Load prediction method and device, program product and storage medium

Abstract

The invention provides a load prediction method and device, a program product and a storage medium, and the method comprises the steps: obtaining a first index data set which comprises load index data affecting the performance of a first platform; inputting the first index data set into a first prediction model, and predicting load condition information of the first platform according to the first index data set through the first prediction model; wherein the first prediction model comprises a multi-scale feature information extraction module, a global time sequence information extraction module and a full connection layer, and the multi-scale feature information extraction module is used for extracting multi-scale feature information of the first index data set; the global time sequence information extraction module is used for extracting global time sequence information of the first index data set according to the multi-scale feature information, and the full connection layer is used for determining a prediction result of the load condition information of the first platform according to the global time sequence information.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to a load prediction method, device, program product and storage medium. BACKGROUND

[0002] Currently, the expansion and contraction of containers and clusters mainly includes passive expansion and contraction relying on manual threshold setting and active expansion and contraction based on load index prediction. The active expansion and contraction strategy relies on accurate prediction of business load. The load mode of cloud applications usually has high dynamicity and complexity, and business peaks can be affected by various factors such as user behavior, time, holiday activities, etc. Common prediction methods include traditional statistical methods such as difference autoregressive moving average (ARIMA), traditional machine learning methods such as Bayesian ridge regression (BRR), and deep learning methods such as recurrent neural network (RNN). These single methods are usually difficult to capture these complex load change patterns, resulting in low prediction accuracy, causing resource overuse or deficiency, and affecting cost and performance. SUMMARY

[0003] To solve the above technical problems, the embodiments of the present application provide a load prediction method, device, program product and storage medium, which can accurately predict load condition information.

[0004] The load prediction method provided by the present application comprises: obtaining a first index data set, the first index data set comprising load index data affecting the performance of a first platform; inputting the first index data set into a first prediction model, and predicting the load condition information of the first platform according to the first index data set through the first prediction model; The first prediction model comprises a multi-scale feature information extraction module, a global time sequence information extraction module and a full connection layer. The multi-scale feature information extraction module is used to extract multi-scale feature information of the first index data set. The global time sequence information extraction module is used to extract global time sequence information of the first index data set according to the multi-scale feature information. The full connection layer is used to determine the prediction result of the load condition information of the first platform according to the global time sequence information.

[0005] The load prediction device provided by the present application comprises a processor and a memory. The memory is used to store a computer program. The processor is used to call and run the computer program stored in the memory to execute the above load prediction method.

[0006] The computer program product provided by the present application comprises a computer program. When the computer program is executed by a processor, the above load prediction method is realized.

[0007] The computer readable storage medium provided in the application is used for storing a computer program, and the computer program enables a computer to execute the load prediction method.

[0008] In the technical solution of the application, a first index data set is obtained, the first index data set including load index data affecting the performance of the first platform; the first index data set is input into a first prediction model, and the first prediction model is used to predict the load condition information of the first platform according to the first index data set; wherein the first prediction model includes a multi-scale feature information extraction module, a global time sequence information extraction module and a full connection layer, the multi-scale feature information extraction module is used to extract multi-scale feature information of the first index data set, the global time sequence information extraction module is used to extract global time sequence information of the first index data set according to the multi-scale feature information, and the full connection layer is used to determine the prediction result of the load condition information of the first platform according to the global time sequence information. In this way, the multi-scale feature information of the load index sequence is extracted by the multi-scale feature information extraction module, and the global time sequence information of the sequence is extracted by the global time sequence information extraction module, so that the multi-scale feature and the global time sequence feature of the sequence are considered during prediction, the complex feature in the load sequence can be accurately extracted, the long-term dependence problem is effectively solved, and more accurate prediction can be realized. BRIEF DESCRIPTION OF DRAWINGS

[0009] The accompanying drawings, which are included to provide a further understanding of the application and constitute a part of this application, illustrate certain illustrative embodiments of the application and together with the description serve to explain the application. The accompanying drawings are intended to be not limiting of the application and are included to illustrate and provide a conceptual base for understanding the application.

[0010] Figure 1 is a flowchart of the load prediction method provided by the embodiments of the application; Figure 2 is a general flowchart of the load prediction method provided by the embodiments of the application; Figure 3 is a schematic diagram of the multi-scale feature information extraction module provided by the embodiments of the application; Figure 4 is a schematic diagram of the global time sequence information extraction module provided by the embodiments of the application; Figure 5 is a schematic diagram of the structure of the load prediction device provided by the embodiments of the application; Figure 6 is a schematic structural diagram of a load prediction device provided by the embodiments of the application; Figure 7 is a schematic structural diagram of a chip according to the embodiments of the application. DETAILED DESCRIPTION

[0011] The technical solutions of the embodiments of this application will now be described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0012] In the following description, the term "some embodiments" refers to a subset of all possible embodiments. However, it is understood that "some embodiments" can be the same or different subsets of all possible embodiments and can be combined with each other without conflict. It should also be noted that the terms "first," "second," and "third" used in the embodiments of this application are only used to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first," "second," and "third" can be interchanged in a specific order or sequence where permissible, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein. The term "and / or" in this document is merely a description of the association relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship. It should also be understood that the "instruction" mentioned in the embodiments of this application can be a direct instruction, an indirect instruction, or an indication of an association relationship. For example, A instructing B can mean that A directly instructs B, for example, B can obtain information through A; it can also mean that A indirectly instructs B, for example, A instructs C, and B can obtain information through C; it can also mean that there is an association between A and B. It should also be understood that the term "correspondence" mentioned in the embodiments of this application can mean that there is a direct or indirect correspondence between the two, or that there is an association between the two, or that there is an instruction and being instructed, configuration and being configured, etc.

[0013] To facilitate understanding of the technical solutions of the embodiments of this application, the relevant technologies of the embodiments of this application are described below. The following relevant technologies are optional solutions and can be combined with the technical solutions of the embodiments of this application in any way, and they all fall within the protection scope of the embodiments of this application.

[0014] Currently, container and cluster scaling is mainly divided into passive scaling, which relies on manually set thresholds, and proactive scaling, which is based on load metric prediction. Proactive scaling strategies depend on accurate prediction of business load. Cloud application load patterns are typically highly dynamic and complex, and business peaks can be affected by various factors, such as user behavior, time, and holiday activities. Common prediction methods include traditional statistical methods such as Autoregressive Moving Average (ARIMA), traditional machine learning methods such as Bayesian Ridge Regression (BRR), and deep learning methods such as Recurrent Neural Networks (RNNs). These single methods often struggle to capture these complex load change patterns, resulting in low prediction accuracy, causing resource over- or under-utilization, and impacting cost and performance.

[0015] Currently, load forecasting based on artificial intelligence algorithms is widely used in scaling up and down tasks. The following are some current load forecasting methods: One approach proposes a cloud platform service load prediction method based on temporal convolutional neural networks (CNNs). This method applies temporal CNNs to load prediction tasks. Temporal CNNs are deep learning models specifically designed for processing sequential data. They combine the advantages of convolutional neural networks and are optimized for time-series data, resulting in higher accuracy compared to traditional time-series prediction models. Compared to recurrent neural networks (RNNs), temporal CNNs can process data in parallel, and their causal convolutional components enable varying receptive fields. However, the network structure needs to be pre-defined, such as the kernel size and number of layers. Deeper network layers achieve larger receptive fields but are prone to overfitting and loss of original information; shallower network layers retain more original information but can only extract small-scale information. Therefore, methods based solely on temporal CNNs cannot flexibly customize the network structure.

[0016] One approach proposes a Kubernetes-based load prediction elastic scaling system and method. This method combines the Prophet model and the LSTM model to predict load sequences, taking into account the influence of load on trend terms, periodic changes, and holiday effects.

[0017] In another approach, a cloud computing resource prediction method based on VMD and SSA-LSTM is proposed. This method addresses the potential non-stationarity and volatility in the original sequence through VMD variational mode decomposition. Furthermore, the Sparrow Search Algorithm (SSA) is used to optimize the parameters in the LSTM network structure, thereby improving prediction accuracy.

[0018] Both of the latter two methods are based on first decomposing the data and then using the LSTM algorithm for prediction. However, due to its characteristics, the LSTM method requires a large amount of computational resources and time for long sequence prediction tasks. Furthermore, for highly nonlinear sequences such as cloud loads, the prediction accuracy often falls short of expectations.

[0019] Therefore, improving the accuracy of load forecasting becomes a problem that needs to be considered. To this end, the following technical solutions based on embodiments of this application are proposed.

[0020] To facilitate understanding of the technical solutions of the embodiments of this application, the technical solutions of this application are described in detail below through specific embodiments. The above-mentioned related technologies are optional solutions and can be arbitrarily combined with the technical solutions of the embodiments of this application, all of which fall within the protection scope of the embodiments of this application. The embodiments of this application include at least some of the following contents.

[0021] Figure 1 This is a flowchart illustrating the load prediction method provided in an embodiment of this application, as shown below. Figure 1 As shown, the load forecasting method includes the following steps: Step 101: Obtain the first indicator data set, which includes load indicator data that affects the performance of the first platform.

[0022] Step 102: Input the first indicator data set into the first prediction model, and use the first prediction model to predict the load information of the first platform based on the first indicator data set.

[0023] The first prediction model includes a multi-scale feature information extraction module, a global time series information extraction module, and a fully connected layer. The multi-scale feature information extraction module is used to extract multi-scale feature information of the first indicator data set. The global time series information extraction module is used to extract global time series information of the first indicator data set based on the multi-scale feature information. The fully connected layer is used to determine the prediction result of the load information of the first platform based on the global time series information.

[0024] In this embodiment of the application, a first prediction model is used to predict the load of a first platform. The first prediction model includes a multi-scale feature information extraction module and a global time series information extraction module. The multi-scale feature information extraction module extracts multi-scale feature information of the load index sequence, and the global time series information extraction module extracts global time series information of the sequence. This can accurately extract complex features in the load sequence, effectively solve the long-term dependency problem, and achieve more accurate prediction.

[0025] In some implementations, load metrics include CPU utilization, memory usage, and traffic. Therefore, the load metric data would be CPU utilization data, memory data, and process data, etc. The first metric data set would then be a set of CPU utilization data, memory data, and process data, etc. The predicted load information for the first platform would then be the predicted CPU utilization, memory usage, and process data of the first platform. This application does not specifically limit the content of the load metrics; it can be determined based on the actual situation.

[0026] In some implementations, if there are multiple load metrics that commonly affect performance, the different load metrics can be weighted and combined into a new load metric variable.

[0027] In some implementations, the parameters of the first prediction model corresponding to each load metric are different; that is, the prediction model structure corresponding to each load metric is similar, but the parameters are different.

[0028] For example, when the first indicator data set is the CPU utilization data set, the CPU utilization data set is input into the first prediction model corresponding to the load indicator. After prediction by the first prediction model corresponding to the load indicator, the load prediction result of the first platform regarding CPU utilization is obtained.

[0029] It should be noted that the first indicator data set is a historical data set of a load indicator, which includes sampling information. For example, for CPU utilization, if sampling is performed once per minute, the CPU utilization in the (i+1)th minute can be predicted based on the data in the 1st minute, the 2nd minute, ... the ith minute.

[0030] In some embodiments, the method further includes: acquiring an initial prediction model and a first training data set, training the initial prediction model using the first training data set to obtain a first prediction model, wherein the initial prediction model includes an initial multi-scale feature information extraction module and an initial global temporal information extraction module.

[0031] In some implementations, to obtain more accurate prediction results, it is first necessary to determine the specific parameters of the first prediction model, i.e., to train the initial prediction model. Specifically, this involves obtaining the initial prediction model and the first training dataset. The initial prediction model is the model before training the first model, while the first prediction model is the model that can be directly used after training. The first training dataset is the dataset used to train the model. The initial prediction model is then trained using the first training dataset to obtain the first prediction model. The initial prediction model includes an initial multi-scale feature information extraction module and an initial global temporal information extraction module. After training, the initial multi-scale feature information extraction module becomes the multi-scale feature information extraction module in the first prediction model, and the initial global temporal information extraction module becomes the global temporal information extraction module in the first prediction model.

[0032] In some implementations, the load metric corresponding to the first training dataset is consistent with the load metric corresponding to the aforementioned first metric dataset, and the initial prediction model has the same structure as the aforementioned first prediction model. For example, the first training dataset is a training dataset of CPU utilization. After training the initial prediction model with the training dataset of CPU utilization, the resulting first prediction model is used to predict the CPU utilization information of the first platform.

[0033] It is understandable that for predictions of different load index data, the first prediction model includes a multi-scale feature information extraction module, a global temporal information extraction module, and a fully connected layer. However, the parameters of the first prediction model are different after training the corresponding initial prediction model with different load index training data. Similarly, the initial prediction model includes an initial multi-scale feature information extraction module, an initial global temporal information extraction module, and an initial fully connected layer.

[0034] In some implementations, obtaining a first training data set includes: obtaining an initial training data set, the initial training data set including historical load index data affecting the first platform; and determining the first training data set based on the initial training data set.

[0035] In some implementations, the initial training dataset consists of historical load metrics affecting the first platform. By mining the relationships between these historical load metrics, the load status of the first platform can be predicted. However, after obtaining the historical load metrics, further processing is required to obtain the first training dataset. Here, the load metrics corresponding to the initial training dataset and the load metrics corresponding to the first training dataset are consistent. For example, when training a first prediction model corresponding to CPU utilization, the first training dataset is the training dataset corresponding to CPU utilization, and the initial training dataset consists of historical load metrics data for CPU utilization, thus determining the first training dataset corresponding to CPU utilization.

[0036] It is understandable that the initial training dataset and the first metric dataset are not historical load metric data from the same time period. The initial training dataset contains earlier historical load metric data, while the first metric dataset contains historical load metric data from before the time period to be predicted. It should also be understood that the sampling period of the historical load metric data used as training data is the same as the sampling period in the first load metric dataset.

[0037] In some implementations, determining the first training data set based on the initial training data set includes: preprocessing the initial training data set to obtain a second training data set; decomposing the second training data set using a first decomposition method to obtain a sub-sequence data set, the sub-sequence data set being used to describe the changes in the second training data set; encoding the data in the second training data set to obtain a first feature vector; and constructing the first training data set based on the second training data set, the sub-sequence data set, and the first feature vector. Here, in order to extract information from the sequence data more accurately, the initial data set needs to be processed to obtain the corresponding first training data set.

[0038] In some implementations, the initial training dataset needs to be preprocessed to obtain a second training dataset. The prediction process includes deleting or replacing outliers, duplicates and missing values ​​in the initial training dataset according to the actual situation, then performing Z-Score normalization, and presetting a sliding window length according to the actual situation to divide the data into training set and test set according to a certain ratio.

[0039] In some implementations, the ratio of training set to test set is 7:3. The specific ratio can be determined according to the actual situation, and this application does not impose any specific limitations on it.

[0040] In some implementations, the first decomposition method is the STL decomposition method, where time series data can be decomposed into a trend term, a seasonal term, and a residual term. Specifically, the STL decomposition method is used to decompose the second training data set to obtain a subsequence data set, which consists of the trend term, seasonal term, and residual term corresponding to the second training data set. These subsequence data sets are used to describe the changes in the second training data set.

[0041] It is understandable that the data corresponding to each load metric is a time series.

[0042] In some implementations, the STL decomposition is shown in equation (1): (1) in, Represents the original sequence. Indicates the trend term. Indicates seasonal items, This represents the residual term. The original sequence and the three subsequences obtained after decomposition are vectorized respectively.

[0043] In some implementations, the encoding method is sine and cosine feature encoding. The time features are constructed by sine and cosine feature encoding, that is, the data in the second training data set is sine and cosine feature encoded to obtain the time features, and the encoded time features are vectorized to obtain the first feature vector.

[0044] In some implementations, the process of sine and cosine encoding the time variable is shown in equations (2) and (3): (2) (3) in, Let t be the value corresponding to the time series. Equations (2) and (3) indicate that by performing sine and cosine encoding on the time variable, the one-dimensional time variable value at time t can be converted into a two-dimensional spatial point. , This achieves continuity on the coordinate axis. The period value represents the length of the period, typically taking values ​​such as 60 minutes, 24 hours, or 7 days per week. Here, period corresponds to a specific load metric. Different load metrics correspond to different periods. The encoded time features are then vectorized.

[0045] In some implementations, the aforementioned second training data set and subsequence data set are vectorized respectively, and the vectorized subsequence data set, the vectorized second training data set, and the vectorized time features are merged to obtain the final first training data set.

[0046] Understandably, each load metric that needs to be predicted requires relevant operations.

[0047] In some implementations, the initial prediction model includes an initial fully connected layer; training the initial prediction model using a first training dataset to obtain a first prediction model includes: inputting the first training dataset into an initial multi-scale feature information extraction module to obtain multi-scale feature information of the first training dataset; inputting the multi-scale feature information of the first training dataset into an initial global temporal information extraction module to obtain global temporal information of the first training dataset; inputting the global temporal information of the first training dataset into an initial fully connected layer, and the initial fully connected layer outputs a first load prediction value; determining a first load error value based on the first load prediction value and the first true load value; adjusting the parameters of the initial prediction model so that the first load error value is the target load error value, thereby obtaining the first prediction model.

[0048] In some implementations, the first prediction model includes a multi-scale feature information extraction module and a global temporal information extraction module. The first training dataset is passed through both the multi-scale feature information extraction module and the global temporal information extraction module, and finally output through a fully connected layer. Therefore, the process of training the initial prediction model using the first training dataset to obtain the first prediction model is as follows: First, the first training dataset is input into the initial multi-scale feature information extraction module to obtain the multi-scale feature information of the first training dataset; then, the multi-scale feature information of the first training dataset is input into the initial global temporal information extraction module to obtain the global temporal information of the first training dataset; the global temporal information of the first training dataset is input into the initial fully connected layer, and the initial fully connected layer outputs the first load prediction value; the error value between the first load prediction value and the first load true value is calculated, i.e., the first load error value; by continuously adjusting the parameters of the initial prediction model until the first load error value is the target load error value, the parameters of the initial prediction model are the final parameters, thus obtaining the final first prediction model.

[0049] It is understandable that the first training data set here is a sequence corresponding to different load indicators. It should also be understood that the sequence of the first training data set here is t×5, where 5 represents 5 columns of data, namely the original data, three subsequence sets and time features.

[0050] In some implementations, the multi-scale feature extraction information module includes convolutional layers, fully connected layers, ReLU activation function layers, and Dropout layers, wherein the convolutional layers are formed by parallel superposition of causal dilation convolutions with different dilation factors.

[0051] In some implementations, inputting the first training data set into the multi-scale feature extraction information module includes: inputting the first training data set into a convolutional layer formed by parallel superimposed causal dilation convolutions with different dilation factors to obtain a first intermediate feature; inputting the first intermediate feature into a fully connected layer to obtain a second intermediate feature; inputting the second intermediate feature into a ReLU activation function layer to obtain a third intermediate feature; inputting the third intermediate feature into a Dropout layer to obtain a fourth intermediate feature; and determining the final multi-scale feature based on the fourth intermediate feature and the first training data set.

[0052] In some implementations, the multi-scale feature information extraction module includes two convolutional layers, two fully connected layers, two ReLU activation function layers, and two Dropout layers, wherein each convolutional layer is formed by parallel superposition of causal dilated convolutions with different dilation factors.

[0053] In some implementations, the structure of the multi-scale feature information extraction module is as follows: a first convolutional layer, a first fully connected layer, a first ReLU activation function layer, a first Dropout layer, a second convolutional layer, a second fully connected layer, a second ReLU activation function layer, and a second Dropout layer.

[0054] In some implementations, inputting the first training dataset into the multi-scale feature extraction information module includes: inputting the first training dataset into a first convolutional layer to obtain a first intermediate feature; inputting the first intermediate feature into a first fully connected layer to obtain a second intermediate feature; inputting the second intermediate feature into a first ReLU activation function layer to obtain a third intermediate feature; inputting the third intermediate feature into a first Dropout layer to obtain a fourth intermediate feature; inputting the fourth intermediate feature into a second convolutional layer to obtain a fifth intermediate feature; inputting the fifth intermediate feature into a second fully connected layer to obtain a sixth intermediate feature; inputting the sixth intermediate feature into a second ReLU activation function layer to obtain a seventh intermediate feature; inputting the seventh intermediate feature into a second Dropout layer to obtain an eighth intermediate feature; and determining the final multi-scale feature based on the eighth intermediate feature and the first training dataset.

[0055] In some implementations, the global temporal information extraction module includes a bidirectional gating unit and a global attention mechanism module.

[0056] In some implementations, inputting multi-scale features into a global temporal information extraction module includes inputting multi-scale features into a bidirectional gating unit to obtain a ninth intermediate feature, and inputting the ninth intermediate feature into a global attention mechanism module to obtain global temporal information.

[0057] In some implementations, global timing information is input into the fully connected layer to obtain the predicted load information of the first platform.

[0058] It is understandable that the initial multi-scale feature extraction information module and the multi-scale feature extraction information module have the same structure but different parameters, and the initial global temporal information extraction module and the global temporal information extraction module have the same structure but different parameters.

[0059] The technical solution provided in this application embodiment obtains a first indicator data set, which includes load indicator data affecting the performance of a first platform. The first indicator data set is input into a first prediction model, which then predicts the load status information of the first platform based on the first indicator data set. The first prediction model includes a multi-scale feature information extraction module, a global temporal information extraction module, and a fully connected layer. The multi-scale feature information extraction module extracts multi-scale feature information from the first indicator data set, the global temporal information extraction module extracts global temporal information from the first indicator data set based on the multi-scale feature information, and the fully connected layer determines the prediction result of the load status information of the first platform based on the global temporal information. Thus, by extracting multi-scale feature information of the load indicator sequence through the multi-scale feature information extraction module and extracting global temporal information of the sequence through the global temporal information extraction module, the prediction takes into account both the multi-scale and global temporal features of the sequence, accurately extracting complex features in the load sequence, effectively solving the long-term dependency problem, and achieving more accurate prediction.

[0060] Based on the foregoing embodiments, the load prediction method provided by the embodiments of this application will be further described.

[0061] To address the current inability to accurately predict cloud platform load, this application proposes the following solution: By improving the residual blocks of the temporal convolutional network model, a multi-scale feature extraction module is proposed. This module overlays multiple residual blocks with different dilation factors and performs causal dilation convolutions with different dilation factors in parallel to extract feature information at different scales. This solves the problem of simultaneously achieving a larger receptive field and preserving original information. A global temporal information extraction module is designed by combining the BiGRU model and the GlobalAttention mechanism. This achieves richer feature representation extraction and more efficient capture of long-distance dependencies. This application combines multi-scale feature extraction and global temporal information extraction to propose a load prediction method that achieves high prediction accuracy. Figure 2 This is a schematic diagram of the overall flow of the load prediction method provided in the embodiments of this application, as shown below. Figure 2As shown, historical load data is acquired, preprocessed, and then decomposed into time-period features. Finally, the model is trained to obtain the prediction result. Specifically, historical load data corresponding to the load metric to be predicted is first obtained from the platform. After data preprocessing, the data enters the sequence decomposition module. Sequence decomposition breaks down the original load sequence into subsequences with different meanings. Time-period features are then constructed, and these data are proportionally divided and fed into the prediction model designed in this proposal for training. Finally, the predicted result is obtained through the trained model.

[0062] It should be noted that different prediction models are used for different load metrics. These different prediction models have the same structure but different parameters. Furthermore, the historical load data is a time series.

[0063] The specific implementation steps are as follows: Step S1: Retrieve historical load data from the database. This typically includes load metrics that primarily affect platform service performance, such as CPU utilization, memory usage, and traffic. If multiple metrics jointly impact performance, they can be weighted and combined into a new metric variable.

[0064] Step S2: Identify outliers, duplicates, and missing values ​​in the original data, and delete or replace them as needed. Then, perform Z-score normalization, preset the sliding window length according to the actual situation, and divide the data into training and test sets in a 7:3 ratio. In some implementations, outliers or missing values ​​can be replaced using the average of adjacent data.

[0065] Step S3: The sequence decomposition process in this step is based on the STL decomposition method. Time series data can be decomposed into: trend, seasonal, and residual terms, and its decomposition formula can generally be expressed as shown in equation (1). The original sequence and the three subsequences obtained after decomposition are vectorized respectively.

[0066] Step S4: This proposal uses sine and cosine feature encoding to construct time features. The specific construction process is shown in equations (2) and (3). Equations (2) and (3) indicate that by performing sine and cosine encoding on the time variable, the one-dimensional time variable value at time t can be converted into a two-dimensional spatial point ( , This achieves continuity on the coordinate axis. The value of `period` represents the length of the period, typically taking values ​​such as 60 minutes, 24 hours, or 7 days per week. The encoded time features are vectorized and merged with the sequence generated in step S3.

[0067] It is understandable that steps S3 and S4 are required for time series of different load indicators.

[0068] Step S5: The prediction model designed in this proposal includes a multi-scale feature information extraction module and a global temporal information extraction module. The merged sequence generated in step S4 is then passed through the multi-scale feature information extraction module and the global temporal information extraction module, and finally output through a fully connected layer.

[0069] S5.1: The multi-scale feature information extraction module improves the TCN model to achieve multi-scale feature information extraction. The data input for this module is the sequence merged and constructed in step 4. Figure 3 This is a schematic diagram of the multi-scale feature information extraction module provided in an embodiment of this application. For example... Figure 3 As shown, the multi-scale feature information extraction module of this application embodiment is a schematic diagram based on the improved TCN model residual block. It includes two convolutional layers formed by parallel superposition of causal dilation convolutions with different dilation factors, two fully connected layers, two ReLU activation function layers, and two Dropout layers, where the improvement lies in the convolutional layers. The data processed by this layer enters the global temporal information extraction module. It can be understood that... Figure 3 The dilation factor n in the convolutional layer is also a parameter that needs to be determined during training.

[0070] S5.2: Figure 4 This is a schematic diagram of the global timing information extraction module provided in an embodiment of this application, as shown below. Figure 4 The diagram shows the global temporal information extraction module, which extracts global temporal information by using a superimposed bidirectional gating unit (BiGRU) and a global attention mechanism (GlobalAttention).

[0071] Step S5.3: Build the above model using code, and repeatedly adjust the parameters by comparing the predicted results with the actual values ​​until the error value is minimized.

[0072] This is the forecasting process for each load metric.

[0073] Based on the above steps, this application implements a cloud platform load prediction method based on the fusion of multi-scale features and global time-series information. Through sequence decomposition and time-period feature construction, the model can better identify periodic features in the load sequence. The multi-scale feature information extraction module extracts complex feature information from the load sequence, and the global time-series information extraction module solves the long-term dependency problem of the load sequence. Compared with traditional prediction models, the load prediction method proposed in this application can effectively improve the accuracy of prediction and provide algorithmic support for container and cluster proactive elastic scaling services.

[0074] The key points of this application's embodiments are the implementation process of the multi-scale feature information extraction module and the global time series information extraction module in the prediction model, and the cloud platform load prediction method based on the fusion of multi-scale features and global time series information.

[0075] Most existing technologies employ prediction methods based on traditional machine learning algorithms or their variants. However, when faced with highly dynamic and complex cloud platform load data, existing prediction methods cannot effectively extract the information contained in the load sequence, thus failing to achieve accurate predictions. This application's embodiments, through sequence decomposition, periodic feature construction, and the design of a multi-scale feature information extraction module and a global time-series information extraction module, can accurately extract multi-scale feature information from the load sequence and effectively solve the long-term dependency problem, enabling more accurate predictions.

[0076] It should be noted that the embodiments of this application employ the STL decomposition method to decompose the load time series into trend, seasonal, and residual terms. This method effectively captures the periodic characteristics of cloud platform load. However, other decomposition methods such as EMD (Empirical Mode Decomposition) or VMD (Variational Mode Decomposition) can also be used in similar scenarios and may provide better performance when dealing with nonlinear or non-stationary series. Different decomposition methods can lead to different technical paths. Specifically, this application does not limit the specific decomposition method and can determine it according to the actual situation.

[0077] It should also be noted that other temporal information extraction models can be used. Besides the BiGRU combined with a global attention mechanism mentioned in the embodiments of this application, Transformer models or BERT models can also be considered. These models have demonstrated powerful temporal dependency information extraction capabilities in natural language processing tasks. Load prediction schemes based on these models are expected to provide higher prediction accuracy when capturing long-sequence dependencies.

[0078] It should also be noted that the multi-scale feature extraction in this embodiment is achieved through an improved TCN model, employing causal convolutional layers with different dilation factors to obtain information at different scales. Multi-resolution analysis methods, such as wavelet transform or multi-level LSTM networks, can also be considered to process the different scale features of time series data. These methods can capture multi-scale information in the data in different ways.

[0079] Most existing cloud platform load forecasting technologies are based on LSTM or its variants, such as LSTM models optimized with VMD and SSA. However, the embodiments in this application focus on the combination of TCN and BiGRU, and achieve multi-scale and global feature extraction through innovative module design. This combination and innovative module design significantly differ from the traditional LSTM architecture.

[0080] To circumvent widely used general load forecasting models, embodiments of this application can also design dedicated forecasting models for specific domains (such as network function virtualization or specific types of cloud services). These models may have unique advantages when processing specific types of load data.

[0081] The cloud platform load prediction method described in this application can effectively improve the prediction accuracy of cloud platform load and enable more accurate and efficient proactive elastic scaling operations. It is a supplement and optimization to existing elastic scaling solutions. It can be used in various cloud platform elastic scaling scenarios and can also be applied to other industries.

[0082] Figure 5 This is a schematic diagram of the structural composition of the load prediction device provided in the embodiments of this application, as shown below. Figure 5 As shown, the load prediction device includes: The acquisition unit 501 is used to acquire a first indicator data set, which includes load indicator data that affects the performance of the first platform; The prediction unit 502 is used to input the first indicator data set into the first prediction model, and predict the load information of the first platform based on the first indicator data set through the first prediction model. The first prediction model includes a multi-scale feature information extraction module, a global time series information extraction module, and a fully connected layer. The multi-scale feature information extraction module is used to extract multi-scale feature information of the first indicator data set. The global time series information extraction module is used to extract global time series information of the first indicator data set based on the multi-scale feature information. The fully connected layer is used to determine the prediction result of the load information of the first platform based on the global time series information.

[0083] In some embodiments, the acquisition unit 501 is used to acquire an initial prediction model and a first training data set; the device further includes a training unit 503; the training unit 503 is used to train the initial prediction model using the first training data set to obtain a first prediction model, wherein the initial prediction model includes an initial multi-scale feature information extraction module and an initial global temporal information extraction module.

[0084] In some implementations, the acquisition unit 501 is used to acquire an initial training data set, which includes historical load index data affecting the first platform; and to determine a first training data set based on the initial training data set.

[0085] In some embodiments, the acquisition unit 501 is used to preprocess the initial training data set to obtain a second training data set; decompose the second training data set using a first decomposition method to obtain a subsequence data set, the subsequence data set being used to describe the changes in the second training data set; encode the data in the second training data set to obtain a first feature vector; and construct a first training data set based on the second training data set, the subsequence data set, and the first feature vector.

[0086] In some implementations, the initial prediction model includes an initial fully connected layer; the training unit 503 is used to input the first training dataset into an initial multi-scale feature information extraction module to obtain multi-scale feature information of the first training dataset; input the multi-scale feature information of the first training dataset into an initial global temporal information extraction module to obtain global temporal information of the first training dataset; input the global temporal information of the first training dataset into the initial fully connected layer, and the initial fully connected layer outputs a first load prediction value; determine a first load error value based on the first load prediction value and the first load true value; adjust the parameters of the initial prediction model so that the first load error value is the target load error value, thereby obtaining the first prediction model.

[0087] In some implementations, the multi-scale feature extraction information module includes convolutional layers, fully connected layers, ReLU activation function layers, and Dropout layers, wherein the convolutional layers are formed by parallel superposition of causal dilation convolutions with different dilation factors.

[0088] In some implementations, the global temporal information extraction module includes a bidirectional gating unit and a global attention mechanism module.

[0089] Those skilled in the art should understand that Figure 5 The functions of each unit in the load prediction device shown can be understood by referring to the relevant description of the aforementioned method. Figure 5 The functions of each unit in the load prediction device shown can be implemented by a program running on a processor or by specific logic circuits.

[0090] Figure 6 This is a schematic structural diagram of a load prediction device 600 provided in an embodiment of this application. Figure 6 The load prediction device 600 shown includes a processor 610, which can call and run computer programs from memory to implement the methods in the embodiments of this application.

[0091] Optionally, such as Figure 6 As shown, the load prediction device 600 may further include a memory 620. The processor 610 can retrieve and run computer programs from the memory 620 to implement the methods described in this embodiment.

[0092] The memory 620 can be a separate device independent of the processor 610, or it can be integrated into the processor 610.

[0093] Optionally, such as Figure 6 As shown, the load prediction device 600 may also include a transceiver 630, which the processor 610 can control to communicate with other devices. Specifically, it can send information or data to other devices or receive information or data sent by other devices.

[0094] The transceiver 630 may include a transmitter and a receiver. The transceiver 630 may further include antennas, and the number of antennas may be one or more.

[0095] The load prediction device 600 can implement the corresponding processes implemented by the load prediction device in the various methods of the embodiments of this application, which will not be described in detail here for the sake of brevity.

[0096] Figure 7 This is a schematic structural diagram of the chip according to an embodiment of this application. Figure 7 The chip 700 shown includes a processor 710, which can call and run computer programs from memory to implement the methods in the embodiments of this application.

[0097] Optionally, such as Figure 7 As shown, chip 700 may further include memory 720. Processor 710 can retrieve and run computer programs from memory 720 to implement the methods described in this embodiment.

[0098] The memory 720 can be a separate device independent of the processor 710, or it can be integrated into the processor 710.

[0099] Optionally, the chip 700 may also include an input interface 730. The processor 710 can control the input interface 730 to communicate with other devices or chips; specifically, it can acquire information or data sent by other devices or chips.

[0100] Optionally, the chip 700 may also include an output interface 740. The processor 710 can control the output interface 740 to communicate with other devices or chips, specifically, to output information or data to other devices or chips.

[0101] This chip can implement the corresponding processes implemented by the load prediction device in the various methods of the embodiments of this application, which will not be described in detail here for the sake of brevity.

[0102] It should be understood that the chip mentioned in the embodiments of this application may also be referred to as a system-on-a-chip, system chip, chip system, or system-on-a-chip, etc.

[0103] It should be understood that the processor in the embodiments of this application may be an integrated circuit chip with signal processing capabilities. In implementation, the steps of the above method embodiments can be completed by integrated logic circuits in the processor's hardware or by instructions in software form. The processor described above can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method.

[0104] It is understood that the memory in the embodiments of this application can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDR SDRAM), Enhanced Synchronous DRAM (ESDRAM), Synchlink DRAM (SLDRAM), and Direct Rambus RAM (DR RAM). It should be noted that the memory used in the systems and methods described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0105] It should be understood that the above-described memory is exemplary and not a limiting description. For example, the memory in the embodiments of this application may also be static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct memory bus RAM (DR RAM), etc. That is to say, the memory in the embodiments of this application is intended to include, but is not limited to, these and any other suitable types of memory.

[0106] This application also provides a computer program product, including a computer program.

[0107] When executed by a processor, the computer program implements the corresponding processes of the load prediction device in the various methods of the embodiments of this application, which will not be described in detail here for the sake of brevity.

[0108] This application also provides a computer-readable storage medium for storing computer programs.

[0109] The computer program causes the computer to execute the corresponding processes implemented by the load prediction device in the various methods of the embodiments of this application, which will not be described in detail here for the sake of brevity.

[0110] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0111] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0112] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0113] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0114] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0115] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0116] The preferred embodiments of this application have been described in detail above with reference to the accompanying drawings. However, this application is not limited to the specific details of the above embodiments. Within the scope of the technical concept of this application, various simple modifications can be made to the technical solutions of this application, and these simple modifications all fall within the protection scope of this application. For example, the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, this application will not describe the various possible combinations separately. Furthermore, various different embodiments of this application can also be arbitrarily combined, as long as they do not violate the spirit of this application, they should also be considered as the content disclosed in this application. Moreover, without conflict, the various embodiments and / or the technical features in the various embodiments described in this application can be arbitrarily combined with the prior art, and the resulting technical solutions should also fall within the protection scope of this application.

[0117] It should be understood that in the various method embodiments of this application, the sequence number of each process does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0118] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A load forecasting method, characterized in that, The method includes: Obtain a first set of indicator data, which includes load indicator data that affects the performance of the first platform; The first indicator data set is input into the first prediction model, and the first prediction model predicts the load information of the first platform based on the first indicator data set. The first prediction model includes a multi-scale feature information extraction module, a global time series information extraction module, and a fully connected layer. The multi-scale feature information extraction module is used to extract multi-scale feature information of the first indicator data set. The global time series information extraction module is used to extract global time series information of the first indicator data set based on the multi-scale feature information. The fully connected layer is used to determine the prediction result of the load information of the first platform based on the global time series information.

2. The method according to claim 1, characterized in that, The method further includes: An initial prediction model and a first training data set are obtained, and the initial prediction model is trained using the first training data set to obtain the first prediction model. The initial prediction model includes an initial multi-scale feature information extraction module and an initial global temporal information extraction module.

3. The method according to claim 2, characterized in that, Obtaining the first training data set includes: Obtain an initial training data set, which includes historical load index data affecting the first platform; The first training data set is determined based on the initial training data set.

4. The method according to claim 3, characterized in that, Determining the first training data set based on the initial training data set includes: The initial training data set is preprocessed to obtain the second training data set; The second training data set is decomposed using a first decomposition method to obtain a subsequence data set, which is used to describe the changes in the second training data set. The data in the second training dataset is encoded to obtain the first feature vector; The first training data set is constructed based on the second training data set, the subsequence data set, and the first feature vector.

5. The method according to claim 2, characterized in that, The initial prediction model includes an initial fully connected layer; training the initial prediction model using the first training dataset to obtain the first prediction model includes: The first training dataset is input into the initial multi-scale feature information extraction module to obtain the multi-scale feature information of the first training dataset. The multi-scale feature information of the first training dataset is input into the initial global temporal information extraction module to obtain the global temporal information of the first training dataset. The global time series information of the first training dataset is input into the initial fully connected layer, and the initial fully connected layer outputs the first load prediction value. The first load error value is determined based on the first predicted load value and the first actual load value; The parameters of the initial prediction model are adjusted so that the first load error value is the target load error value, thus obtaining the first prediction model.

6. The method according to any one of claims 1 to 5, characterized in that, The multi-scale feature extraction information module includes a convolutional layer, a fully connected layer, a ReLU activation function layer, and a Dropout layer. The convolutional layer is formed by parallel superposition of causal dilation convolutions with different dilation factors.

7. The method according to any one of claims 1 to 5, characterized in that, The global temporal information extraction module includes a bidirectional gating unit and a global attention mechanism module.

8. A load prediction device, characterized in that, include: A processor and a memory for storing a computer program, the processor for calling and running the computer program stored in the memory to perform the method as described in any one of claims 1 to 7.

9. A computer program product, characterized in that, include: A computer program that, when executed by a processor, implements the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store a computer program that causes a computer to perform the method as described in any one of claims 1 to 7.

Images