to form environmental parameters, and connect to the application layer gateway to parse the number of requests and the total number of concurrent connections included in the business load data;

[0014] S13: Establish a unified clock reference to align the timestamps of all heterogeneous data sources, and use linear interpolation to complete the missing environmental parameters or business load data within the preset aggregation window, generating a standardized multi-dimensional time-series data stream.

[0015] As a further aspect of the present invention, step S2 specifically comprises:

[0016] S21: Set the time span of the aggregation window to regulate the standardized multi-dimensional time-series data stream and ensure that all data points are mapped to the same time scale when calculating total energy consumption;

[0017] S22: Traverse all the real-time current and voltage values ​​in the aggregation window and calculate their product to obtain the instantaneous power. Add the instantaneous power and the input and output power of the power module to calculate the total energy consumption of the data center.

[0018] S23: Filter the operating power of core components related to IT equipment in the standardized multi-dimensional time-series data stream, and perform a summation operation on the operating power of the core components in the aggregation window to obtain the energy consumption value of IT equipment.

[0019] As a further aspect of the present invention, step S3 specifically comprises:

[0020] S31: Using the total energy consumption of the data center as the dividend and the energy consumption of the IT equipment as the divisor, perform standard division to obtain a preliminary power efficiency value.

[0021] S32: Calculate the load-environment correction factor based on the environmental parameters and the service load data, using the formula:

[0022] ;

[0023] The power efficiency value is dynamically calibrated to generate the optimal power efficiency value;

[0024] in, This represents the optimal power efficiency value. This represents the total energy consumption of the data center. Represents the energy consumption of IT equipment. , These represent the preset weighting coefficients for workload and air temperature, respectively. Represents the normalized current business load. This represents the normalized baseline workload. Represents the normalized current air temperature. This represents the normalized reference air temperature.

[0025] As a further aspect of the present invention, step S4 specifically comprises:

[0026] S41: Using the aggregation window as a unit, combine the total energy consumption value of the data center, the optimal power efficiency value, the number of requests, and the total number of concurrent connections by column, and associate them with the timestamp to construct a historical energy consumption sequence;

[0027] S42: The historical energy consumption sequence is processed using the Z-Score normalization method to eliminate the dimensional differences between differential features, and the historical energy consumption sequence is divided into a training set and a validation set. The training set is used for the initial weight training of the long short-term memory network.

[0028] S43: Input the training set into the long short-term memory network for iterative regression training, use the validation set to evaluate the convergence and prediction accuracy of the network model, and solidify the network model parameters when the prediction accuracy meets the preset threshold.

[0029] S44: Input the latest time-series data slice into the long short-term memory network with the network model parameters already fixed, perform regression prediction operation, and obtain the future energy consumption expectation curve.

[0030] As a further aspect of the present invention, the alignment and data completion process of the timestamp in S13 includes:

[0031] The Network Time Protocol (NTP) service is used to force time synchronization of all data acquisition endpoints to ensure that the timestamp error of the real-time current and voltage values, the power module input and output power, the core component operating power, the environmental parameters, and the service load data is less than a preset 10 milliseconds.

[0032] Invalid data points or null points in the standardized multidimensional time-series data stream are detected, and cubic spline interpolation algorithm is applied to replace the linear interpolation method to smoothly complete the continuously missing environmental parameters within the aggregation window.

[0033] As a further aspect of the present invention, the step of setting the temporal span of the aggregation window in S21 specifically includes:

[0034] Monitor the real-time fluctuation variance of the number of requests in the business load data to determine whether the data center is in a peak or off-peak period.

[0035] When the real-time fluctuation variance is greater than the preset high fluctuation threshold, the time span of the aggregation window is dynamically shortened to one minute to capture the details of rapid changes in energy consumption.

[0036] When the real-time fluctuation variance is less than the preset low fluctuation threshold, the time span of the aggregation window is dynamically extended to ten minutes to smooth out energy consumption spikes and reduce the computational complexity of the long short-term memory network, thereby obtaining the time span.

[0037] As a further aspect of the present invention, the process of inputting the historical energy consumption sequence into the long short-term memory network for regression prediction includes:

[0038] Construct a time sliding window and extract the total energy consumption value of the data center, the optimal power efficiency value, the number of requests, and the total number of concurrent connections for the past N time steps from the historical energy consumption sequence processed by the Z-Score normalization method to form an input feature vector;

[0039] The input feature vector is fed into the input layer of the long short-term memory network. Through the synergistic effect of the forget gate, input gate, and output gate, the cell state is updated and time-dependent features are extracted.

[0040] The output of the hidden layer of the Long Short-Term Memory network is passed to the fully connected layer, which maps the high-dimensional features to the energy consumption prediction values ​​for the next M time steps. The energy consumption prediction values ​​are combined to obtain the expected future energy consumption curve.

[0041] A data center PUE real-time monitoring system is provided, the data center PUE real-time monitoring system being used to implement the above-mentioned data center PUE real-time monitoring method, the system comprising:

[0042] The multi-source data acquisition module is used to acquire real-time current and voltage values, power module input and output power, core component operating power, environmental parameters, and service load data of the intelligent power distribution unit. The environmental parameters consist of air temperature, humidity, and flow rate values, and the service load data consists of the number of requests and the total number of concurrent connections, all of which are associated with timestamps. The module then transmits the real-time current and voltage values, power module input and output power, and core component operating power to the sub-item energy consumption calculation module, and the service load data to the energy consumption trend prediction module.

[0043] The component energy consumption calculation module receives the real-time current and voltage values, the power module input and output power, and the core component operating power. It sets an aggregation window, maps the real-time current and voltage values, power module input and output power, and core component operating power based on the timestamp, sums the real-time current and voltage values ​​and the power module input and output power within the aggregation window to obtain the total energy consumption value of the data center, sums the core component operating power to obtain the IT equipment energy consumption value, and transmits the total data center energy consumption value and the IT equipment energy consumption value to the PUE performance evaluation module. Finally, it transmits the total data center energy consumption value to the energy consumption trend prediction module.

[0044] The PUE performance evaluation module is used to receive the total energy consumption value of the data center and the energy consumption value of the IT equipment, obtain the power usage efficiency value by dividing the total energy consumption value of the data center by the energy consumption value of the IT equipment, and transmit the power usage efficiency value to the energy consumption trend prediction module.

[0045] The energy consumption trend prediction module is used to receive the total energy consumption value of the data center, the optimal power efficiency value, and the business load data, store the total energy consumption value of the data center, the optimal power efficiency value, and the business load data in time sequence to construct a historical energy consumption sequence, input the historical energy consumption sequence into the long short-term memory network for regression prediction, and obtain the expected future energy consumption curve.

[0046] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0047] In this invention, by acquiring current and voltage values, power module input and output power, and core component operating power in real time, a precise mapping between total data center energy consumption and IT equipment energy consumption is established. This overcomes the problem of one-sided PUE results caused by the limitations of traditional measurement points, and obtains instantaneous and precise power utilization efficiency. Furthermore, by associating energy consumption values ​​with business load data to construct historical energy consumption sequences, and utilizing the regression prediction capabilities of time-series data, a future energy consumption expectation curve is output. This solves the deficiency of existing technologies that can only perform post-event statistics and cannot predict trends, and provides decision support for proactive energy efficiency optimization. Attached Figure Description

[0048] Figure 1 This is the overall flowchart of the data center PUE real-time monitoring method of the present invention;

[0049] Figure 2 This is a flowchart of the multidimensional time-series data acquisition process of the present invention;

[0050] Figure 3 This is a flowchart of the energy consumption numerical calculation process of the present invention;

[0051] Figure 4 This is a flowchart of the power supply efficiency calibration process of the present invention;

[0052] Figure 5 This is a flowchart illustrating the energy consumption prediction curve of this invention. Detailed Implementation

[0053] 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.

[0054] 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.

[0055] Please see Figure 1 and Figure 2 This invention provides a technical solution: a method for real-time monitoring of PUE in a data center, comprising the following steps:

[0056] S1: Obtain real-time current and voltage values ​​of the intelligent power distribution unit, input and output power of the power module, operating power of core components, environmental parameters, and business load data. Environmental parameters consist of air temperature, humidity, and flow rate values, while business load data consists of the number of requests and the total number of concurrent connections, all of which are associated with timestamps.

[0057] Environmental parameters include air temperature, humidity, and airflow rate values; business load data includes the number of requests and the total number of concurrent connections.

[0058] The specific steps of S1 are as follows:

[0059] S11: Periodically poll the intelligent power distribution unit through the Modbus-TCP protocol interface to obtain real-time current and voltage values ​​associated with timestamps, and simultaneously capture the SNMPMIB library data of the power module to obtain the input and output power of the power module.

[0060] S12: Deploy distributed sensor nodes to collect air temperature, humidity, and flow rate values ​​to form environmental parameters, and connect to the application layer gateway to parse the number of requests and the total number of concurrent connections included in the business load data;

[0061] S13: Establish a unified clock reference to align the timestamps of all heterogeneous data sources, and use linear interpolation to complete the missing environmental parameters or business load data in the preset aggregation window to generate a standardized multi-dimensional time-series data stream.

[0062] The timestamp alignment and data completion process in S13 includes:

[0063] The network time protocol service is used to force time synchronization of all data acquisition endpoints to ensure that the timestamp error of real-time current and voltage values, power module input and output power, core component operating power, environmental parameters and business load data is less than the preset 10 milliseconds.

[0064] Invalid or null data points are detected in a standardized multidimensional time-series data stream. A cubic spline interpolation algorithm is applied to replace the linear interpolation method to smoothly complete the continuously missing environmental parameters within a preset aggregation window.

[0065] The system periodically polls the intelligent power distribution unit via the Modbus-TCP protocol interface to obtain real-time current and voltage values ​​associated with timestamps, and simultaneously captures SNMPMIB library data from the power module to obtain the power module's input and output power. Distributed sensor nodes are deployed to collect air temperature, humidity, and flow rate values ​​to form environmental parameters, and these parameters are connected to the application layer gateway to parse the number of requests and the total number of concurrent connections included in the business load data. A unified clock reference is established to align the timestamps of all heterogeneous data sources, and linear interpolation is used to complete the missing environmental parameters or business load data within the preset aggregation window, generating a standardized multidimensional time-series data stream.

[0066] The data acquisition controller initializes the Modbus-TCP protocol stack and establishes a TCP connection with the intelligent power distribution unit (PDU), setting the target port to 502. The controller sends a request message with function code 03 (read holding register) to the PDU, reads register data from address 0x1000 to 0x100F, parses the returned hexadecimal byte stream, and obtains the real-time voltage (volts) and current (amperes) values ​​for phases A, B, and C. Simultaneously, the controller sends a Get-Request message to the power module via the SNMPv2c protocol to query the node with OID 1.3.6.1.4.1.2021.1.1 to obtain the power input power (watts) and output power (watts). Upon acquiring these power parameters, the controller records the system kernel time as the raw timestamp. .

[0067] Meanwhile, distributed sensor nodes (ZigBee terminals) deployed in the cold and hot aisles of the server racks collect environmental data every second. The sensors read the resistance changes of thermistors and convert them into air temperature values ​​(degrees Celsius), read the capacitance values ​​of capacitive humidity sensors and convert them into relative humidity values ​​(percentages), and read the voltage output of hot-wire anemometers and convert them into airflow speed values ​​(meters per second). These environmental parameters are aggregated through a gateway and timestamped with the gateway's timestamp. The application-layer gateway captures the inbound traffic of the switch using port mirroring technology, parses HTTP request headers using deep packet inspection (DPI) technology, counts the number of HTTP GET / POST requests per second as the request count, and counts the number of session entries with the TCP flag set to ESTABLISHED as the total number of concurrent connections.

[0068] The data processing server runs an NTP (Network Time Protocol) client, specifying a Stratum-1 level atomic clock server as the upstream time source. It performs a clock smoothing calibration every 60 seconds to keep the deviation between the local system clock and standard UTC time within 1 millisecond. After receiving power parameters, environmental parameters, and business load data, the server performs a forced time synchronization operation. If the timestamp of a data packet... Compared with the reference time deviation If the time is less than a millisecond, the packet is discarded and marked as a packet loss point.

[0069] For data that passes validity validation, the server constructs a timeline with a granularity of 100 milliseconds. When a point on the timeline is detected... The environmental parameters or business load data are missing, and the valid data points adjacent to the missing point are... and When linear interpolation is performed, if more than three consecutive time points of data (environmental parameters or business load data) are missing (i.e., there are missing data points), then linear interpolation is performed. If the value is empty, it is determined to be a non-linear drift interval, and the cubic spline interpolation algorithm is called. The server selects two valid data points before and after the missing interval (a total of four points) to construct a cubic polynomial function. The polynomial coefficients were determined by solving the three-moment equation system, and the fill-in values ​​for missing points were calculated to complete the standardization process of the data stream. Table 1 shows a partial multidimensional data segment after acquisition and alignment at a certain moment.

[0070] Table 1. Snapshot Table of Multidimensional Time Series Data Acquisition:

[0071] ;

[0072] As shown in Table 1, all physical quantities have been time-aligned and there are no null values, forming the basic input stream for subsequent calculations.

[0073] Please see Figure 1 and Figure 3 S2: Set up an aggregation window, map real-time current and voltage values, power module input and output power, and core component operating power based on timestamps, sum the real-time current and voltage values ​​and power module input and output power in the aggregation window to get the total energy consumption value of the data center, and sum the core component operating power to get the IT equipment energy consumption value.

[0074] The total energy consumption of a data center includes real-time current and voltage values, and the input and output power of power modules. The energy consumption of IT equipment specifically refers to the operating power of core components.

[0075] The specific steps of S2 are as follows:

[0076] S21: Set the time span of the aggregation window to regulate and standardize the multi-dimensional time-series data stream, ensuring that all data points are mapped to the same time scale when calculating total energy consumption;

[0077] The steps for setting the time span of the aggregation window in S21 specifically include:

[0078] Monitor the real-time fluctuation variance of the number of requests in the business load data to determine whether the data center is in a peak or off-peak period.

[0079] When the real-time fluctuation variance exceeds the preset high fluctuation threshold, the time span of the aggregation window is dynamically shortened to one minute to capture the details of rapid changes in energy consumption.

[0080] When the real-time fluctuation variance is less than the preset low fluctuation threshold, the time span of the aggregation window is dynamically extended to ten minutes to smooth out energy consumption spikes and reduce the computational complexity of the long short-term memory network, thereby obtaining the time span.

[0081] S22: Traverse all real-time current and voltage values ​​in the aggregation window and calculate their product to obtain the instantaneous power. Add the instantaneous power and the input and output power of the power module to calculate the total energy consumption of the data center.

[0082] S23: Filter the operating power of core components related to IT equipment in the standardized multi-dimensional time-series data stream, perform a summation operation on the operating power of core components within the preset aggregation window, and obtain the energy consumption value of IT equipment.

[0083] An aggregation window is set up to map real-time current and voltage values, power module input and output power, and core component operating power based on timestamps. The total energy consumption of the data center is obtained by summing the real-time current and voltage values ​​and the power module input and output power within the aggregation window, and the energy consumption of the IT equipment is obtained by summing the core component operating power. The time span of the aggregation window is set to standardize the multi-dimensional time-series data stream, ensuring that all data points are mapped to the same time scale when calculating the total energy consumption. All real-time current and voltage values ​​within the aggregation window are traversed and their products are calculated to obtain the instantaneous power. The instantaneous power is accumulated and summed with the power module input and output power to calculate the total energy consumption of the data center. The operating power of the core components related to the IT equipment in the standardized multi-dimensional time-series data stream is filtered out, and the operating power of the core components within the preset aggregation window is summed to obtain the energy consumption of the IT equipment.

[0084] The data processing engine first initializes a sliding window buffer, requiring the window's duration (time span) to be determined. The system then reads the request count sequence from the business load data within the last 5 minutes. Calculate the real-time volatility variance of the sequence. The calculation process follows the formula ,in This is the sequence mean. The system has a preset low volatility threshold. The threshold value is 50, indicating a high volatility. The threshold is set at 200. These two thresholds are based on statistical analysis of a normal business cycle over a historical month: during the nighttime off-peak period, the mean variance of the request count fluctuation is 30. The value is 50; during stress testing of major promotional events such as "Double Eleven," the mean variance of the request count fluctuation is 180. It is 200.

[0085] Example 1: Assume the current 5 request data points are... ;

[0086] 1. Calculate the mean ;

[0087] 2. Calculate the variance ;

[0088] 3. Judgment: ;

[0089] 4. Decision: The system determines that the current business is in a stable period and dynamically adjusts the time span of the aggregation window. Set to 10 minutes (600 seconds).

[0090] Example 2: If the subsequent intercepted request data points suddenly change... ;

[0091] 1. Calculate the mean ;

[0092] 2. Calculate the variance ;

[0093] 3. Judgment: ;

[0094] 4. Decision: The system determines that the current period is one of drastic business fluctuations and immediately adjusts the aggregation window. Reduced to 1 minute (60 seconds).

[0095] Determine window span (Taking 1 minute as an example) After that, the engine locks all data frames within that window. (Regarding the total energy consumption of the data center...) The calculation involves the system iterating through each power data record within the window. For the first record... Records, extract voltage and current Calculate instantaneous apparent power Suppose at a certain moment... ,but Simultaneously, the input power of the power module at that moment is read. (e.g., 2300W) and output power (For example, 2100W). The system uses the trapezoidal integral method to integrate the power to obtain the energy consumption. However, when there are discrete data points and the sampling frequency is high (e.g., 10Hz), the summation is used directly for approximate calculation. in For the mains power side at time Total input power, The sampling time interval (e.g., 100 milliseconds). The value is determined by the input power of all power modules. Operating power of all refrigeration units (such as air conditioners and chillers) The summation is obtained. In Within a window of seconds, assuming co-sampling Data points ( If the average input power of all power modules in 600 sampling points is... The average operating power of all refrigeration units The average total power within the window. Total energy consumption of this window To standardize the units, convert it to kWh: .

[0096] Energy consumption of IT equipment The calculation: The system uses a device whitelist to filter out the operating power of core components belonging to the IT load (servers, storage, network switches). If the average power of the server cluster within the window is... Average power of network devices Then the total average power of IT Calculated Ultimately, it will and It is then passed to the next processing module.

[0097] Please see Figure 1 and Figure 4 S3: Obtain the power efficiency value by dividing the total energy consumption of the data center by the energy consumption of the IT equipment;

[0098] Power efficiency figures include total energy consumption figures for data centers and energy consumption figures for IT equipment.

[0099] The specific steps for S3 are as follows:

[0100] S31: Use the total energy consumption of the data center as the dividend and the energy consumption of IT equipment as the divisor to perform standard division to obtain a preliminary power efficiency value.

[0101] S32: Calculate the load-environment correction factor based on environmental parameters and service load data, using the formula:

[0102] ;

[0103] The power efficiency values ​​are dynamically calibrated to generate the optimal power efficiency values.

[0104] in, This represents the optimal power efficiency value. This represents the total energy consumption of the data center. Represents the energy consumption of IT equipment. , These represent the preset weighting coefficients for workload and air temperature, respectively. Represents the normalized current business load. This represents the normalized baseline workload. Represents the normalized current air temperature. This represents the normalized reference air temperature.

[0105] The power efficiency value is obtained by dividing the total energy consumption of the data center by the energy consumption of IT equipment; the load-environment correction factor is calculated based on environmental parameters and business load data, and the power efficiency value is dynamically calibrated by formula to generate the optimal power efficiency value.

[0106] The calculation module received and Then, perform basic division operations first:

[0107] Substituting the result into S2: This value is the uncorrected static PUE value.

[0108] Subsequently, the system activates the calibration logic and calls the calibration formula:

[0109] ;

[0110] Parameter details and logic explanation:

[0111] : The optimal power efficiency value, i.e., the corrected PUE;

[0112] : Represents the total energy consumption of the data center;

[0113] : Represents the energy consumption of IT equipment;

[0114] : Business load weighting coefficient. This coefficient reflects the slope of the power supply and distribution system efficiency curve caused by changes in load rate;

[0115] Normalized current business load. Calculation method: ;

[0116] Normalized baseline workload. Typically set as the load rate corresponding to the power system's highest efficiency point;

[0117] Air temperature weighting coefficient. This coefficient reflects the degree of impact of temperature deviation on the coefficient of performance (COP) of the refrigeration system;

[0118] : Normalized current air temperature. Calculation method: ;

[0119] Normalized reference air temperature. Typically set to the ASHRAE recommended normalized value for optimal operating temperature;

[0120] Term in the formula This constitutes a dynamic correction factor. Through multiplication, this factor scales the original PUE based on the degree of deviation of the current operating condition from the optimal operating condition.

[0121] Parameter setting experiment process and example: 1. Determine the weighting coefficients This study employed multiple linear regression analysis on the data center's operational data over the past year. PUE fluctuation was used as the dependent variable, with load factor variation and temperature variation as independent variables. Experimental data showed that, under constant temperature conditions, for every 10% deviation of the load factor from its optimal point, PUE deteriorated by approximately 1.5%. Therefore, the study set... Under constant load, for every 1°C increase in ambient temperature (approximately 2.5% of the measurement range), cooling energy consumption increases by about 0.2% in PUE. After conversion to a normalized factor, the following settings are applied: All of the above settings passed the correlation verification with an R-squared value greater than 0.85.

[0122] Determine the benchmark value According to the power module technical manual, the highest efficiency is achieved at 50% load rate, therefore the setting is... According to ASHRAETC 9.9 standard, the recommended inlet air temperature is 22℃, and the warning upper limit is 40℃. .

[0123] Substitute actual data into the calculation:

[0124] Current number of concurrent connections The system is designed to have a maximum number of connections. ,but ;

[0125] Current average temperature upper limit of warning temperature ,but ;

[0126] The foregoing ;

[0127] Weight ;

[0128] Substitute the values ​​into the formula:

[0129] ;

[0130] ;

[0131] ;

[0132] ;

[0133] ;

[0134] Explanation of the formula's innovativeness: The advantage of this formula lies in the introduction of... and The correction term expands the simple energy consumption ratio into a multi-dimensional evaluation index that incorporates load characteristics and environmental stress. In this example, the results... The results indicate that although the calculated PUE for the current instantaneous energy consumption is 1.50, the system actually suffers additional energy efficiency penalties considering that the current load is slightly higher than the optimal load point (0.6 > 0.5) and the temperature is higher than the recommended value (26℃ > 22℃). This corrected value more accurately characterizes the energy efficiency degradation trend of infrastructure under non-ideal operating conditions.

[0135] Please see Figure 1 and Figure 5 S4: Store the total energy consumption of the data center, the optimal power efficiency, and the business load data in time sequence to construct a historical energy consumption sequence. Input the historical energy consumption sequence into the long short-term memory network for regression prediction to obtain the expected future energy consumption curve.

[0136] Historical energy consumption data includes total data center energy consumption, power efficiency, and workload data.

[0137] The specific steps for S4 are as follows:

[0138] S41: Using an aggregated window as a unit, combine the total energy consumption of the data center, the optimal power efficiency, the number of requests, and the total number of concurrent connections by column, and associate them with timestamps to construct a historical energy consumption sequence.

[0139] S42: The Z-Score normalization method is used to process the historical energy consumption sequence, eliminate the dimensional differences between differential features, and divide the historical energy consumption sequence into a training set and a validation set. The training set is used for the initial weight training of the long short-term memory network.

[0140] S43: Input the training set into the Long Short-Term Memory network for iterative regression training, use the validation set to evaluate the convergence and prediction accuracy of the network model, and solidify the network model parameters when the prediction accuracy meets the preset threshold.

[0141] S44: Input the latest time series data slices into the long short-term memory network with solidified network model parameters, perform regression prediction operation, and obtain the future energy consumption expectation curve;

[0142] The process of inputting historical energy consumption sequences into a long short-term memory network for regression prediction includes:

[0143] Construct a time sliding window to extract the total energy consumption of the data center over the past N time steps, the optimal power efficiency, the number of requests, and the total number of concurrent connections from the historical energy consumption sequence processed by the Z-Score normalization method, and form the input feature vector;

[0144] The input feature vector is fed into the input layer of the Long Short-Term Memory network. Through the synergistic effect of the forget gate, input gate and output gate, the cell state is updated and time-dependent features are extracted.

[0145] The output of the hidden layer of the Long Short-Term Memory network is passed to the fully connected layer. The fully connected layer maps the high-dimensional features to the energy consumption prediction values ​​for the next M time steps. The combined energy consumption prediction values ​​yield the expected energy consumption curve for the future.

[0146] The historical energy consumption sequence is constructed by storing the total energy consumption of the data center in time sequence, the optimal power efficiency, and the business load data. The historical energy consumption sequence is then input into the long short-term memory network for regression prediction to obtain the expected future energy consumption curve.

[0147] The data management module first restructures the data output from S3 using the aggregation window (e.g., 10 minutes) determined in S2 as the unit, constructing a feature matrix. Each line represents a time step. The column vectors include:

[0148] .

[0149] For example, selecting the past Data at each time step. Due to... The unit is kWh (order of magnitude) The number of requests is measured in units of times per second (on the order of magnitude). The dimensions of these features differ greatly, and direct input can lead to vanishing or exploding gradients. Therefore, the system performs Z-score normalization on each column of features separately.

[0150] Z-Score Standardization Example: For a sequence of requests Assuming the sample mean Standard deviation For the requested value at a certain moment .

[0151] Calculate standardized values Similarly, for Given a sequence (mean 1.5, standard deviation 0.1), if the value at a certain time is 1.5345, then the standardized value is... The processed matrix Most of the elements are distributed in Within the range.

[0152] The dataset was split in an 8:2 ratio, with the first 800 time steps serving as the training set and the last 200 as the validation set. The system constructed a Long Short-Term Memory (LSTM) network, with 4 nodes in the input layer (corresponding to 4 features), 2 stacked LSTM layers in the hidden layer, 64 neurons in each layer, and 1 node in the output layer (predicting future energy consumption).

[0153] LSTM Input and Processing: Construct a time sliding window, with the window length set to... That is, the input feature vector. Includes from arrive Ten sets of standardized data at each moment. After being fed into the network, the LSTM unit performs the following gating operations: 1. Forget gate: 1. Decide which old historical energy consumption information to discard. 2. Input Gate: 3. Cell Status Update: This determines which new load fluctuation information to update. 4. Output gate: And calculate the current hidden state. .

[0154] Training and Parameter Fixation: The loss function is set to Mean Squared Error (MSE), the optimizer is Adam, and the learning rate is set to 0.001. Iterative training begins, and the MSE on the validation set is calculated after each epoch (full dataset traversal). A convergence threshold is set. Experimental data verification: Table 2 shows the changes in loss during model training.

[0155] Table 2: LSTM Model Training Convergence Process

[0156] ;

[0157] When the validation set loss drops to 0.009, which is less than the preset threshold of 0.01, the system determines that the model has converged, immediately stops training, and solidifies all weight matrices. and bias vector .

[0158] Regression Prediction Example: Prediction using a fixed model. The standardized data from the 10 most recent time steps are extracted as the input vector. . After network forward propagation, the fully connected layer outputs a scalar value. (Standardized predicted values). Perform inverse normalization to restore physical dimensions: predict energy consumption. .like mean Standard deviation .but The system outputs this value as the expected energy consumption point for the next time window, and connects multiple consecutive points to form a future energy consumption expectation curve.

[0159] A data center PUE real-time monitoring system is provided for executing the above-described data center PUE real-time monitoring method. The system includes:

[0160] The multi-source data acquisition module is used to acquire real-time current and voltage values, power module input and output power, core component operating power, environmental parameters, and business load data of the intelligent power distribution unit. The environmental parameters consist of air temperature, humidity, and flow rate values, and the business load data consists of the number of requests and the total number of concurrent connections. All of these are associated with timestamps. The real-time current and voltage values, power module input and output power, and core component operating power are transmitted to the sub-item energy consumption calculation module, and the business load data is transmitted to the energy consumption trend prediction module.

[0161] The sub-item energy consumption calculation module receives real-time current and voltage values, power module input and output power, and core component operating power. It sets an aggregation window, maps real-time current and voltage values, power module input and output power, and core component operating power based on timestamps, sums the real-time current and voltage values ​​and power module input and output power within the aggregation window to obtain the total energy consumption value of the data center, sums the core component operating power to obtain the IT equipment energy consumption value, and transmits the total data center energy consumption value and IT equipment energy consumption value to the PUE performance evaluation module and the total data center energy consumption value to the energy consumption trend prediction module.

[0162] The PUE performance assessment module receives the total energy consumption value of the data center and the energy consumption value of IT equipment. By dividing the total energy consumption value of the data center by the energy consumption value of IT equipment, the power usage efficiency value is obtained, and the power usage efficiency value is transmitted to the energy consumption trend prediction module.

[0163] The energy consumption trend prediction module receives the total energy consumption value of the data center, the optimal power efficiency value, and the business load data. It stores the total energy consumption value, the optimal power efficiency value, and the business load data in time sequence to construct a historical energy consumption sequence. The historical energy consumption sequence is then input into the Long Short-Term Memory network for regression prediction to obtain the expected future energy consumption curve.

[0164] 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 scope of protection defined by the claims of the present invention.

Claims

1. A method for real-time monitoring of PUE in a data center, characterized in that, Includes the following steps: S1: Obtain real-time current and voltage values, power module input and output power, core component operating power, environmental parameters, and service load data of the intelligent power distribution unit. The environmental parameters consist of air temperature, humidity, and flow rate values, and the service load data consists of the number of requests and the total number of concurrent connections, both of which are associated with timestamps. S2: Set an aggregation window, map the real-time current and voltage values, the power module input and output power, and the core component operating power based on the timestamp, sum the real-time current and voltage values ​​and the power module input and output power in the aggregation window to obtain the total energy consumption value of the data center, and sum the core component operating power to obtain the IT equipment energy consumption value. S3: Obtain the power efficiency value by dividing the total energy consumption of the data center by the energy consumption of the IT equipment; The specific steps of S3 are as follows: S31: Using the total energy consumption of the data center as the dividend and the energy consumption of the IT equipment as the divisor, perform standard division to obtain a preliminary power efficiency value. S32: Calculate the load-environment correction factor based on the environmental parameters and the service load data, using the formula: ； The power efficiency value is dynamically calibrated to generate the optimal power efficiency value; in, This represents the optimal power efficiency value. This represents the total energy consumption of the data center. Represents the energy consumption of IT equipment. and These represent the preset weighting coefficients for workload and air temperature, respectively. Represents the normalized current business load. This represents the normalized baseline workload. Represents the normalized current air temperature. Represents the normalized reference air temperature; S4: Store the total energy consumption value of the data center and the optimal power efficiency value in time sequence, and construct a historical energy consumption sequence with the business load data. Input the historical energy consumption sequence into the long short-term memory network for regression prediction to obtain the expected future energy consumption curve.

2. The data center PUE real-time monitoring method according to claim 1, characterized in that, The specific steps of S1 are as follows: S11: Periodically poll the intelligent power distribution unit through the Modbus-TCP protocol interface to obtain real-time current and voltage values ​​associated with timestamps, and simultaneously capture the SNMPMIB library data of the power module to obtain the input and output power of the power module. S12: Deploy distributed sensor nodes to collect air temperature, humidity, and flow rate values ​​to form environmental parameters, and connect to the application layer gateway to parse the number of requests and the total number of concurrent connections included in the business load data; S13: Establish a unified clock reference to align the timestamps of all heterogeneous data sources, and use linear interpolation to complete the missing environmental parameters or business load data within the preset aggregation window, generating a standardized multi-dimensional time-series data stream.

3. The data center PUE real-time monitoring method according to claim 2, characterized in that, The specific steps of S2 are as follows: S21: Set the time span of the aggregation window to regulate the standardized multi-dimensional time-series data stream and ensure that all data points are mapped to the same time scale when calculating total energy consumption; S22: Traverse all the real-time current and voltage values ​​in the aggregation window and calculate their product to obtain the instantaneous power. Add the instantaneous power and the input and output power of the power module to calculate the total energy consumption of the data center. S23: Filter the operating power of core components related to IT equipment in the standardized multi-dimensional time-series data stream, and perform a summation operation on the operating power of the core components in the aggregation window to obtain the energy consumption value of IT equipment.

4. The data center PUE real-time monitoring method according to claim 1, characterized in that, The specific steps of S4 are as follows: S41: Using the aggregation window as a unit, combine the total energy consumption value of the data center, the optimal power efficiency value, the number of requests, and the total number of concurrent connections by column, and associate them with the timestamp to construct a historical energy consumption sequence; S42: The historical energy consumption sequence is processed using the Z-Score normalization method to eliminate the dimensional differences between differential features, and the historical energy consumption sequence is divided into a training set and a validation set. The training set is used for the initial weight training of the long short-term memory network. S43: Input the training set into the long short-term memory network for iterative regression training, use the validation set to evaluate the convergence and prediction accuracy of the network model, and solidify the network model parameters when the prediction accuracy meets the preset threshold. S44: Input the latest time-series data slice into the long short-term memory network with the network model parameters already fixed, perform regression prediction operation, and obtain the future energy consumption expectation curve.

5. The data center PUE real-time monitoring method according to claim 2, characterized in that, The alignment and data completion process for the timestamp in S13 includes: The Network Time Protocol (NTP) service is used to force time synchronization of all data acquisition endpoints to ensure that the timestamp error of the real-time current and voltage values, the input and output power of the power module, the operating power of the core components, the environmental parameters, and the business load data is less than a preset 10 milliseconds. Invalid data points or null points in the standardized multidimensional time-series data stream are detected, and cubic spline interpolation algorithm is applied to replace the linear interpolation method to smoothly complete the continuously missing environmental parameters within the aggregation window.

6. The data center PUE real-time monitoring method according to claim 3, characterized in that, The step of setting the temporal span of the aggregation window in S21 specifically includes: Monitor the real-time fluctuation variance of the number of requests in the business load data to determine whether the data center is in a peak or off-peak period. When the real-time fluctuation variance is greater than the preset high fluctuation threshold, the time span of the aggregation window is dynamically shortened to one minute to capture the details of rapid changes in energy consumption. When the real-time fluctuation variance is less than the preset low fluctuation threshold, the time span of the aggregation window is dynamically extended to ten minutes to smooth out energy consumption spikes and reduce the computational complexity of the long short-term memory network, thereby obtaining the time span.

7. The data center PUE real-time monitoring method according to claim 4, characterized in that, The process of inputting the historical energy consumption sequence into the long short-term memory network regression prediction includes: Construct a time sliding window and extract the total energy consumption value of the data center, the optimal power efficiency value, the number of requests, and the total number of concurrent connections for the past N time steps from the historical energy consumption sequence processed by the Z-Score normalization method to form an input feature vector; The input feature vector is fed into the input layer of the long short-term memory network. Through the synergistic effect of the forget gate, input gate, and output gate, the cell state is updated and time-dependent features are extracted. The output of the hidden layer of the Long Short-Term Memory network is passed to the fully connected layer, which maps the high-dimensional features to the energy consumption prediction values ​​for the next M time steps. The energy consumption prediction values ​​are combined to obtain the expected future energy consumption curve.

8. A real-time PUE monitoring system for a data center, characterized in that, The system is used to implement the real-time PUE monitoring method for data centers according to any one of claims 1-7, the system comprising: The multi-source data acquisition module is used to acquire real-time current and voltage values, power module input and output power, core component operating power, environmental parameters, and service load data of the intelligent power distribution unit. The environmental parameters consist of air temperature, humidity, and flow rate values, and the service load data consists of the number of requests and the total number of concurrent connections, all of which are associated with timestamps. The module then transmits the real-time current and voltage values, the power module input and output power, and the core component operating power to the sub-item energy consumption calculation module, and the service load data to the energy consumption trend prediction module. The component energy consumption calculation module receives the real-time current and voltage values, the power module input and output power, and the core component operating power. It sets an aggregation window, maps the real-time current and voltage values, power module input and output power, and core component operating power based on the timestamp, sums the real-time current and voltage values ​​and the power module input and output power within the aggregation window to obtain the total energy consumption value of the data center, sums the core component operating power to obtain the IT equipment energy consumption value, and transmits the total data center energy consumption value and the IT equipment energy consumption value to the PUE performance evaluation module. Finally, it transmits the total data center energy consumption value to the energy consumption trend prediction module. The PUE performance evaluation module is used to receive the total energy consumption value of the data center and the energy consumption value of the IT equipment, obtain the power usage efficiency value by dividing the total energy consumption value of the data center by the energy consumption value of the IT equipment, calculate the load-environment correction factor based on environmental parameters and business load data, dynamically calibrate the power usage efficiency value through a formula, generate the optimal power usage efficiency value, and transmit the optimal power usage efficiency value to the energy consumption trend prediction module. The energy consumption trend prediction module is used to receive the total energy consumption value of the data center, the optimal power efficiency value, and the business load data, store the total energy consumption value of the data center, the optimal power efficiency value, and the business load data in time sequence to construct a historical energy consumption sequence, input the historical energy consumption sequence into the long short-term memory network for regression prediction, and obtain the expected future energy consumption curve.

Images