Load power consumption prediction method and computer readable storage medium

Abstract

The invention relates to a load power consumption prediction method and a computer readable storage medium, and the method comprises the steps: dividing the operation state of a load into a plurality of types, simulating each type of operation state of the load, and obtaining the power consumption performance data and power consumption data of the load in each type of operation state; all the obtained power consumption performance data are preprocessed, all the preprocessed time feature sequences are sorted according to the sequence from strong fluctuation to weak fluctuation, VMD decomposition and reconstruction are conducted on the first N preprocessed time feature sequences, and finally the reconstructed time feature sequences and other time feature sequences are combined to form a data set; and constructing a load power consumption prediction model, inputting the training set into the constructed load power consumption prediction model in batches, and inputting a test sample selected from the test set into the trained load power consumption prediction model, namely predicting to obtain power consumption data corresponding to the test sample. The method improves the accuracy of load power consumption prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of power consumption preset, and in particular to a load power consumption prediction method and a computer-readable storage medium. Background Art

[0002] With the rapid development of the cloud computing industry, data centers, as important carriers of information, have ushered in a new wave of construction. The rapid expansion has also brought about problems such as operating costs, energy consumption, and environmental protection.

[0003] Building a power consumption prediction model can effectively monitor the power consumption of servers. The power consumption prediction model refers to a function model that maps system state-related variables to system power consumption. It generally includes one or more performance indicators with a certain granularity (such as CPU utilization, instruction cycle). The output of the power consumption prediction model is the system power consumption over a period of time or the power consumption value at a certain moment.

[0004] With the rapid development of computer technology, especially the improvement of computing power, deep learning has received more and more attention and has also been applied to server power consumption modeling in recent years. However, most existing deep learning modeling methods directly normalize the data and then send it to the deep learning network training. They rarely consider the short-term fluctuations in server operation, and consider GPU-intensive and hybrid-intensive tasks less or in a relatively simple manner. Therefore, the accuracy of power consumption prediction in existing technologies is not high. Therefore, further improvements to existing technologies are needed. Summary of the invention

[0005] The first technical problem to be solved by the present invention is to provide a load power consumption prediction method which can improve the power consumption prediction accuracy in view of the above-mentioned prior art.

[0006] The second technical problem to be solved by the present invention is to provide a computer-readable storage medium for running the above-mentioned load power consumption prediction method.

[0007] The technical solution adopted by the present invention to solve the above-mentioned first technical problem is: a load power consumption prediction method, characterized by comprising the following steps:

[0008] Step 1: Obtain the power consumption performance index of each component of the load;

[0009] Step 2: Divide the operating state of the load into multiple types, simulate each type of operating state of the load, and obtain the power consumption performance data and power consumption data of the load in each type of operating state. The power consumption performance data of the load in a certain type of operating state is the time characteristic sequence corresponding to the power consumption performance index of each component of the load;

[0010] Step 3: preprocess all power consumption performance data acquired in step 2 to obtain multiple preprocessed time feature sequences;

[0011] Step 4: Analyze the volatility of each preprocessed time feature sequence in step 3, sort each preprocessed time feature sequence in the order of volatility from strong to weak, and obtain the first N preprocessed time feature sequences after sorting; N is a positive integer;

[0012] Step 5, performing VMD decomposition on the N time feature sequences obtained in step 4 respectively to obtain multiple stable modal components, reconstructing all stable modal components to obtain reconstructed time feature sequences, and combining the reconstructed time feature sequences with other time feature sequences except the sorted first N time feature sequences in step 4 to form a data set, and dividing the data set into a training set and a test set, wherein the training set includes multiple training samples, each training sample includes a group of time feature sequences and power consumption data corresponding to the group of time feature sequences; the test set includes multiple test samples, each test sample corresponds to a group of time feature sequences;

[0013] Step 6: Construct a load power consumption prediction model, input the training set into the constructed load power consumption prediction model in batches, obtain the load power consumption prediction model after training, and select any test sample in the test set, input the test sample into the load power consumption prediction model after training, that is, predict the power consumption data corresponding to the test sample.

[0014] Preferably, the load is a server or a PC.

[0015] Preferably, there are four components in the load, namely: CPU, GPU, memory and hard disk.

[0016] Preferably, the power consumption performance index of each component of the load in step 1 corresponds to:

[0017] CPU utilization, CPU operating frequency and CPU temperature are used as CPU power consumption performance indicators;

[0018] GPU utilization, GPU graphics frequency, GPU memory frequency and GPU temperature are used as GPU power consumption performance indicators;

[0019] Use memory usage as a memory power consumption performance indicator;

[0020] The hard disk read rate and write rate are used as hard disk power consumption performance indicators.

[0021] Taking into account the complexity of the load during operation, the operating state of the load in step 2 is divided into five types: CPU intensive, memory intensive, I / O intensive, GPU intensive and mixed intensive, where the mixed intensive type is a mixture of the four components of CPU, memory, I / O and GPU.

[0022] Preferably, the load power consumption prediction model in step 6 is a CNN-LSTM network, and the CNN-LSTM network includes a CNN network and an LSTM network connected in sequence.

[0023] Preferably, the CNN network includes a 1D convolution layer, an activation function layer and a regularization layer connected in sequence, and the LSTM network adopts a bidirectional LSTM layer.

[0024] Preferably, the power consumption data in step 2 is acquired by a non-invasive acquisition device, and the non-invasive acquisition device includes a current transformer, a smart meter, a communication module and a data storage module. The load is electrically connected to the power supply, the current transformer is used to collect the current signal on the power line connecting the load and the power supply, the smart meter is electrically connected to the current transformer, and the smart meter is communicatively connected to the data storage module through the communication module.

[0025] In order to eliminate the dimensional differences between features, the preprocessing in step 3 is a normalization process.

[0026] The technical solution adopted by the present invention to solve the above-mentioned second technical problem is: a computer-readable storage medium, which stores a computer program and can be read and executed by a processor, characterized in that: when the computer program is executed by the processor, it implements the load power consumption prediction method as mentioned above.

[0027] Compared with the prior art, the advantages of the present invention are: analyzing the volatility of time feature sequences, sorting them in order from strong to weak fluctuations, selecting the first N preprocessed time feature sequences, and then performing VMD decomposition and reconstruction on the selected time feature sequences, and finally combining the reconstructed time feature sequences with other time feature sequences to form a data set. Therefore, this method takes into account the fluctuation situation, and respectively obtains the power consumption performance data and power consumption data of each type of load operating state, thereby improving the accuracy of load power consumption prediction, and the method is simple, efficient and highly adaptable. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a block diagram of a load power consumption prediction method in an embodiment of the present invention;

[0029] Figure 2 This is a structural diagram of a CNN-LSTM network in an embodiment of the present invention;

[0030] Figure 3 Schematic diagram of a non-invasive data collection device in an embodiment of the present invention. DETAILED DESCRIPTION

[0031] The present invention is further described in detail below with reference to the accompanying drawings.

[0032] The load power consumption prediction method in this embodiment includes the following steps:

[0033] Step 1: Obtain the power consumption performance index of each component of the load;

[0034] The load is a server or PC. There are four components in the load: CPU, GPU, memory, and hard disk.

[0035] The power consumption performance indicators of each component of the load correspond to: CPU utilization, CPU operating frequency and CPU temperature as CPU power consumption performance indicators; GPU utilization, GPU graphics frequency, GPU memory frequency and GPU temperature as GPU power consumption performance indicators; memory usage as memory power consumption performance indicator; hard disk read rate and write rate as hard disk power consumption performance indicator;

[0036] Step 2: Divide the operating state of the load into multiple types, simulate each type of operating state of the load, and obtain the power consumption performance data and power consumption data of the load in each type of operating state. The power consumption performance data of the load in a certain type of operating state is the time characteristic sequence corresponding to the power consumption performance index of each component of the load;

[0037] In this embodiment, the operating states of the load are divided into five categories: CPU intensive, memory intensive, I / O intensive, GPU intensive and mixed intensive, where the mixed intensive is a mixture of the four components of CPU, memory, I / O and GPU;

[0038] Step 3: preprocess all power consumption performance data acquired in step 2 to obtain multiple preprocessed time feature sequences;

[0039] In this embodiment, the preprocessing is normalization processing. Normalization is a common data preprocessing technology that aims to unify the numerical ranges between different features to a similar scale. Normalization can eliminate the dimensional differences between features, which helps to improve the training effect of the following load power consumption prediction model. In this embodiment, the minimum-maximum scaling (Min-MaxScaling) is used to scale the power consumption performance data to the range of [0,1]. The formula is as follows:

[0040]

[0041] Where X is the data to be normalized, X′ is the data after normalization of X, X_min is the minimum value of X, and X_max is the maximum value of X;

[0042] Step 4: Analyze the volatility of each preprocessed time feature sequence in step 3, sort each preprocessed time feature sequence in the order of volatility from strong to weak, and obtain the first N preprocessed time feature sequences after sorting; N is a positive integer;

[0043] In this embodiment, N=3, and the first three pre-processed time feature sequences after sorting are the time feature sequences of CPU utilization, CPU frequency, and GPU utilization respectively;

[0044] Step 5, performing VMD decomposition on the N time feature sequences obtained in step 4 respectively to obtain multiple stable modal components, reconstructing all stable modal components to obtain reconstructed time feature sequences, and combining the reconstructed time feature sequences with other time feature sequences except the sorted first N time feature sequences in step 4 to form a data set, and dividing the data set into a training set and a test set, wherein the training set includes multiple training samples, each training sample includes a group of time feature sequences and power consumption data corresponding to the group of time feature sequences; the test set includes multiple test samples, each test sample corresponds to a group of time feature sequences;

[0045] In this embodiment, VMD decomposition is variational mode decomposition. VMD decomposition and reconstruction of the decomposed modal components are both prior arts and will not be described in detail here.

[0046] Step 6: construct a load power consumption prediction model, input the training set into the constructed load power consumption prediction model in batches, obtain the load power consumption prediction model after training, and select any test sample in the test set, input the test sample into the load power consumption prediction model after training, that is, predict the power consumption data corresponding to the test sample;

[0047] like Figure 1 As shown, in this embodiment, the load power consumption prediction model is a CNN-LSTM network, such as Figure 2As shown in the figure, the CNN-LSTM network includes a CNN network and an LSTM network connected in sequence; the CNN network includes a 1D convolution layer, an activation function layer and a regularization layer connected in sequence, the activation function layer adopts the ReLu activation function, the regularization layer is Dropout, the 1D convolution layer uses a 1D convolution kernel to extract features from the data, and batch normalization can speed up the training of the network; the ReLu activation function is used to enhance the nonlinearity of the network, the pooling layer is discarded to retain the features, and Dropout is used to reduce the overfitting of the network; the LSTM network adopts a bidirectional LSTM layer; the bidirectional LSTM is used to fully explore the time dimension characteristics of the sequence, improve the accuracy of the model load power consumption prediction, and finally enter the fully connected prediction power consumption.

[0048] In this embodiment, the method of the present invention is described by taking the server as an example of the load. In step 2, a benchmark tool is used to simulate various work tasks of the server, wherein the CPU-Z stress test tool is used to simulate the CPU load state, the Memtest64 tool is used to apply different running memories to test the memory performance, and the CrystalDiskMark is used to detect the hard disk read and write speed. In addition, the GPU load is controlled by running different deep learning tasks; in mixed intensive tasks, the above four benchmark tools are used at the same time to simulate the concurrency of multiple types of tasks. Among them, the CPU utilization is designed from no load to full load in five gears in CPU intensive tasks, the memory is designed from 12G to 24G in four gears in memory intensive tasks, the mechanical hard disk is designed to read and write 64K sequentially in I / 0 intensive tasks, the GPU utilization is designed from 37% to 96% in GPU intensive tasks, and 9 mixed modes are designed in mixed intensive tasks. Under no load conditions, the CPU utilization and GPU utilization are below 5%, the memory occupancy is 8G, and the hard disk read and write are basically below 100Kb / s.

[0049] Table 1 Specific design of each operating state

[0050]

[0051] In step 2, the power consumption performance data of the server in each type of operation state is obtained by calling the performance interface of Windows and Nvidia through Python script. The power consumption data is obtained by non-intrusive acquisition devices, such as Figure 3 As shown, the non-intrusive data acquisition device includes a current transformer, a smart meter, a communication module and a data storage module. The above-mentioned load is electrically connected to the power supply, the current transformer is used to collect the current signal on the power line connecting the load and the power supply, the smart meter is electrically connected to the current transformer, and the smart meter is communicated with the data storage module through the communication module.

[0052] The smart meter collects the voltage of the load and the current scaled by the current transformer. The smart meter has a built-in digital-to-analog conversion module that can convert the voltage and current signals into digital signals. The communication module transmits the data of the smart meter to the data storage device. The power consumption data of the smart meter is transmitted to the data storage device through the RS485 communication port using the modbus protocol. The data storage module is used to receive and store data and includes some basic data processing functions, such as converting the received binary data into decimal, setting the sampling frequency, and selecting the storage format of csv, txt, etc. according to the situation.

[0053] This embodiment also involves a computer-readable storage medium, which stores a computer program and can be read and executed by a processor. When the computer program is executed by the processor, the load power consumption prediction method as described above is implemented.

[0054] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A load power consumption prediction method, characterized in that The steps include: Step 1: Obtain the power consumption performance index of each component of the load; Step 2: Divide the operating state of the load into multiple types, simulate each type of operating state of the load, and obtain the power consumption performance data and power consumption data of the load in each type of operating state. The power consumption performance data of the load in a certain type of operating state is the time characteristic sequence corresponding to the power consumption performance index of each component of the load; Step 3: preprocess all power consumption performance data acquired in step 2 to obtain multiple preprocessed time feature sequences; Step 4: Analyze the volatility of each preprocessed time feature sequence in step 3, sort each preprocessed time feature sequence in the order of volatility from strong to weak, and obtain the first N preprocessed time feature sequences after sorting; N is a positive integer; Step 5, performing VMD decomposition on the N time feature sequences obtained in step 4 respectively to obtain multiple stable modal components, reconstructing all stable modal components to obtain reconstructed time feature sequences, and combining the reconstructed time feature sequences with other time feature sequences except the sorted first N time feature sequences in step 4 to form a data set, and dividing the data set into a training set and a test set, wherein the training set includes multiple training samples, each training sample includes a group of time feature sequences and power consumption data corresponding to the group of time feature sequences; the test set includes multiple test samples, each test sample corresponds to a group of time feature sequences; Step 6: Construct a load power consumption prediction model, input the training set into the constructed load power consumption prediction model in batches, obtain the load power consumption prediction model after training, and select any test sample in the test set, input the test sample into the load power consumption prediction model after training, that is, predict the power consumption data corresponding to the test sample.

2. The load power consumption prediction method according to claim 1, characterized in that: The load is a server or a PC.

3. The load power consumption prediction method according to claim 2, characterized in that: There are four components in the load, namely: CPU, GPU, memory and hard disk.

4. The load power consumption prediction method according to claim 3, characterized in that: The power consumption performance index of each component of the load in step 1 corresponds to: CPU utilization, CPU operating frequency and CPU temperature are used as CPU power consumption performance indicators; GPU utilization, GPU graphics frequency, GPU memory frequency and GPU temperature are used as GPU power consumption performance indicators; Use memory usage as a memory power consumption performance indicator; The hard disk read rate and write rate are used as hard disk power consumption performance indicators.

5. The load power consumption prediction method according to claim 4, characterized in that: The operating status of the load in step 2 is divided into five types: CPU intensive, memory intensive, I / O intensive, GPU intensive and mixed intensive, where the mixed intensive type is a mixture of the four components of CPU, memory, I / O and GPU.

6. The load power consumption prediction method according to claim 4, characterized in that: The load power consumption prediction model in step 6 is a CNN-LSTM network, and the CNN-LSTM network includes a CNN network and a LSTM network connected in sequence.

7. The load power consumption prediction method according to claim 6, characterized in that: The CNN network includes a 1D convolution layer, an activation function layer and a regularization layer connected in sequence, and the LSTM network adopts a bidirectional LSTM layer.

8. The load power consumption prediction method according to any one of claims 1 to 7, characterized in that: The power consumption data in step 2 is acquired by a non-invasive acquisition device, which includes a current transformer, a smart meter, a communication module and a data storage module. The load is electrically connected to a power supply, the current transformer is used to collect a current signal on a power line connecting the load to the power supply, the smart meter is electrically connected to the current transformer, and the smart meter is communicatively connected to the data storage module via the communication module.

9. The load power consumption prediction method according to any one of claims 1 to 7, characterized in that: The preprocessing in step 3 is normalization processing.

10. A computer-readable storage medium storing a computer program and capable of being read and executed by a processor, characterized in that: When the computer program is executed by a processor, the load power consumption prediction method according to any one of claims 1 to 9 is implemented.

Images