Cloud data center load prediction method and system based on improved TimesNet

Abstract

The invention provides an improved TimesNet-based cloud data center load prediction method and system, and the method comprises the steps: carrying out the multi-period feature extraction of a time sequence of resource utilization data, determining a plurality of most significant periods, carrying out the two-dimensional time sequence change modeling, obtaining the change data of the time sequence at different time scales, and carrying out the prediction of the load of a cloud data center. Therefore, the most significant features of a plurality of periods are extracted, and the features are adaptively fused into a ResNet-based Times Block. A plurality of Times Blocks are combined and stacked to form an improved Times Net network, so that the resource demand load of a cloud platform in the future time is predicted, the multi-periodicity and volatility of the load can be more accurately captured, the long-term prediction precision is improved, two-dimensional convolution is performed by using ResNet, the rapid change of the load of a cloud data center can be more rapidly adapted, and the resource demand load of the cloud platform in the future time can be predicted. The resource scheduling lagging or redundancy is reduced through accurate load prediction, the resource utilization efficiency is improved, the operation cost caused by default and waste is reduced, and the load prediction has better adaptability and generalization ability.

Description

Technical Field

[0001] The present invention relates to the field of cloud computing, and in particular to a cloud data center load prediction method and system based on improved TimesNet. Background Art

[0002] In the cloud computing environment, cloud vendors widely use time series load forecasting technology to predict resource changes in the next period of time based on historical CPU or memory resource utilization, so as to perform elastic scaling in advance, improve resource utilization efficiency, and reduce business bottlenecks. Traditional load forecasting technology mainly relies on classic time series analysis methods, such as ARIMA and seasonal decomposition time series forecasting (STL). These methods work well when dealing with simple cycles and linear trends, but are often difficult to accurately predict when faced with high volatility and complex patterns of cloud data center loads. In addition, although deep learning methods such as LSTM and Informer have higher accuracy, they are insufficient in processing multi-period cloud load data and have large resource overhead. It can be seen that the existing load forecasting methods have shortcomings such as insufficient waveform processing, limited prediction accuracy, and low computational efficiency. They cannot accurately predict loads with low overhead when the load presents nonlinear and multi-periodic characteristics. Summary of the invention

[0003] The purpose of the present invention is to provide a cloud data center load prediction method and system based on improved TimesNet, which extracts multi-period features of the time series of resource utilization data, determines the most significant several periods, and uses this to perform two-dimensional time series change modeling to obtain change data of the time series on different time scales, thereby extracting the characteristics of the most significant several periods, and adaptively merging the characteristics into a TimesBlock based on ResNet; multiple TimesBlocks are combined and stacked to form an improved TimesNet network, so as to predict the resource demand load of the cloud platform in the future, which can more accurately capture the multi-periodicity and volatility of the load, improve the accuracy of long-term prediction, use ResNet for two-dimensional convolution, and can adapt to the rapid changes in the cloud data center load more quickly, reduce resource scheduling lag or redundancy through accurate load prediction, improve resource utilization efficiency, reduce operating costs caused by default and waste, and make load prediction have better adaptability and generalization ability.

[0004] The present invention is achieved through the following technical solutions: The cloud data center load prediction method based on the improved TimesNet includes: The improved TimesNet model is used to extract multi-period features of the time series corresponding to the resource utilization data of the cloud platform, and the most significant periods in the time series are determined; based on the most significant periods, two-dimensional time series change modeling is performed to obtain the change data of the time series on different time scales; Processing the change data of the time series at different time scales, extracting the features of the most significant cycles, and adaptively fusing the features of the most significant cycles, thereby obtaining a TimesBlock based on ResNet; Multiple TimesBlocks are combined and stacked to form an improved TimesNet network; the improved TimesNet network is then used to perform load prediction on the resource utilization data of the cloud platform to obtain the resource demand load of the cloud platform in the future.

[0005] Optionally, the improved TimesNet model is used to perform multi-period feature extraction on the time series corresponding to the resource utilization data of the cloud platform to determine the most significant periods in the time series; two-dimensional time series change modeling is performed based on the most significant periods to obtain change data of the time series on different time scales, including: Obtain historical resource utilization data of a cloud platform, process the historical resource utilization data based on the generation time of the historical resource utilization data, and obtain a time series corresponding to the resource utilization data of the cloud platform; wherein the historical resource utilization data includes CPU and / or memory historical utilization data; perform fast Fourier transform on the time series using an improved TimesNet model to obtain the frequency component intensity of the time series, which is used as a multi-period feature of the time series; and then determine the most significant number of periods in the time series based on the frequency component intensity; According to the most significant cycles, the time series is folded according to the length corresponding to each cycle, so as to convert the one-dimensional time series into a two-dimensional tensor; wherein the width of the two-dimensional tensor is equal to the length of the corresponding cycle; and then a two-dimensional convolutional network is used to capture the temporal variation data of the two-dimensional tensor within and between cycles, which is used as the variation data of the time series on different time scales.

[0006] Optionally, the change data of the time series at different time scales are processed to extract the features of the most significant cycles, and the features of the most significant cycles are adaptively fused to obtain a TimesBlock based on ResNet, including: Performing two-dimensional convolution processing on the change data of the time series at different time scales through ResNet to extract the features of each of the most significant cycles; According to the frequency component strengths corresponding to all the periods under the time series, the features of the most significant periods are weightedly fused to obtain a TimesBlock based on ResNet.

[0007] Optionally, performing operation abnormality determination on the processes of the most significant cycles in the time series includes: Extracting and obtaining the period acquisition durations of several most significant periods in the time series; Comparing the period acquisition duration with a preset duration threshold; When the periodic acquisition duration exceeds a preset duration threshold, the processing duration of the historical resource utilization data and the frequency component intensity acquisition duration in the historical records are retrieved; Determine the process evaluation coefficients of the most significant several periods in the time series by using the processing time of the historical resource utilization data and the frequency component intensity acquisition time in the historical records; The process evaluation coefficient is obtained by the following formula: Wherein, S represents the process evaluation coefficient; n represents the number of times the most significant cycles in the time series are determined in the historical records except for the current determination of the most significant cycles in the time series; T 01i and T 02i They represent the processing time of historical resource utilization data and the acquisition time of frequency component intensity corresponding to the determination of the i-th period respectively; T zmax Indicates the maximum value of the cycle duration determined in n cycles; T z Indicates that the periodic acquisition duration corresponding to the preset duration threshold is exceeded; S 01 and S 02 denote the first coefficient and the second coefficient respectively, and the first coefficient and the second coefficient are obtained by the following formula: Among them, S 01 and S 02 Represent the first coefficient and the second coefficient respectively; T 01i and T 02i They represent the processing time of historical resource utilization data and the acquisition time of frequency component intensity corresponding to the determination of the i-th period respectively; T zmax Indicates the maximum value of the cycle duration determined in n cycles; T zmin Indicates the minimum value of the cycle acquisition duration in n cycles; Comparing the process evaluation coefficient with a preset evaluation coefficient; When the process evaluation coefficient is not lower than the preset evaluation coefficient, it is determined that there are abnormalities in the processes of the most significant several cycles in the time series, and an abnormality alarm is issued.

[0008] Optionally, multiple TimesBlocks are combined and stacked to form an improved TimesNet network; and the improved TimesNet network is then used to perform load prediction on the resource utilization data of the cloud platform to obtain the resource demand load of the cloud platform in the future, including: Analyze multiple TimesBlocks, determine the residuals of all TimesBlocks, and then link multiple TimesBlock residuals to form an improved TimesNet network; The resource utilization data of the cloud platform is input into the improved TimesNet network to perform load prediction for periods of different lengths, so as to obtain the resource demand load corresponding to the cloud platform in time intervals of different lengths in the future; wherein the resource demand load includes the data demand load of the CPU and / or memory.

[0009] The cloud data center load prediction system based on the improved TimesNet includes: A multi-period feature extraction module is used to extract multi-period features from the time series corresponding to the resource utilization data of the cloud platform using the improved TimesNet model, and determine the most significant periods in the time series; A two-dimensional time series change modeling module is used to perform two-dimensional time series change modeling according to the most significant cycles to obtain change data of the time series at different time scales; A two-dimensional convolution processing module is used to process the change data of the time series at different time scales to extract the characteristics of each of the most significant cycles; An adaptive fusion module, used for adaptively fusing the features of the most significant cycles, thereby obtaining a TimesBlock based on ResNet; The combination stacking module is used to combine and stack multiple TimesBlocks to form an improved TimesNet network; The load prediction module is used to use the improved TimesNet network to perform load prediction on the resource utilization data of the cloud platform to obtain the resource demand load of the cloud platform in the future.

[0010] Optionally, the multi-cycle feature extraction module is used to perform multi-cycle feature extraction on the time series corresponding to the resource utilization data of the cloud platform using the improved TimesNet model, and determine the most significant cycles in the time series, including: Obtain historical resource utilization data of a cloud platform, process the historical resource utilization data based on the generation time of the historical resource utilization data, and obtain a time series corresponding to the resource utilization data of the cloud platform; wherein the historical resource utilization data includes CPU and / or memory historical utilization data; perform fast Fourier transform on the time series using an improved TimesNet model to obtain the frequency component intensity of the time series, which is used as a multi-period feature of the time series; and then determine the most significant number of periods in the time series based on the frequency component intensity; The two-dimensional time series change modeling module is used to perform two-dimensional time series change modeling according to the most significant cycles to obtain change data of the time series on different time scales, including: According to the most significant cycles, the time series is folded according to the length corresponding to each cycle, so as to convert the one-dimensional time series into a two-dimensional tensor; wherein the width of the two-dimensional tensor is equal to the length of the corresponding cycle; and then a two-dimensional convolutional network is used to capture the temporal variation data of the two-dimensional tensor within and between cycles, which is used as the variation data of the time series on different time scales.

[0011] Optionally, performing operation abnormality determination on the processes of the most significant cycles in the time series includes: Extracting and obtaining the period acquisition durations of several most significant periods in the time series; Comparing the period acquisition duration with a preset duration threshold; When the periodic acquisition duration exceeds a preset duration threshold, the processing duration of the historical resource utilization data and the frequency component intensity acquisition duration in the historical records are retrieved; Determine the process evaluation coefficients of the most significant several periods in the time series by using the processing time of the historical resource utilization data and the frequency component intensity acquisition time in the historical records; The process evaluation coefficient is obtained by the following formula: Wherein, S represents the process evaluation coefficient; n represents the number of times the most significant cycles in the time series are determined in the historical records except for the current determination of the most significant cycles in the time series; T 01i and T 02i They represent the processing time of historical resource utilization data and the acquisition time of frequency component intensity corresponding to the determination of the i-th period respectively; Tzmax Indicates the maximum value of the cycle duration determined in n cycles; T z Indicates that the periodic acquisition duration corresponding to the preset duration threshold is exceeded; S 01 and S 02 denote the first coefficient and the second coefficient respectively, and the first coefficient and the second coefficient are obtained by the following formula: Among them, S 01 and S 02 Represent the first coefficient and the second coefficient respectively; T 01i and T 02i They represent the processing time of historical resource utilization data and the acquisition time of frequency component intensity corresponding to the determination of the i-th period respectively; T zmax Indicates the maximum value of the cycle duration determined in n cycles; T zmin Indicates the minimum value of the cycle acquisition duration in n cycles; Comparing the process evaluation coefficient with a preset evaluation coefficient; When the process evaluation coefficient is not lower than the preset evaluation coefficient, it is determined that there are abnormalities in the processes of the most significant several cycles in the time series, and an abnormality alarm is issued.

[0012] Optionally, the two-dimensional convolution processing module is used to process the change data of the time series at different time scales to extract the characteristics of each of the most significant cycles, including: Performing two-dimensional convolution processing on the change data of the time series at different time scales through ResNet to extract the features of each of the most significant cycles; The adaptive fusion module is used to adaptively fuse the features of the most significant cycles to obtain a TimesBlock based on ResNet, including: According to the frequency component strengths corresponding to all the periods under the time series, the features of the most significant periods are weightedly fused to obtain a TimesBlock based on ResNet.

[0013] Optionally, the combination stacking module is used to combine and stack multiple TimesBlocks to form an improved TimesNet network, including: Analyze multiple TimesBlocks, determine the residuals of all TimesBlocks, and then link multiple TimesBlock residuals to form an improved TimesNet network; The load prediction module is used to use the improved TimesNet network to perform load prediction on the resource utilization data of the cloud platform to obtain the resource demand load of the cloud platform in the future, including: The resource utilization data of the cloud platform is input into the improved TimesNet network to perform load prediction for periods of different lengths, so as to obtain the resource demand load corresponding to the cloud platform in time intervals of different lengths in the future; wherein the resource demand load includes the data demand load of the CPU and / or memory.

[0014] Compared with the prior art, the present invention has the following beneficial effects: The cloud data center load prediction method and system based on improved TimesNet provided in the present application extract multi-period features of the time series of resource utilization data, determine the most significant several periods, and use this to perform two-dimensional time series change modeling to obtain the change data of the time series at different time scales, thereby extracting the characteristics of each of the most significant several periods, and adaptively merging the characteristics into a TimesBlock based on ResNet; multiple TimesBlocks are combined and stacked to form an improved TimesNet network, so as to predict the resource demand load of the cloud platform in the future, which can more accurately capture the multi-periodicity and volatility of the load, improve the accuracy of long-term prediction, and use ResNet for two-dimensional convolution, which can adapt to the rapid changes in the cloud data center load more quickly, reduce resource scheduling lag or redundancy through accurate load prediction, improve resource utilization efficiency, reduce operating costs caused by default and waste, and make load prediction have better adaptability and generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. Among them: Figure 1 A schematic diagram of the flow of a cloud data center load prediction method based on improved TimesNet provided by the present invention.

[0016] Figure 2 A schematic diagram of the structure of a cloud data center load prediction system based on improved TimesNet provided by the present invention. DETAILED DESCRIPTION

[0017] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are described in detail below in conjunction with the accompanying drawings. It is to be understood that the specific embodiments described herein are only used to explain the present application, rather than to limit the present application. It should also be noted that, for ease of description, only some structures related to the present application are shown in the accompanying drawings, rather than all structures. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0018] The terms "include" and "have" and any variations thereof in this application are intended to cover non-exclusive inclusions. For example, a process, method, method, product or device comprising a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products or devices.

[0019] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0020] See also Figure 1 As shown, an embodiment of the present application provides a cloud data center load prediction method based on an improved TimesNet. The cloud data center load prediction method based on an improved TimesNet includes: The improved TimesNet model is used to extract multi-period features from the time series corresponding to the resource utilization data of the cloud platform, and the most significant periods in the time series are determined. Based on the most significant periods, two-dimensional time series change modeling is performed to obtain the change data of the time series at different time scales. The change data of the time series at different time scales are processed to extract the features of the most significant cycles, and the features of the most significant cycles are adaptively fused to obtain TimesBlock based on ResNet. Multiple TimesBlocks are combined and stacked to form an improved TimesNet network. The improved TimesNet network is then used to perform load prediction on the resource utilization data of the cloud platform to obtain the resource demand load of the cloud platform in the future.

[0021] The beneficial effects of the above embodiments are as follows: the cloud data center load prediction method based on the improved TimesNet extracts multi-period features of the time series of resource utilization data, determines the most significant periods, and uses this to perform two-dimensional time series change modeling to obtain change data of the time series at different time scales, thereby extracting the characteristics of the most significant periods, and adaptively merging the characteristics into a TimesBlock based on ResNet; multiple TimesBlocks are combined and stacked to form an improved TimesNet network, thereby predicting the resource demand load of the cloud platform in the future, which can more accurately capture the multi-periodicity and volatility of the load, improve the accuracy of long-term predictions, and use ResNet for two-dimensional convolution, which can adapt to the rapid changes in the cloud data center load more quickly, reduce resource scheduling lags or redundancies through accurate load predictions, improve resource utilization efficiency, reduce operating costs caused by default and waste, and make load predictions have better adaptability and generalization capabilities.

[0022] In another embodiment, the improved TimesNet model is used to extract multi-period features from the time series corresponding to the resource utilization data of the cloud platform, and the most significant periods in the time series are determined; based on the most significant periods, two-dimensional time series change modeling is performed to obtain change data of the time series on different time scales, including: Obtain historical resource utilization data of the cloud platform, process the historical resource utilization data based on the generation time of the historical resource utilization data, and obtain the corresponding time series of the resource utilization data of the cloud platform; wherein the historical resource utilization data includes historical CPU and / or memory utilization data; perform fast Fourier transform on the time series using the improved TimesNet model to obtain the frequency component intensity of the time series, which is used as the multi-period feature of the time series; and then determine the most significant number of periods in the time series based on the frequency component intensity; According to the most significant cycles, the time series is folded according to the length corresponding to each cycle, so as to convert the one-dimensional time series into a two-dimensional tensor; wherein the width of the two-dimensional tensor is equal to the length of the corresponding cycle; and then a two-dimensional convolutional network is used to capture the temporal variation data of the two-dimensional tensor within and between cycles, which is used as the variation data of the time series on different time scales.

[0023] The beneficial effect of the above-mentioned embodiment is that the cloud platform needs to call different types of resources such as CPU and memory in the process of processing different tasks. The number and speed of calling these resources directly affect the reliability of the cloud platform's task processing. However, it is not the case that the more and faster the number and speed of calling resources such as CPU and memory in the process of processing tasks, the better. This depends on the number of resources actually required in the process of processing tasks. Calling resources too much and too quickly will cause redundant waste of resources. For this reason, it is necessary to accurately predict the number of resources that the cloud platform will use in the future based on the resource utilization data of the cloud platform in the historical operation process. In order to ensure the matching of the historical resource utilization data of the cloud platform with the time series prediction, it is necessary to process the historical resource data in advance, and divide the historical resource utilization data in chronological order based on the generation time of the historical resource utilization data, so as to obtain the corresponding time series of the resource utilization data of the cloud platform, so as to facilitate the subsequent direct processing of the time series. The improved TimesNet model is also used to perform a fast Fourier transform (FFT) on the time series to obtain the frequency component intensity of the time series, which is used as the multi-period feature of the time series. The frequency component intensity of the time series obtained by FFT is a conventional technical means in this field and will not be described in detail here. The frequency component intensity is used to characterize the importance of the data in the period corresponding to the time series. The greater the frequency component intensity, the greater the importance of the data in the corresponding period. At this time, the periods corresponding to the first k largest frequency component intensities are taken as the most significant periods in the time series, that is, , thereby capturing the multi-periodicity of the time series and realizing the multi-periodic feature extraction of the time series, so that the subsequent load forecasting can adapt to the high volatility and multi-periodicity of different scenarios. Then, according to the most significant k periods, the time series is divided into (where i = 1, 2, ..., k) to fold the time series with a one-dimensional form into a width The two-dimensional tensor is transformed into a two-dimensional tensor, and a two-dimensional convolutional network is used to capture the time series change data within the cycle and between periodic parts. Through the above transformation, the change patterns of time series on different time scales can be simultaneously analyzed and learned.

[0024] In another embodiment, determining the most significant number of cycles in the time series for a process of performing an operational abnormality determination includes: Extracting and obtaining the period acquisition durations of several most significant periods in the time series; Comparing the period acquisition duration with a preset duration threshold; When the periodic acquisition duration exceeds a preset duration threshold, the processing duration of the historical resource utilization data and the frequency component intensity acquisition duration in the historical records are retrieved; Determine the process evaluation coefficients of the most significant several periods in the time series by using the processing time of the historical resource utilization data and the frequency component intensity acquisition time in the historical records; The process evaluation coefficient is obtained by the following formula: Wherein, S represents the process evaluation coefficient; n represents the number of times the most significant cycles in the time series are determined in the historical records except for the current determination of the most significant cycles in the time series; T 01i and T 02i They represent the processing time of historical resource utilization data and the acquisition time of frequency component intensity corresponding to the determination of the i-th period respectively; T zmax Indicates the maximum value of the cycle duration determined in n cycles; T z Indicates that the periodic acquisition duration corresponding to the preset duration threshold is exceeded; S 01 and S 02 denote the first coefficient and the second coefficient respectively, and the first coefficient and the second coefficient are obtained by the following formula: Among them, S 01 and S 02 Represent the first coefficient and the second coefficient respectively; T 01i and T 02i They represent the processing time of historical resource utilization data and the acquisition time of frequency component intensity corresponding to the determination of the i-th period respectively; T zmax Indicates the maximum value of the cycle duration determined in n cycles; T zmin Indicates the minimum value of the cycle acquisition duration in n cycles; Comparing the process evaluation coefficient with a preset evaluation coefficient; When the process evaluation coefficient is not lower than the preset evaluation coefficient, it is determined that there are abnormalities in the processes of the most significant several cycles in the time series, and an abnormality alarm is issued.

[0025] The beneficial effects of the above embodiments are that the technical solution can monitor the determination process of the most significant cycles in the time series in real time, and timely discover possible anomalies by extracting and comparing the cycle acquisition duration with the preset duration threshold. When the cycle acquisition duration exceeds the preset threshold, the solution can retrieve relevant data in the historical records for auxiliary determination. This includes the processing duration of historical resource utilization data and the acquisition duration of frequency component intensity, which provide rich reference information for subsequent anomaly determination. By introducing the process evaluation coefficient S, the technical solution can quantitatively evaluate whether there are anomalies in the process of determining the most significant cycles in the time series. The calculation of the process evaluation coefficient S takes into account the multiple cycle determination durations in the historical records, as well as the weights of the first coefficient S01 and the second coefficient S02, making the evaluation more comprehensive and accurate. The calculation method of the first coefficient S01 and the second coefficient S02 takes into account the maximum and minimum values ​​of the cycle acquisition duration, which enables the process evaluation coefficient S to be dynamically adjusted according to different historical data, enhancing the adaptability and flexibility of the technical solution. When the process evaluation coefficient is not lower than the preset evaluation coefficient, the technical solution can determine that there are anomalies in the process of determining the most significant cycles in the time series, and perform an abnormal alarm. This helps to detect problems in a timely manner and take appropriate countermeasures to reduce potential risks and losses.

[0026] In summary, this technical solution effectively improves the accuracy and efficiency of abnormality determination in determining the most significant cycles in the time series through real-time monitoring, historical data-assisted determination, quantitative evaluation of abnormality, dynamic adjustment and adaptability, and timely alarm and response. This is of great significance for ensuring the stable operation of the system and responding to potential problems in a timely manner.

[0027] In another embodiment, the change data of the time series at different time scales are processed to extract the features of the most significant cycles, and the features of the most significant cycles are adaptively fused to obtain a TimesBlock based on ResNet, including: The ResNet is used to perform two-dimensional convolution processing on the change data of the time series at different time scales to extract the features of the most significant cycles. According to the frequency component intensity corresponding to all the periods under the time series, the features of the most significant periods are weighted and fused to obtain the ResNet-based TimesBlock.

[0028] The beneficial effect of the above embodiment is that ResNet performs two-dimensional convolution processing on the change data of the time series at different time scales, and utilizes ResNet's own powerful feature extraction capabilities and convergence speed advantages to improve the prediction accuracy and training efficiency of the model. In addition, according to the frequency component intensity corresponding to all periods under the time series, the features of the most significant periods are weighted and fused to obtain a TimesBlock based on ResNet; wherein the greater the frequency component intensity corresponding to a certain period, the greater the corresponding weighted proportion of the period in the weighted fusion process. By referencing the adaptive fusion strategy of weighted fusion, the features of different periods are weighted and fused, so that subsequent load prediction of different period lengths can be performed, with better adaptability and generalization capabilities.

[0029] In another embodiment, multiple TimesBlocks are combined and stacked to form an improved TimesNet network; the improved TimesNet network is then used to perform load prediction on the resource utilization data of the cloud platform to obtain the resource demand load of the cloud platform in the future, including: Analyze multiple TimesBlocks, determine the residuals of all TimesBlocks, and then link multiple TimesBlock residuals to form an improved TimesNet network; The resource utilization data of the cloud platform is input into the improved TimesNet network to perform load prediction for periods of different lengths, and the resource demand load corresponding to the cloud platform in future time intervals of different lengths is obtained; wherein the resource demand load includes the data demand load of the CPU and / or memory.

[0030] The beneficial effect of the above embodiment is that each TimesBlock can perform load prediction in the corresponding period. In order to accurately predict the multi-periodicity and volatility of load demand, multiple TimesBlocks are analyzed to determine the residuals of all TimesBlocks, and then multiple TimesBlock residuals are linked to form an improved TimesNet network, so that the improved TimesNet network can achieve global accurate prediction of loads with multi-periodicity and volatility. Then the resource utilization data of the cloud platform is input into the improved TimesNet network to perform load prediction for periods of different lengths, and the resource demand load corresponding to the cloud platform in different time intervals in the future is obtained, which effectively reduces the scheduling lag and redundancy of load prediction, can adapt to the rapid changes in the cloud data center load more quickly, and improve the accuracy of long-term load prediction.

[0031] See also Figure 2As shown, an embodiment of the present application provides a cloud data center load prediction system based on an improved TimesNet. The cloud data center load prediction system based on an improved TimesNet includes: The multi-period feature extraction module is used to extract multi-period features from the time series corresponding to the resource utilization data of the cloud platform using the improved TimesNet model, and determine the most significant periods in the time series; A two-dimensional time series change modeling module is used to perform two-dimensional time series change modeling based on the most significant cycles to obtain change data of the time series at different time scales; A two-dimensional convolution processing module is used to process the change data of the time series at different time scales and extract the characteristics of the most significant cycles; The adaptive fusion module is used to adaptively fuse the features of the most significant cycles to obtain a TimesBlock based on ResNet; The combination stacking module is used to combine and stack multiple TimesBlocks to form an improved TimesNet network; The load prediction module is used to use the improved TimesNet network to perform load prediction on the resource utilization data of the cloud platform to obtain the resource demand load of the cloud platform in the future.

[0032] The beneficial effects of the above embodiments are as follows: the cloud data center load prediction system based on the improved TimesNet extracts multi-period features of the time series of resource utilization data, determines the most significant periods, and uses this to perform two-dimensional time series change modeling to obtain change data of the time series at different time scales, thereby extracting the characteristics of the most significant periods, and adaptively merging the characteristics into a TimesBlock based on ResNet; multiple TimesBlocks are combined and stacked to form an improved TimesNet network, thereby predicting the resource demand load of the cloud platform in the future, which can more accurately capture the multi-periodicity and volatility of the load, improve the accuracy of long-term predictions, and use ResNet for two-dimensional convolution, which can adapt to the rapid changes in the cloud data center load more quickly, reduce resource scheduling lags or redundancies through accurate load predictions, improve resource utilization efficiency, reduce operating costs caused by default and waste, and make load predictions have better adaptability and generalization capabilities.

[0033] In another embodiment, the multi-cycle feature extraction module is used to perform multi-cycle feature extraction on the time series corresponding to the resource utilization data of the cloud platform using the improved TimesNet model, and determine the most significant cycles in the time series, including: Obtain historical resource utilization data of the cloud platform, process the historical resource utilization data based on the generation time of the historical resource utilization data, and obtain the corresponding time series of the resource utilization data of the cloud platform; wherein the historical resource utilization data includes historical CPU and / or memory utilization data; perform fast Fourier transform on the time series using the improved TimesNet model to obtain the frequency component intensity of the time series, which is used as the multi-period feature of the time series; and then determine the most significant number of periods in the time series based on the frequency component intensity; The two-dimensional time series change modeling module is used to perform two-dimensional time series change modeling according to the most significant cycles to obtain the change data of the time series on different time scales, including: According to the most significant cycles, the time series is folded according to the length corresponding to each cycle, so as to convert the one-dimensional time series into a two-dimensional tensor; wherein the width of the two-dimensional tensor is equal to the length of the corresponding cycle; and then a two-dimensional convolutional network is used to capture the temporal variation data of the two-dimensional tensor within and between cycles, which is used as the variation data of the time series on different time scales.

[0034] The beneficial effect of the above-mentioned embodiment is that the cloud platform needs to call different types of resources such as CPU and memory in the process of processing different tasks. The number and speed of calling these resources directly affect the reliability of the cloud platform's task processing. However, it is not the case that the more and faster the number and speed of calling resources such as CPU and memory in the process of processing tasks, the better. This depends on the number of resources actually required in the process of processing tasks. Calling resources too much and too quickly will cause redundant waste of resources. For this reason, it is necessary to accurately predict the number of resources that the cloud platform will use in the future based on the resource utilization data of the cloud platform in the historical operation process. In order to ensure the matching of the historical resource utilization data of the cloud platform with the time series prediction, it is necessary to process the historical resource data in advance, and divide the historical resource utilization data in chronological order based on the generation time of the historical resource utilization data, so as to obtain the corresponding time series of the resource utilization data of the cloud platform, so as to facilitate the subsequent direct processing of the time series. The improved TimesNet model is also used to perform a fast Fourier transform (FFT) on the time series to obtain the frequency component intensity of the time series, which is used as the multi-period feature of the time series. The frequency component intensity of the time series obtained by FFT is a conventional technical means in this field and will not be described in detail here. The frequency component intensity is used to characterize the importance of the data in the period corresponding to the time series. The greater the frequency component intensity, the greater the importance of the data in the corresponding period. At this time, the periods corresponding to the first k largest frequency component intensities are taken as the most significant periods in the time series, that is, , thereby capturing the multi-periodicity of the time series and realizing the multi-periodic feature extraction of the time series, so that the subsequent load forecasting can adapt to the high volatility and multi-periodicity of different scenarios. Then, according to the most significant k periods, the time series is divided into (where i = 1, 2, ..., k) to fold the time series with a one-dimensional form into a width The two-dimensional tensor is transformed into a two-dimensional tensor, and a two-dimensional convolutional network is used to capture the time series change data within the cycle and between periodic parts. Through the above transformation, the change patterns of time series on different time scales can be simultaneously analyzed and learned.

[0035] In another embodiment, determining the most significant number of cycles in the time series for a process of performing an operational abnormality determination includes: Extracting and obtaining the period acquisition durations of several most significant periods in the time series; Comparing the period acquisition duration with a preset duration threshold; When the periodic acquisition duration exceeds a preset duration threshold, the processing duration of the historical resource utilization data and the frequency component intensity acquisition duration in the historical records are retrieved; Determine the process evaluation coefficients of the most significant several periods in the time series by using the processing time of the historical resource utilization data and the frequency component intensity acquisition time in the historical records; The process evaluation coefficient is obtained by the following formula: Wherein, S represents the process evaluation coefficient; n represents the number of times the most significant cycles in the time series are determined in the historical records except the current one; T 01i and T 02i They represent the processing time of historical resource utilization data and the acquisition time of frequency component intensity corresponding to the determination of the i-th period respectively; T zmax Indicates the maximum value of the cycle duration determined in n cycles; T z Indicates that the periodic acquisition duration corresponding to the preset duration threshold is exceeded; S 01 and S 02 denote the first coefficient and the second coefficient respectively, and the first coefficient and the second coefficient are obtained by the following formula: Among them, S 01 and S 02 Represent the first coefficient and the second coefficient respectively; T 01i and T 02i They represent the processing time of historical resource utilization data and the acquisition time of frequency component intensity corresponding to the determination of the i-th period respectively; T zmaxIndicates the maximum value of the cycle duration determined in n cycles; T zmin Indicates the minimum value of the cycle acquisition duration in n cycles; Comparing the process evaluation coefficient with a preset evaluation coefficient; When the process evaluation coefficient is not lower than the preset evaluation coefficient, it is determined that there are abnormalities in the processes of the most significant several cycles in the time series, and an abnormality alarm is issued.

[0036] The beneficial effects of the above embodiments are that the technical solution can monitor the determination process of the most significant cycles in the time series in real time, and timely discover possible anomalies by extracting and comparing the cycle acquisition duration with the preset duration threshold. When the cycle acquisition duration exceeds the preset threshold, the solution can retrieve relevant data in the historical records for auxiliary determination. This includes the processing duration of historical resource utilization data and the acquisition duration of frequency component intensity, which provide rich reference information for subsequent anomaly determination. By introducing the process evaluation coefficient S, the technical solution can quantitatively evaluate whether there are anomalies in the process of determining the most significant cycles in the time series. The calculation of the process evaluation coefficient S takes into account the multiple cycle determination durations in the historical records, as well as the weights of the first coefficient S01 and the second coefficient S02, making the evaluation more comprehensive and accurate. The calculation method of the first coefficient S01 and the second coefficient S02 takes into account the maximum and minimum values ​​of the cycle acquisition duration, which enables the process evaluation coefficient S to be dynamically adjusted according to different historical data, enhancing the adaptability and flexibility of the technical solution. When the process evaluation coefficient is not lower than the preset evaluation coefficient, the technical solution can determine that there are anomalies in the process of determining the most significant cycles in the time series, and perform an abnormal alarm. This helps to detect problems in a timely manner and take appropriate countermeasures to reduce potential risks and losses.

[0037] In summary, this technical solution effectively improves the accuracy and efficiency of abnormality determination in determining the most significant cycles in the time series through real-time monitoring, historical data-assisted determination, quantitative evaluation of abnormality, dynamic adjustment and adaptability, and timely alarm and response. This is of great significance for ensuring the stable operation of the system and responding to potential problems in a timely manner.

[0038] In another embodiment, the two-dimensional convolution processing module is used to process the change data of the time series at different time scales to extract the characteristics of each of the most significant cycles, including: The ResNet is used to perform two-dimensional convolution processing on the change data of the time series at different time scales to extract the features of the most significant cycles. The adaptive fusion module is used to adaptively fuse the features of the most significant cycles to obtain a ResNet-based TimesBlock, including: According to the frequency component intensity corresponding to all the periods under the time series, the features of the most significant periods are weighted and fused to obtain the ResNet-based TimesBlock.

[0039] The beneficial effect of the above embodiment is that ResNet performs two-dimensional convolution processing on the change data of the time series at different time scales, and utilizes ResNet's own powerful feature extraction capabilities and convergence speed advantages to improve the prediction accuracy and training efficiency of the model. In addition, according to the frequency component intensity corresponding to all periods under the time series, the features of the most significant periods are weighted and fused to obtain a TimesBlock based on ResNet; wherein the greater the frequency component intensity corresponding to a certain period, the greater the corresponding weighted proportion of the period in the weighted fusion process. By referencing the adaptive fusion strategy of weighted fusion, the features of different periods are weighted and fused, so that subsequent load prediction of different period lengths can be performed, with better adaptability and generalization capabilities.

[0040] In another embodiment, the combination stacking module is used to combine and stack multiple TimesBlocks to form an improved TimesNet network, including: Analyze multiple TimesBlocks, determine the residuals of all TimesBlocks, and then link multiple TimesBlock residuals to form an improved TimesNet network; The load prediction module is used to use the improved TimesNet network to perform load prediction on the resource utilization data of the cloud platform to obtain the resource demand load of the cloud platform in the future, including: The resource utilization data of the cloud platform is input into the improved TimesNet network to perform load prediction for periods of different lengths, and the resource demand load corresponding to the cloud platform in future time intervals of different lengths is obtained; wherein the resource demand load includes the data demand load of the CPU and / or memory.

[0041] The beneficial effect of the above embodiment is that each TimesBlock can perform load prediction in the corresponding period. In order to accurately predict the multi-periodicity and volatility of load demand, multiple TimesBlocks are analyzed to determine the residuals of all TimesBlocks, and then multiple TimesBlock residuals are linked to form an improved TimesNet network, so that the improved TimesNet network can achieve global accurate prediction of loads with multi-periodicity and volatility. Then the resource utilization data of the cloud platform is input into the improved TimesNet network to perform load prediction for periods of different lengths, and the resource demand load corresponding to the cloud platform in different time intervals in the future is obtained, which effectively reduces the scheduling lag and redundancy of load prediction, can adapt to the rapid changes in the cloud data center load more quickly, and improve the accuracy of long-term load prediction.

[0042] In general, the cloud data center load prediction method and system based on the improved TimesNet extracts multi-period features of the time series of resource utilization data, determines the most significant periods, and uses this to perform two-dimensional time series change modeling to obtain the change data of the time series at different time scales, thereby extracting the characteristics of the most significant periods, and adaptively merging the characteristics into a TimesBlock based on ResNet; multiple TimesBlocks are combined and stacked to form an improved TimesNet network, so as to predict the resource demand load of the cloud platform in the future, which can more accurately capture the multi-periodicity and volatility of the load, improve the accuracy of long-term prediction, and use ResNet for two-dimensional convolution, which can adapt to the rapid changes in the cloud data center load more quickly, reduce resource scheduling lags or redundancies through accurate load prediction, improve resource utilization efficiency, reduce operating costs caused by default and waste, and make load prediction have better adaptability and generalization capabilities.

[0043] The above is only a specific implementation of the present invention, and any other improvements made based on the concept of the present invention are considered to be within the protection scope of the present invention.

Claims

1. A cloud data center load prediction method based on improved TimesNet, characterized in that: include: The improved TimesNet model is used to extract multi-period features of the time series corresponding to the resource utilization data of the cloud platform, and the most significant periods in the time series are determined; based on the most significant periods, two-dimensional time series change modeling is performed to obtain the change data of the time series on different time scales; Processing the change data of the time series at different time scales, extracting the features of the most significant cycles, and adaptively fusing the features of the most significant cycles, thereby obtaining a TimesBlock based on ResNet; Multiple TimesBlocks are combined and stacked to form an improved TimesNet network; the improved TimesNet network is then used to perform load prediction on the resource utilization data of the cloud platform to obtain the resource demand load of the cloud platform in the future.

2. The cloud data center load prediction method based on improved TimesNet according to claim 1, characterized in that: The improved TimesNet model is used to extract multi-period features of the time series corresponding to the resource utilization data of the cloud platform, and the most significant periods in the time series are determined; based on the most significant periods, two-dimensional time series change modeling is performed to obtain the change data of the time series on different time scales, including: Obtain historical resource utilization data of a cloud platform, process the historical resource utilization data based on the generation time of the historical resource utilization data, and obtain a time series corresponding to the resource utilization data of the cloud platform; wherein the historical resource utilization data includes CPU and / or memory historical utilization data; perform fast Fourier transform on the time series using an improved TimesNet model to obtain the frequency component intensity of the time series, which is used as a multi-period feature of the time series; and then determine the most significant number of periods in the time series based on the frequency component intensity; According to the most significant cycles, the time series is folded according to the length corresponding to each cycle, so as to convert the one-dimensional time series into a two-dimensional tensor; wherein the width of the two-dimensional tensor is equal to the length of the corresponding cycle; and then a two-dimensional convolutional network is used to capture the temporal variation data of the two-dimensional tensor within and between cycles, which is used as the variation data of the time series on different time scales.

3. The cloud data center load prediction method based on improved TimesNet according to claim 2, characterized in that: The process of determining the most significant cycles in the time series is subjected to operation abnormality determination, including: Extracting and obtaining the period acquisition durations of several most significant periods in the time series; Comparing the period acquisition duration with a preset duration threshold; When the periodic acquisition duration exceeds a preset duration threshold, the processing duration of the historical resource utilization data and the frequency component intensity acquisition duration in the historical records are retrieved; Determine the process evaluation coefficients of the most significant several periods in the time series by using the processing time of the historical resource utilization data and the frequency component intensity acquisition time in the historical records; The process evaluation coefficient is obtained by the following formula: Wherein, S represents the process evaluation coefficient; n represents the number of times the most significant cycles in the time series are determined in the historical records except the current one; T 01i and T 02i They represent the processing time of historical resource utilization data and the acquisition time of frequency component intensity corresponding to the determination of the i-th period respectively; T zmax Indicates the maximum value of the cycle duration determined in n cycles; T z Indicates that the periodic acquisition duration corresponding to the preset duration threshold is exceeded; S 01 and S 02 denote the first coefficient and the second coefficient respectively, and the first coefficient and the second coefficient are obtained by the following formula: Among them, S 01 and S 02 Represent the first coefficient and the second coefficient respectively; T 01i and T 02i They represent the processing time of historical resource utilization data and the acquisition time of frequency component intensity corresponding to the determination of the i-th period respectively; T zmax Indicates the maximum value of the cycle duration determined in n cycles; T zmin Indicates the minimum value of the cycle acquisition duration in n cycles; comparing the process evaluation coefficient with a preset evaluation coefficient; When the process evaluation coefficient is not lower than the preset evaluation coefficient, it is determined that there are abnormalities in the processes of the most significant several cycles in the time series, and an abnormality alarm is issued.

4. The cloud data center load prediction method based on improved TimesNet according to claim 1, characterized in that: The change data of the time series at different time scales are processed to extract the features of the most significant cycles, and the features of the most significant cycles are adaptively fused to obtain a TimesBlock based on ResNet, including: Performing two-dimensional convolution processing on the change data of the time series at different time scales through ResNet to extract the features of each of the most significant cycles; According to the frequency component intensities corresponding to all the periods under the time series, the features of the most significant periods are weightedly fused to obtain a TimesBlock based on ResNet.

5. The cloud data center load prediction method based on improved TimesNet according to claim 1, characterized in that: Multiple TimesBlocks are combined and stacked to form an improved TimesNet network; the improved TimesNet network is then used to perform load prediction on the resource utilization data of the cloud platform to obtain the resource demand load of the cloud platform in the future, including: Analyze multiple TimesBlocks, determine the residuals of all TimesBlocks, and then link multiple TimesBlock residuals to form an improved TimesNet network; The resource utilization data of the cloud platform is input into the improved TimesNet network to perform load prediction for periods of different lengths, so as to obtain the resource demand load corresponding to the cloud platform in time intervals of different lengths in the future; wherein the resource demand load includes the data demand load of the CPU and / or memory.

6. A cloud data center load prediction system based on improved TimesNet, characterized by: A multi-period feature extraction module is used to extract multi-period features from the time series corresponding to the resource utilization data of the cloud platform using the improved TimesNet model, and determine the most significant periods in the time series; A two-dimensional time series change modeling module is used to perform two-dimensional time series change modeling according to the most significant cycles to obtain change data of the time series at different time scales; A two-dimensional convolution processing module is used to process the change data of the time series at different time scales to extract the characteristics of each of the most significant cycles; An adaptive fusion module, used for adaptively fusing the features of the most significant cycles, thereby obtaining a TimesBlock based on ResNet; The combination stacking module is used to combine and stack multiple TimesBlocks to form an improved TimesNet network; The load prediction module is used to use the improved TimesNet network to perform load prediction on the resource utilization data of the cloud platform to obtain the resource demand load of the cloud platform in the future.

7. The cloud data center load prediction system based on improved TimesNet according to claim 6, characterized in that: The multi-cycle feature extraction module is used to perform multi-cycle feature extraction on the time series corresponding to the resource utilization data of the cloud platform using the improved TimesNet model, and determine the most significant cycles in the time series, including: Obtain historical resource utilization data of a cloud platform, process the historical resource utilization data based on the generation time of the historical resource utilization data, and obtain a time series corresponding to the resource utilization data of the cloud platform; wherein the historical resource utilization data includes CPU and / or memory historical utilization data; perform fast Fourier transform on the time series using an improved TimesNet model to obtain the frequency component intensity of the time series, which is used as a multi-period feature of the time series; and then determine the most significant number of periods in the time series based on the frequency component intensity; The two-dimensional time series change modeling module is used to perform two-dimensional time series change modeling according to the most significant cycles to obtain change data of the time series on different time scales, including: According to the most significant cycles, the time series is folded according to the length corresponding to each cycle, so as to convert the one-dimensional time series into a two-dimensional tensor; wherein the width of the two-dimensional tensor is equal to the length of the corresponding cycle; and then a two-dimensional convolutional network is used to capture the temporal variation data of the two-dimensional tensor within and between cycles, which is used as the variation data of the time series on different time scales.

8. The cloud data center load prediction system based on improved TimesNet according to claim 7, characterized in that: The process of determining the most significant cycles in the time series is subjected to operation abnormality determination, including: Extracting and obtaining the period acquisition durations of several most significant periods in the time series; Comparing the period acquisition duration with a preset duration threshold; When the periodic acquisition duration exceeds a preset duration threshold, the processing duration of the historical resource utilization data and the frequency component intensity acquisition duration in the historical records are retrieved; Determine the process evaluation coefficients of the most significant several periods in the time series by using the processing time of the historical resource utilization data and the frequency component intensity acquisition time in the historical records; The process evaluation coefficient is obtained by the following formula: Wherein, S represents the process evaluation coefficient; n represents the number of times the most significant cycles in the time series are determined in the historical records except the current one; T 01i and T 02i They represent the processing time of historical resource utilization data and the acquisition time of frequency component intensity corresponding to the determination of the i-th period respectively; T zmax Indicates the maximum value of the cycle duration determined in n cycles; T z Indicates that the periodic acquisition duration corresponding to the preset duration threshold is exceeded; S 01 and S 02 denote the first coefficient and the second coefficient respectively, and the first coefficient and the second coefficient are obtained by the following formula: Among them, S 01 and S 02 Represent the first coefficient and the second coefficient respectively; T 01i and T 02i They represent the processing time of historical resource utilization data and the acquisition time of frequency component intensity corresponding to the determination of the i-th period respectively; T zmax Indicates the maximum value of the cycle duration determined in n cycles; T zmin Indicates the minimum value of the cycle acquisition duration in n cycles; comparing the process evaluation coefficient with a preset evaluation coefficient; When the process evaluation coefficient is not lower than the preset evaluation coefficient, it is determined that there are abnormalities in the processes of the most significant several cycles in the time series, and an abnormality alarm is issued.

9. The cloud data center load prediction system based on improved TimesNet according to claim 6, characterized in that: The two-dimensional convolution processing module is used to process the change data of the time series at different time scales to extract the characteristics of each of the most significant cycles, including: Performing two-dimensional convolution processing on the change data of the time series at different time scales through ResNet to extract the features of each of the most significant cycles; The adaptive fusion module is used to adaptively fuse the features of the most significant cycles to obtain a TimesBlock based on ResNet, including: According to the frequency component intensities corresponding to all the periods under the time series, the features of the most significant periods are weightedly fused to obtain a TimesBlock based on ResNet.

10. The cloud data center load prediction system based on improved TimesNet according to claim 6, characterized in that: The combination stacking module is used to combine and stack multiple TimesBlocks to form an improved TimesNet network, including: Analyze multiple TimesBlocks, determine the residuals of all TimesBlocks, and then link multiple TimesBlock residuals to form an improved TimesNet network; The load prediction module is used to use the improved TimesNet network to perform load prediction on the resource utilization data of the cloud platform to obtain the resource demand load of the cloud platform in the future, including: The resource utilization data of the cloud platform is input into the improved TimesNet network to perform load prediction for periods of different lengths, so as to obtain the resource demand load corresponding to the cloud platform in time intervals of different lengths in the future; wherein the resource demand load includes the data demand load of the CPU and / or memory.

Images