A centralized group control optimization method and system for machine room equipment

Abstract

The application relates to the field of data processing, in particular to a centralized group control optimization method and system for machine room equipment, which comprises the following steps: acquiring total load data of the machine room equipment in time sequence; performing ordered sample clustering on the total load data to obtain a plurality of clustering clusters, each clustering cluster being load data of a time period between two time points, constructing a subsequence according to the load data to obtain a plurality of time periods and a plurality of time period corresponding subsequences; for any time period, calculating the similarity of the subsequence of any historical day and the current day, obtaining effective similar time according to the similarity; calculating the periodicity evaluation of the effective similar time, screening out a credible period as a training set according to the periodicity evaluation, training a neural network to obtain a prediction model to predict the load; dynamically adjusting the air conditioner power based on the predicted load, and migrating the tasks corresponding to the high-load equipment to the low-load equipment for processing. The application reduces the probability of local hot spots, and improves the accuracy and timeliness of equipment task scheduling.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data processing, in particular to a centralized group control optimization method and system for equipment in a computer room. BACKGROUND

[0002] In a data center computer room, the energy consumption and load of key equipment such as servers and air conditioners are driven by multiple factors, including business peak and valley, work calendar, financial nodes, and seasonal climate, showing multi-period fluctuation characteristics. Such load changes not only exist in daily time period differences, weekly differences, and monthly differences, but are also affected by environmental conditions (such as high temperature and humidity in summer leading to an increase in air conditioning energy consumption). However, traditional fixed threshold or simple sliding window group control strategies are difficult to capture these fuzzy periods, resulting in air conditioning refrigeration and task allocation lagging behind real load changes, and failing to achieve fine-tuned advance regulation, thereby causing energy waste or local hotspot risks.

[0003] The main problem of existing centralized group control methods is that they cannot accurately identify the fuzzy period boundaries of the computer room load in daily, weekly, monthly, and seasonal scales, lack quantitative evaluation of the coupling relationship between period reliability and refrigeration efficiency, resulting in blind selection of the training set of the prediction model, insufficient prediction accuracy, and ultimately leading to lag of air conditioning power regulation and task scheduling strategies, making it difficult to reduce energy use efficiency and energy consumption while ensuring equipment reliability. SUMMARY

[0004] To solve the technical problem that the existing centralized group control method cannot accurately identify the fuzzy period boundaries of the computer room load and quantitatively evaluate the coupling relationship between period reliability and refrigeration efficiency, resulting in insufficient prediction model accuracy and lag of the regulation strategy, and making it difficult to achieve fine-tuned energy optimization and task scheduling, the present application provides a centralized group control optimization method and system for equipment in a computer room.

[0005] In a first aspect, the present application provides a centralized group control optimization method for equipment in a computer room, which adopts the following technical solution:

[0006] A centralized group control optimization method for equipment in a computer room, comprising the steps of:

[0007] Obtaining total load data of the computer room equipment in time sequence; performing ordered sample clustering on the total load data to obtain a plurality of clustering clusters, each clustering cluster being load data of a time period between two time points, constructing a subsequence according to the load data to obtain a plurality of time periods and a plurality of subsequences corresponding to the time periods; for any time period, calculating the similarity of the subsequence of any historical day and the current day, obtaining an effective similar time according to the similarity; calculating the periodicity evaluation of the effective similar time, screening out a reliable period as a training set according to the periodicity evaluation, and using the training set to train a neural network to obtain a prediction model to predict the load; dynamically adjusting the air conditioning power based on the predicted load, and migrating tasks corresponding to high-load equipment to low-load equipment for processing.

[0008] Optionally, for any time period, the similarity between the historical any day and the current day sub-sequence is calculated, including the steps of: calculating the size matching degree and the trend similarity of the sub-sequence of the historical time period and the sub-sequence of the current time period, the historical time period being the historical time period of the historical any day, and the current time period being the current time period of the current day; and taking the product of the trend similarity and the size matching degree as the similarity of the sub-sequence.

[0009] Optionally, the calculation method of the size matching degree is: calculating the ratio of the mean values of the corresponding sub-sequences of the historical time period and the current time period; calculating the absolute difference value of the ratio and 1; and taking the difference value between 1 and the absolute difference value as the size matching degree.

[0010] Optionally, the calculation method of the trend similarity is: taking the Pearson correlation coefficient of the sub-sequence of the historical time period and the sub-sequence of the current time period as the trend similarity.

[0011] Optionally, the effective similar time is obtained according to the similarity, including the steps of: obtaining the similarity of the historical days and the current day, obtaining a similarity sequence, performing spline curve fitting on the similarity sequence to obtain a mapping function of the similarity sequence; calculating the second derivative of the mapping function, finding the day corresponding to the point where the second derivative of the mapping function is 0, and taking the time between the nearest day and the current day as the effective similar time.

[0012] Optionally, the method for obtaining the effective similar time according to the similarity is: in response to the similarity being less than a preset similarity threshold, taking the time between the historical time period and the current time period as the effective similar time, and fitting the sub-sequence in the effective similar time to obtain the mapping function.

[0013] Optionally, the periodicity evaluation of the effective similar time is calculated, and the calculation method of the periodicity evaluation is: in any time period, calculating a weekly fluctuation evaluation, constructing a weekly fluctuation evaluation sequence, and taking the product of the negative correlation normalization result of the standard deviation of the weekly fluctuation evaluation sequence and the positive correlation normalization result of the effective similar time as the periodicity evaluation.

[0014] Optionally, the calculation method of the weekly fluctuation evaluation is: calculating the distance between each load data of the sub-sequence of the historical time period and the mapping function, and taking the cumulative sum of each distance as the weekly fluctuation evaluation.

[0015] In a second aspect, the application provides a centralized group control optimization system for equipment rooms, which adopts the following technical scheme:

[0016] A centralized group control optimization system for equipment rooms, a processor and a memory, the memory storing computer program instructions, which, when executed by the processor, implement the centralized group control optimization method for equipment rooms described above.

[0017] The beneficial effect is that the computer program generated by the centralized group control optimization method of the machine room equipment is stored in the memory to be loaded and executed by the processor, so that the system is made according to the memory and the processor, and the use is facilitated.

[0018] The present application has the following technical effects:

[0019] 1. By ordering sample clustering, similarity calculation and periodic evaluation of load data, the fuzzy period boundary of the load can be accurately identified, providing a reliable data basis for subsequent air conditioning power regulation and task scheduling, thereby realizing fine energy consumption optimization. By screening out the reliable period as the training set to train the neural network, the prediction model obtained has high predictability for the actual load situation, avoiding the shortcomings of traditional fixed threshold or simple sliding window strategy, and significantly improving the accuracy of the prediction model.

[0020] 2. By predicting the load in advance and adjusting the refrigeration power, temperature fluctuations or energy waste caused by lag control are avoided, and the goal of reducing energy use efficiency and energy consumption while ensuring equipment reliability is achieved.

[0021] 3. Based on the predicted load, the air conditioning power is dynamically adjusted, and the task is migrated and redistributed to ensure more accurate and timely task scheduling, improving overall operation efficiency. By migrating high-load device corresponding tasks to low-load device processing, the occurrence of local hot spot problems is effectively avoided, improving the stability of the system and the service life of the equipment. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 is a method flowchart of a centralized group control optimization method of machine room equipment according to an embodiment of the present application. DETAILED DESCRIPTION

[0023] An embodiment of the present application discloses a centralized group control optimization method of machine room equipment, referring to Figure 1 , including steps S1-S5, as follows:

[0024] S1: Obtain the total load data of the machine room equipment in time sequence.

[0025] The total load data can be the load data of the equipment room device in time sequence, such as the CPU usage rate, memory usage rate, bandwidth usage rate of the server, and refrigeration power of the air conditioning system. These load data are affected by business rules and environmental changes, and show periodic characteristics in the intra-day, intra-week, intra-month, and seasonal scales. For example, the load significantly increases during the business peak period (such as 9:00-18:00), and decreases to a low point during the late night to early morning (0:00-6:00); the load is continuously high during weekdays, and decreases by 30%-50% during weekends; the load of the database and storage changes due to the financial settlement at the beginning of the month and the data archiving at the end of the month; and the air conditioning energy consumption ratio increases in the summer due to the high temperature and high humidity environment.

[0026] Due to different strategies of the load, the periodicity of the equipment room device does not completely occur at fixed time nodes, that is, both the peak and the valley are a fuzzy time range. Therefore, when the centralized group control of the equipment room device is performed, the periodicity of the load needs to be monitored.

[0027] In an embodiment, the collected total load data (such as the CPU usage rate, memory usage rate, bandwidth usage rate, and air conditioning refrigeration power) of the equipment room device is preprocessed, which is denoising and smoothing. The prior art will not be described again.

[0028] S2: The total load data is subjected to ordered sample clustering to obtain a plurality of clustering clusters, each clustering cluster is the load data of a period between two time points, a subsequence is constructed according to the load data, and a plurality of periods and a plurality of subsequences corresponding to the periods are obtained.

[0029] The total load data in a day is divided into a plurality of clustering clusters, each clustering cluster corresponds to a period of time-continuous load data, which is recorded as a subsequence, and each subsequence represents the load change of a period (for example, from the time point to the time point).

[0030] The purpose of clustering is to divide the total load data in a day into a plurality of subsequences with similar characteristics according to time sequence, so as to analyze the periodicity of the load subsequently. The ordered sample clustering method is used for clustering in the present application. The ordered sample clustering is a clustering method specially for time series data, which not only considers the numerical characteristics of the data points, but also retains the time sequence relationship of the data points. This method is particularly suitable for analyzing time series data with periodicity or trend, and can divide the total load data in a day into a plurality of subsequences with similar characteristics. The prior art will not be described again.

[0031] S3: For any period, the similarity between the subsequence of any historical day and the subsequence of the current day is calculated, and the effective similar time is obtained according to the similarity.

[0032] In this system, any given time period is defined as a historical time period on any given day and the current time period on the current day. The system calculates the size matching degree and trend similarity between the subsequences of the historical time period and the subsequences of the current time period. The product of the trend similarity and the size matching degree is used as the subsequence similarity. By calculating the similarity, the system can accurately identify the fuzzy periodic boundaries of the load, avoiding misjudgments caused by sudden load surges or environmental changes. This provides a reliable data foundation for subsequent periodic evaluation and prediction model training, thereby enabling refined air conditioning power regulation and task scheduling optimization.

[0033] In one embodiment, the scale matching degree is calculated as follows: the ratio of the mean of the corresponding subsequence of the historical period to that of the current period is calculated; the absolute difference between the ratio and 1 is calculated; and the difference between 1 and the absolute difference is taken as the scale matching degree. When the ratio of the mean deviates from 1, the absolute difference increases and the scale matching degree decreases; when the ratio of the mean approaches 1, the absolute difference decreases and the scale matching degree increases.

[0034] In one embodiment, the trend similarity is calculated by using the Pearson correlation coefficient between the subsequences of historical periods and the subsequences of the current period. The Pearson correlation coefficient is existing technology and will not be elaborated upon here.

[0035] With a first Time to the Taking a subsequence at time step as an example, calculate the similarity of the sequence:

[0036] ;in, To push forward from the current day The first day Time to the The degree of similarity between time subsequences.

[0037] To push forward from the current day The first day Time to the Subsequence of time, The current day's Time to the A subsequence of time points. Factors This indicates the degree of scale matching, calculated from the current day backwards. The first day Time to the Subsequence of time and the current day's number Time to the The case where the subsequences at time points are similar in size to the overall numerical value.

[0038] The Pearson correlation coefficient is used. Calculate the mean. Factors. Indicates trend similarity, calculated from the current day backwards. The first day Time to the Subsequence of time and the current day's number Time to the The degree of similarity in the trends of numerical changes among the subsequences at time points. The closer it is to 1, the more historical subsequences and the current subsequence The more similar their trends are, the less similar they are; conversely, the less similar they are.

[0039] In one embodiment, obtaining the effective similarity time based on similarity includes the following steps: acquiring the similarity between historical days and the current day to obtain a similarity sequence; performing spline curve fitting on the similarity sequence using MATLAB to obtain a mapping function for the similarity sequence; calculating the second derivative of the mapping function; finding the day corresponding to the point where the second derivative of the mapping function is 0; and using the time between the nearest day to the current day and the current day as the effective similarity time. The point where the second derivative is 0 usually corresponds to an inflection point in the similarity sequence, indicating that the similarity begins to change significantly at this point. If the similarity changes significantly on a previous day (i.e., the similarity decreases with distance from the current day), finding the day when the similarity decreases helps avoid dissimilarity caused by a sudden increase in load on a particular day, which could affect the overall prediction. Spline curve fitting can flexibly adapt to changes in different load patterns; this embodiment is suitable for complex multi-cycle load scenarios.

[0040] In other embodiments, the method for obtaining the effective similar time based on similarity can also be as follows: in response to a similarity less than a preset similarity threshold, the time between the historical time period and the current time period is taken as the effective similar time, and a mapping function is fitted to the subsequences within the effective similar time period. For example, the similarity threshold can be 0.7, and the similarity threshold can be adjusted according to the actual application scenario, which will not be elaborated here. Directly determining the similarity between the historical time period and the current time period is suitable for scenarios with high real-time requirements.

[0041] S4: Calculate the periodic evaluation of effective similarity time, select reliable periods based on the periodic evaluation as training sets, and use them to train the neural network to obtain a prediction model to predict the load.

[0042] Based on the effective similarity time obtained in step S3, the periodicity evaluation of the corresponding time period is further calculated to assess whether the load within that time period has stable periodic characteristics. Since the load subsequence may be disturbed in actual operation (such as sudden tasks or environmental changes), resulting in load surges or fluctuations, a comprehensive evaluation in conjunction with cooling efficiency is also required.

[0043] Specifically, at any given time period (e.g., the first...) Time to the (Time), calculate the weekly fluctuation evaluation, construct the weekly fluctuation evaluation sequence, and use the product of the negative correlation normalization result of the standard deviation of the weekly fluctuation evaluation sequence and the positive correlation normalization result of the effective similar time as the periodic evaluation.

[0044] The calculation method for weekly fluctuation evaluation is as follows: calculate the distance between each load data of the subsequence of the historical period and the mapping function, and use the sum of each distance as the weekly fluctuation evaluation.

[0045] Taking a time period as an example, the periodic evaluation satisfies the following polynomial:

[0046] ; In the formula, For the first Time to the Periodic evaluation of time, The distance from the current day The day before yesterday Time to the Weekly fluctuation evaluation at any given time For mapping functions, The distance from the point to the line. for The number of elements in the sequence This is a valid similar time for that time period.

[0047] Set a periodic evaluation threshold; when the... Time to the When the periodic evaluation at time 1 is greater than the periodic evaluation threshold, it indicates that the periodic evaluation at time 2 is greater than the periodic evaluation threshold. Time to the The load at any given time exhibits periodicity. This load period can serve as a training set for training a neural network, and the resulting neural network demonstrates high predictive accuracy for reality. If the load at the given time... Time to the If the periodic evaluation at any given time is less than or equal to the periodic evaluation threshold, the load period at this time cannot be used as the training set for training the neural network, and the resulting neural network will not be predictive of reality. For example, the periodic evaluation threshold can be 0.6. The setting of the periodic evaluation threshold can be adjusted according to the actual application scenario, which will not be elaborated here.

[0048] All loads within the effective similar time period corresponding to the periodic time period are sequentially fed into the LSTM (Long Short-Term Memory) network model for training. Each training sample is a load data sequence within the time period on any day within the effective similar time period. The training completed load prediction network is obtained when the loss is less than the preset value or the number of training iterations is reached.

[0049] For example, the preset value can be 0.01, and the number of training iterations can be 1000. The preset value and the number of training iterations can be adjusted according to the actual application scenario, which will not be elaborated here.

[0050] The load sequence within a real-time time period is fed into a trained LSTM load prediction network to obtain the predicted load size. Based on the predicted load size, the air conditioning power is dynamically adjusted: when the load increases, the cooling power is proactively increased to meet demand; when the load decreases, the cooling power is reduced to save energy. Adjusting the cooling power in advance based on the predicted load avoids temperature fluctuations or energy waste caused by delayed control. Task allocation can be performed based on the predicted load.

[0051] S5: Dynamically adjust air conditioning power based on predicted load, and migrate tasks corresponding to high-load devices to low-load devices for processing.

[0052] Using a trained LSTM load prediction network, the load sequence within a real-time time period is predicted to obtain the load magnitude of each device in the future. By migrating tasks, the load of each device is balanced, avoiding local hotspots and improving overall operating efficiency.

[0053] For each device, calculate the difference between its current load and its maximum load capacity. This difference is the device's spare load, which reflects how many additional tasks the device can still handle.

[0054] Load is allocated based on the real-time unit load and the available load of each device. A device's weight is calculated by dividing its available load by the sum of the available loads of all devices, and the assigned load is the weight multiplied by the predicted load. By rationally allocating tasks, localized hotspots are avoided, improving equipment utilization and system stability. Continuous monitoring of system operation and timely adjustments to strategies ensure efficient operation and energy savings for data center equipment.

[0055] This application also discloses a centralized group control optimization system for data center equipment, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the centralized group control optimization method for data center equipment according to this application is implemented.

[0056] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0057] In this application, the aforementioned memory can be any tangible medium that contains or stores a program that can be used or combined with an instruction execution system, apparatus, or device. For example, a computer-readable storage medium can be any suitable magnetic or magneto-optical storage medium, such as resistive random access memory (DRAM), dynamic random access memory (DRAM), static random access memory (SRAM), etc., or any other medium that can be used to store desired information and can be accessed by an application program, module, or both. Any such computer storage medium can be part of a device or accessible to or connected to a device.

[0058] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for centralized group control optimization of computer room equipment, characterized in that, Including the following steps: Obtain the total load data of the data center equipment in time sequence; The total load data is subjected to ordered sample clustering to obtain multiple clusters. Each cluster contains load data for a time period between two time periods. Subsequences are constructed based on the load data to obtain multiple time periods and their corresponding subsequences. For any given time period, calculate the similarity between any historical day and the current emperor sequence, and obtain the effective similar time based on the similarity. Calculate the periodic evaluation of effective similarity time, select reliable periods based on the periodic evaluation as training set, and use them to train the neural network to obtain a prediction model to predict the load. Based on predicted load, the air conditioning power is dynamically adjusted, and tasks corresponding to high-load devices are migrated to low-load devices for processing. The calculation method for periodic evaluation is as follows: at any time period, calculate the periodic fluctuation evaluation, construct the periodic fluctuation evaluation sequence, and use the product of the negative correlation normalization result of the standard deviation of the periodic fluctuation evaluation sequence and the positive correlation normalization result of the effective similar time as the periodic evaluation. The calculation method for weekly fluctuation evaluation is as follows: calculate the distance between each load data of the subsequence of the historical period and the mapping function, and use the sum of each distance as the weekly fluctuation evaluation.

2. The centralized group control optimization method for computer room equipment according to claim 1, characterized in that, For any given time period, calculate the similarity between any historical day and the current emperor sequence, including the following steps: Any time period is defined as a historical time period on any day in history and as a current time period on the current day. The size matching degree and trend similarity between the subsequences of the historical time period and the subsequences of the current time period are calculated. The product of trend similarity and size matching is used as the similarity of subsequences.

3. The centralized group control optimization method for computer room equipment according to claim 2, characterized in that, The method for calculating the scale matching degree is as follows: calculate the ratio of the mean of the corresponding subsequence of the historical period to that of the current period; Calculate the absolute difference between the ratio and 1; use the difference between 1 and the absolute difference as the scale matching degree.

4. The centralized group control optimization method for computer room equipment according to claim 2, characterized in that, The trend similarity is calculated by using the Pearson correlation coefficient between the subsequence of the historical period and the subsequence of the current period as the trend similarity.

5. The centralized group control optimization method for computer room equipment according to claim 1, characterized in that, The effective similarity time is obtained based on the similarity, including the following steps: Obtain the similarity between each historical day and the current day to get a similarity sequence. Then, perform spline curve fitting on the similarity sequence using MATLAB to obtain the mapping function of the similarity sequence. Calculate the second derivative of the mapping function, find the day corresponding to the point where the second derivative of the mapping function is 0, and take the time between the nearest day to the current day and the current day as the effective similar time.

6. The centralized group control optimization method for computer room equipment according to claim 1, characterized in that, The method for obtaining the effective similar time based on similarity is as follows: in response to the similarity being less than a preset similarity threshold, the time between the historical time period and the current time period is taken as the effective similar time, and a mapping function is obtained by fitting the subsequence within the effective similar time period.

7. A centralized group control and optimization system for computer room equipment, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement the centralized group control optimization method for computer room equipment according to any one of claims 1-6.

Images