Cloud resource load prediction method and device based on data features, equipment and medium

Abstract

The invention discloses a cloud resource load prediction method and device based on data features, equipment and a medium, and relates to the technical field of cloud computing technologies and data centers, and the method comprises the steps: obtaining operation data in a preset time period during the operation of a cloud platform; inputting the operation data into a pre-trained forest selection model to obtain key parameters output by the forest selection model; performing decomposition operation on the key parameters based on a modal decomposition algorithm to obtain a plurality of subsequences; constructing a two-dimensional sub-sequence, and inputting the obtained two-dimensional sub-sequence into a pre-trained load prediction model to obtain a prediction value corresponding to the two-dimensional sub-sequence; and performing summation processing on each predicted value, and determining an obtained summation result as a cloud resource load predicted value. The method and the device are used for solving the problem of poor accuracy of a prediction result during cloud resource load prediction in the prior art, and the accuracy of the prediction result is improved by utilizing combination of multiple networks and extraction of hidden features.

Description

Technical Field

[0001] The present application relates to the fields of cloud computing technology and data center technology, and in particular to a method, device, equipment and medium for predicting cloud resource load based on data features. Background Art

[0002] Cloud computing has powerful network computing capabilities and can process large amounts of data in a short period of time based on distributed parallel computing. With the growing demand for cloud services, the resource allocation of cloud platforms has brought challenges to the security, service quality and cost of distributed cloud platforms, and the increase in workloads has put tremendous pressure on the data storage and memory capacity of distributed cloud platforms. Therefore, how to achieve efficient resource management is the main problem that cloud platforms need to solve.

[0003] Existing technologies use cloud resource load prediction to efficiently manage cloud resources, avoid over-allocation or under-allocation of resources, and help cloud companies make decisions in advance. However, in cloud computing, various types of computing resources are consumed unevenly, and due to the high dynamics and irregularity of cloud resource load characteristics, cloud resource load prediction is challenging.

[0004] Existing technologies predict cloud resource load by training prediction models, such as traditional time series models, machine learning models, and deep learning models. However, time series models are mainly used for linear stationary time series, and have poor prediction performance for highly volatile, non-stationary, and long time series; machine learning models are not robust to load data with long time series, and have poor prediction performance; deep learning models have the problem of long time series loss, resulting in inaccurate prediction results. Summary of the invention

[0005] In response to the above-mentioned problems and technical needs, the applicant has proposed a cloud resource load prediction method, device, equipment and medium based on data features to solve the problem of poor prediction accuracy of cloud resource load prediction in the prior art, and to improve the accuracy of prediction results by utilizing the combination of multiple networks and the extraction of hidden features.

[0006] The present application provides a cloud resource load prediction method based on data features, the method comprising:

[0007] Obtaining operation data within a preset time period when the cloud platform is running, wherein the operation data includes: CPU utilization, memory utilization, incoming network packets, outgoing network packets and disk space utilization;

[0008] Inputting the operating data into a pre-trained forest selection model to obtain key parameters output by the forest selection model, wherein the key parameter is any item in the operating data, and the forest selection model is trained based on operating data samples and key parameter samples;

[0009] Decomposing the key parameters based on a modal decomposition algorithm to obtain a plurality of subsequences;

[0010] Constructing a two-dimensional subsequence, inputting the obtained two-dimensional subsequence into a pre-trained load prediction model, and obtaining hidden features corresponding to the two-dimensional subsequence through an attention mechanism embedded in a convolutional neural network; inputting the hidden features into a long short-term memory network to obtain a predicted value corresponding to the two-dimensional subsequence, wherein the load prediction model is trained based on two-dimensional subsequence samples and predicted value samples;

[0011] Each of the predicted values ​​is summed up, and the summed result is determined as the cloud resource load prediction value.

[0012] According to a cloud resource load prediction method based on data features in one embodiment of the present application, the hidden features corresponding to the two-dimensional subsequence are obtained by using an attention mechanism embedded in a convolutional neural network, including:

[0013] Obtaining a first feature map through a first convolutional layer of a convolutional neural network, and obtaining a second feature map through a second convolutional layer of the convolutional neural network;

[0014] Generate an average feature corresponding to a global average pooling channel and a maximum feature corresponding to a maximum pooling channel based on the first feature map and the second feature map;

[0015] Based on the attention mechanism, the hidden feature is obtained by using the average feature and the maximum feature.

[0016] According to a cloud resource load prediction method based on data features in one embodiment of the present application, the hidden features corresponding to the two-dimensional subsequence are obtained by using an attention mechanism embedded in a convolutional neural network, including:

[0017] Obtaining the hidden features through a feature extraction formula;

[0018] Wherein, the feature extraction formula includes:

[0019] F 1 =Conv 1 (IMF(t));

[0020] F 2 =Conv 2 (F 1 );

[0021]

[0022] F c =[F avg ,F max ];

[0023] S c =σ(W 2 δ(W 1 F c ));

[0024] F cout =S c ⊙F 2 ;;

[0025] F out =Conv next (F cout );

[0026] Among them, F 1 represents the first feature map, IMF(t) represents a two-dimensional subsequence, Conv 1 (IMF(t)) represents the convolution operation on the two-dimensional subsequence through the first convolution layer, F 2 Represents the second feature map, Conv 2 (F 1 ) indicates that the first feature map is convolved by the second convolutional layer, F avg represents the average feature, F max represents the maximum feature, F 2 [:,i,j] represents the feature values ​​of all channels at the (i, j) position of the second feature map, F c Indicates F avg and F max The union of c represents the channel attention weight, which is used to characterize the importance of the channel, σ represents the sigmoid function, and W 2 represents the weight matrix of the second fully connected layer in the convolutional neural network, δ represents the activation function, W 1 represents the weight matrix of the first fully connected layer in the convolutional neural network, F cout represents the new feature, F out represents the final hidden features.

[0027] According to a cloud resource load prediction method based on data features in one embodiment of the present application, the hidden features are input into a long short-term memory network to obtain a prediction value corresponding to the two-dimensional subsequence, including:

[0028] Obtaining the predicted value through a prediction formula;

[0029] Wherein, the prediction formula includes:

[0030] f t =σ(W (f) IMFP (t) +U (f) h (t-1) );

[0031] i t =σ(W ih h t-1 +W iIMFP IMFP (t) +b i );

[0032] o t =o(W (o) IMFP (t) +U (o) h (t-1) );

[0033]

[0034] h t =o t ·tanh(c t );

[0035]

[0036] IMFY k (t) = W p +h t +b p ;

[0037] Among them, IMFY k (t) represents the predicted value of the kth MF component at time t, h (t-1) Indicates the hidden state of the previous moment, IMFP (t) represents the one-dimensional feature corresponding to the hidden feature, W ih is the weight in the input gate weight matrix, used to connect IMFP (t) and h (t-1) The interaction, i t Indicates the state of the input gate, o t Indicates the state of the output gate, W (f) represents the weight matrix of the forget gate, U (f) represents the recursive weight matrix of the forget gate, W ( i ) represents the weight matrix of the input gate, W (o) represents the weight matrix of the output gate, U (o) represents the recursive weight matrix of the output gate, W (c) The weight matrix representing the candidate cell state, U (c)Represents the recursive weight matrix of candidate cell states, b i represents the bias term of the input gate, i t represents the activation value of the input gate, W p represents the weight matrix, b p represents the bias term, C t represents the cell state at time t, h t represents the hidden state at the current moment, W (i)IMFP represents another weight matrix of the input gate, represents the candidate cell state, C (t-1) Indicates the cell state at the previous moment.

[0038] According to a cloud resource load prediction method based on data features in one embodiment of the present application, the operation data is input into a pre-trained forest selection model to obtain key parameters output by the forest selection model, including:

[0039] Inputting the operating data into the forest selection model, determining the importance score of each parameter through the forest selection model, and determining the parameter corresponding to the maximum importance score as the key parameter;

[0040] Among them, the importance score of each parameter is determined by the scoring formula:

[0041]

[0042] Among them, MSE before represents the mean square error before splitting, N represents the number of running data in the g-th forest node, and y i represents the target value corresponding to the g-th forest node, Represents the sample mean corresponding to the g-th forest node, MSE after represents the mean square error after splitting, N L represents the number of running data in the left child node of the g-th forest node, N R represents the number of running data in the right child node of the g-th forest node, MSE L Represents the mean square error corresponding to the left child node, MSE R represents the mean square error corresponding to the right child node, f i represents the importance score, T i Represents all forest nodes that undergo splitting.

[0043] According to a cloud resource load prediction method based on data features in one embodiment of the present application, the key parameters include: the CPU utilization;

[0044] The key parameters are decomposed based on a modal decomposition algorithm to obtain multiple subsequences, including:

[0045] Adding white noise to the CPU utilization to form superimposed data; decomposing the superimposed data using an empirical mode decomposition method to obtain a first IMF and a residual component, and using an average value to eliminate the influence of the white noise;

[0046] The residual component is added with white noise for the CPU utilization to form the superimposed data; the superimposed data is decomposed by using the empirical mode decomposition method to obtain IMF and residual components, and the influence of the white noise is eliminated by using the average value, until the residual component cannot be decomposed.

[0047] According to a cloud resource load prediction method based on data features in one embodiment of the present application, the key parameters are decomposed based on a modal decomposition algorithm to obtain multiple subsequences, including:

[0048] Based on the decomposition formula, multiple subsequences are obtained;

[0049] Wherein, the decomposition formula includes:

[0050] IMF 1 i (t)=cpu(t)+ε 0 ω i (t),i=1,2,3,...,n;

[0051]

[0052] in, represents the first IMF component obtained after adding white noise for the ith time, cpu(t) represents the CPU utilization, ε 0 represents the signal-to-noise ratio, t represents the time point, i represents white noise, ω i (t) represents the white noise sequence obtained after adding white noise for the ith time, n represents the total number of times white noise is added, It represents the average value of the first intrinsic mode function (IMF) obtained by averaging the added results after adding white noise multiple times. It means that the IMF components obtained after adding white noise each time are averaged to obtain a smooth result, EMD 1 (f i (t)) represents the process of using empirical mode decomposition after adding white noise for the i-th time, which is used to extract the first IMF. represents the average value of the kth intrinsic mode function (IMF) obtained by averaging the added results after adding white noise multiple times, K represents the total number of IMFs, r 1 (t) represents the first residual component, r k-1 (t) k-1th residual component, r k(t) represents the kth residual component, R(t) represents the final residual component, ε k-1 represents adaptive white noise.

[0053] The embodiment of the present application also provides a cloud resource load prediction device based on data features, including:

[0054] An acquisition module is used to acquire operation data within a preset time period when the cloud platform is running, wherein the operation data includes: CPU utilization, memory utilization, incoming network packets, outgoing network packets and disk space utilization;

[0055] A key feature extraction module, used for inputting the operation data into a pre-trained forest selection model to obtain key parameters output by the forest selection model, wherein the key parameter is any item in the operation data, and the forest selection model is trained based on the operation data sample and the key parameter sample;

[0056] A decomposition module, used for performing a decomposition operation on the key parameters based on a modal decomposition algorithm to obtain a plurality of subsequences;

[0057] A prediction module is used to construct a two-dimensional subsequence, input the obtained two-dimensional subsequence into a pre-trained load prediction model, obtain hidden features corresponding to the two-dimensional subsequence through an attention mechanism embedded in a convolutional neural network; input the hidden features into a long short-term memory network to obtain a predicted value corresponding to the two-dimensional subsequence, wherein the load prediction model is trained based on two-dimensional subsequence samples and predicted value samples;

[0058] The summing module is used to sum each of the predicted values ​​and determine the summed result as the cloud resource load prediction value.

[0059] An embodiment of the present application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of the cloud resource load prediction method based on data features as described in any one of the above items are implemented.

[0060] An embodiment of the present application also provides a non-transitory computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the cloud resource load prediction method based on data features as described in any of the above items are implemented.

[0061] The cloud resource load prediction method, device, equipment and medium based on data features provided in the embodiments of the present application obtain the operation data within a preset time period when the cloud platform is running, wherein the operation data includes: CPU utilization, memory utilization, incoming network packets, outgoing network packets and disk space utilization; the operation data is input into a pre-trained forest selection model to obtain key parameters output by the forest selection model, wherein the key parameters are any item in the operation data. It can be seen that the present application uses the forest selection model to select important features, i.e., key parameters, from the operation data; then, the key parameters are decomposed based on the modal decomposition algorithm to obtain multiple subsequences; a two-dimensional subsequence is constructed, and the obtained two-dimensional subsequence is input into a pre-trained load prediction model, and the hidden features corresponding to the two-dimensional subsequence are obtained through the attention mechanism embedded in the convolutional neural network; the hidden features are input into a long short-term memory network to obtain the prediction value corresponding to the two-dimensional subsequence, and the present application uses the attention mechanism to effectively extract the hidden features of the data to reduce the interference of invalid features, thereby improving the accuracy of the prediction results; each prediction value is summed, and the summed result is determined as the cloud resource load prediction value, and finally the combination of multiple networks and the extraction of hidden features are realized to improve the accuracy of the prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0063] Figure 1 This is one of the flow charts of the cloud resource load prediction method based on data features provided in the embodiment of the present application;

[0064] Figure 2 It is a schematic diagram of the influence of various parameters in the operating data provided in the embodiment of the present application on the load;

[0065] Figure 3 This is the second flow chart of the cloud resource load prediction method based on data features provided in the embodiment of the present application;

[0066] Figure 4 It is a structural diagram of a cloud resource load prediction device based on data features provided in an embodiment of the present application;

[0067] Figure 5 It is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0068] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0069] The embodiment of the present application provides a cloud resource load prediction method based on data features. The method can be applied to smart terminals and servers. Some other descriptions in the embodiments of the present application are for illustrative purposes only and are not intended to limit the scope of protection of the present application. They will not be described one by one. The specific implementation of the method is as follows Figure 1 As shown:

[0070] Step 101, obtaining the operation data of the cloud platform within a preset time period when the cloud platform is running.

[0071] The operation data includes: CPU utilization, memory utilization, incoming network packets, outgoing network packets and disk space utilization.

[0072] Step 102, inputting the operating data into a pre-trained forest selection model to obtain key parameters output by the forest selection model.

[0073] Among them, the key parameter is any item in the operating data, and the forest selection model is trained based on the operating data samples and the key parameter samples.

[0074] Step 103, performing a decomposition operation on the key parameters based on a modal decomposition algorithm to obtain a plurality of subsequences.

[0075] Step 104, construct a two-dimensional subsequence, input the obtained two-dimensional subsequence into a pre-trained load prediction model, obtain hidden features corresponding to the two-dimensional subsequence through the attention mechanism embedded in the convolutional neural network; input the hidden features into the long short-term memory network to obtain the predicted value corresponding to the two-dimensional subsequence.

[0076] The load prediction model is trained based on two-dimensional subsequence samples and predicted value samples.

[0077] Step 105: sum each predicted value and determine the sum result as the cloud resource load prediction value.

[0078] The cloud resource load prediction method based on data features provided in the embodiment of the present application obtains the operation data within a preset time period when the cloud platform is running, wherein the operation data includes: CPU utilization, memory utilization, incoming network packets, outgoing network packets and disk space utilization; the operation data is input into a pre-trained forest selection model to obtain key parameters output by the forest selection model, wherein the key parameters are any item in the operation data. It can be seen that the present application uses the forest selection model to select important features, namely key parameters, from the operation data; then, the key parameters are decomposed based on the modal decomposition algorithm to obtain multiple subsequences; a two-dimensional subsequence is constructed, and the obtained two-dimensional subsequence is input into a pre-trained load prediction model, and the hidden features corresponding to the two-dimensional subsequence are obtained through the attention mechanism embedded in the convolutional neural network; the hidden features are input into a long short-term memory network to obtain the prediction value corresponding to the two-dimensional subsequence, and the present application uses the attention mechanism to effectively extract the hidden features of the data to reduce the interference of invalid features, thereby improving the accuracy of the prediction results; each prediction value is summed, and the summed result is determined as the cloud resource load prediction value, and finally the combination of multiple networks and the extraction of hidden features are realized to improve the accuracy of the prediction results.

[0079] In a specific embodiment, the specific implementation of obtaining the operation data within a preset time period when the cloud platform is running includes:

[0080] Acquire initial operation data of a preset number of servers in the cloud platform within a preset time period; perform preprocessing operations on the initial operation data to obtain operation data.

[0081] The preprocessing operation includes: deleting data including missing and incomplete information, and deleting data without preset fields.

[0082] Among them, the preset fields include: normalized memory bandwidth and cache misses per thousand instructions.

[0083] This application provides an effective data basis for improving the accuracy of model prediction by preprocessing the initial operation data.

[0084] In a specific embodiment, the operating data is input into a pre-trained forest selection model to obtain key parameters output by the forest selection model, including:

[0085] The running data is input into the forest selection model, and the forest selection model is used to determine the importance of each parameter, and the parameter corresponding to the largest importance score is determined as the key parameter.

[0086] Among them, key parameters include CPU utilization.

[0087] Specifically, determining the importance score of each parameter includes: performing an importance score on each parameter using an importance score formula.

[0088] The importance scoring formula includes: a first scoring formula, a second scoring formula and a third scoring formula.

[0089] Specifically, the mean square error before the forest node split is calculated by the first scoring formula, as shown in formula (1):

[0090]

[0091] Among them, MSE before represents the mean square error before splitting, N represents the number of running data in the g-th forest node, and y i represents the target value corresponding to the g-th forest node, Represents the sample mean corresponding to the g-th forest node.

[0092] The target value is obtained through historical data, including the CPU utilization in the historical data, and the sample mean refers to the average value of all historical data in each forest node, including the average value of CPU utilization in the historical data.

[0093] Specifically, the mean square error after the forest node split is calculated by the second scoring formula, which can be seen in formula (2):

[0094]

[0095] Among them, MSE after represents the mean square error after splitting, N L represents the number of running data in the left child node of the g-th forest node, N R represents the number of running data in the right child node of the g-th forest node, MSE L Represents the mean square error corresponding to the left child node, MSE R Represents the mean square error corresponding to the right child node.

[0096] Specifically, the output result of the first scoring formula and the output result of the second scoring formula are input into the third scoring formula to obtain the key parameters output by the third scoring formula, as shown in formula (3):

[0097]

[0098] Among them, f i represents the importance score, T i Represents all forest nodes that undergo splitting.

[0099] Specifically, the key parameter determined by the forest selection model of this application is CPU utilization. In addition, this application verifies the effectiveness of the key parameters selected by the forest selection model, such as Figure 2 As shown in Figure 2, the impact of key parameters in cloud resource load prediction is demonstrated. Figure 2 In the example, the correlation coefficient between the corresponding labels on the axis indicates the correlation between each parameter and the cloud resource load, which is represented by Figure 2 It can be seen that only the correlation coefficient of CPU utilization is greater than 0.8, which proves the effectiveness and reliability of the forest selection model.

[0100] In a specific embodiment, the key parameters are decomposed based on a modal decomposition algorithm to obtain a plurality of subsequences, which includes:

[0101] White noise is added to the CPU utilization to form superimposed data; the superimposed data is decomposed using the empirical mode decomposition method to obtain the first IMF and residual component, and the average value is used to eliminate the influence of white noise;

[0102] The residual component is subjected to the steps of adding white noise to CPU utilization to form superposition data; decomposing the superposition data by using the empirical mode decomposition method to obtain the first IMF and the residual component, and eliminating the influence of the white noise by using the average value, until the residual component cannot be decomposed.

[0103] This application obtains the key parameter CPU utilization through the forest selection model, but the CPU utilization still has strong mutation and volatility, resulting in a certain degree of noise. This application uses a modal decomposition algorithm to process nonlinear and high-frequency fluctuating signals and achieve effective signal noise reduction. Therefore, the modal decomposition algorithm is first used to preprocess the CPU utilization.

[0104] Specifically, white noise is added to the CPU utilization to form superimposed data, as shown in formula (4):

[0105]

[0106] in, represents the first IMF component obtained after adding white noise for the ith time, cpu(t) represents the CPU utilization, ε 0 represents the signal-to-noise ratio, t represents the time point, i represents white noise, ω i (t) represents the white noise sequence obtained after adding white noise for the ith time, and n represents the total number of times white noise is added.

[0107] The superimposed data includes multiple IMF components.

[0108] Specifically, the empirical mode decomposition (EMD) method is used to decompose the superimposed data to obtain the first IMF and residual component, and the average value is used to eliminate the influence of white noise, as shown in formula (5):

[0109]

[0110] in, It represents the average value of the first intrinsic mode function (IMF) obtained by averaging the added results after adding white noise multiple times. It means that the IMF components obtained after adding white noise each time are averaged to obtain a smooth result, EMD 1 (f i (t)) represents the process of using empirical mode decomposition after adding white noise for the i-th time, which is used to extract the first IMF.

[0111] Specifically, adaptive white noise is added to the residual component to form new superposition data, and then EMD decomposition is performed to obtain IMF and residual components, and so on. Finally, the operation is stopped when the residual component cannot be decomposed. For details, please refer to formula (6)-formula (9):

[0112]

[0113] in, represents the average value of the kth intrinsic mode function (IMF) obtained by averaging the added results after adding white noise multiple times, K represents the total number of IMFs, r 1 (t) represents the first residual component, r k-1 (t) k-1th residual component, r k (t) represents the kth residual component, R(t) represents the final residual component, ε k-1 represents adaptive white noise.

[0114] In a specific embodiment, the specific implementation of obtaining the hidden features corresponding to the two-dimensional subsequence through the attention mechanism embedded in the convolutional neural network includes:

[0115] A first feature map is obtained through the first convolution layer of the convolutional neural network, and a second feature map is obtained through the second convolution layer of the convolutional neural network; an average feature corresponding to a global average pooling channel and a maximum feature corresponding to a maximum pooling channel are generated based on the first feature map and the second feature map; based on the attention mechanism, hidden features are obtained using the average features and the maximum features.

[0116] Specifically, a attention network (CBAM) is embedded in a two-layer convolutional neural network (CNN), and the channel attention mechanism (CAM) in CBAM is partially embedded in CNN to realize the extraction of hidden features.

[0117] Specifically, multiple subsequences are converted into a two-dimensional structure to obtain a two-dimensional subsequence.

[0118] The two-dimensional subsequence is in the form of C×H×W.

[0119] Among them, C represents the number of channels, H represents the height, and W represents the width.

[0120] The two-dimensional subsequence is input into the load prediction model and a double-layer convolution operation is performed. For details, see formula (10) and formula (11):

[0121] F 1 =Conv 1 (IMF(t))…………………………………………(10)

[0122] F 2 =Conv 2 (F 1 )…………………………………………………………(11)

[0123] Specifically, the first characteristic map is obtained by formula (10), and the second characteristic map is obtained by formula (11).

[0124] Among them, F 1 represents the first feature map, IMF(t) represents a two-dimensional subsequence, Conv 1 (IMF(t)) represents the convolution operation on the two-dimensional subsequence through the first convolution layer, F 2 Represents the second feature map, Conv 2 (F 1 ) indicates that the first feature map is convolved through the second convolutional layer.

[0125] Specifically, the channel is generated by formula (12) and formula (13):

[0126]

[0127] Among them, F avg represents the average feature, F max represents the maximum feature, F 2 [:,i,j] represents the feature values ​​of all channels at the (i, j) position of the second feature map.

[0128] Specifically, the specific implementation of further processing hidden features based on the channel attention mechanism is as follows:

[0129] The weights are generated by formula (14) and formula (15):

[0130] F c =[F avg,F max ]………………………………………………………………(14)

[0131] S c =σ(W 2 δ(W 1 F c ))……………………………………………………(15)

[0132] Among them, F c Indicates F avg and F max The union of c represents the channel attention weight, which is used to characterize the importance of the channel, σ represents the sigmoid function, and W 2 represents the weight matrix of the second fully connected layer in the convolutional neural network, δ represents the activation function, W 1 Represents the weight matrix of the first fully connected layer in a convolutional neural network.

[0133] Among them, the activation function δ is used to introduce nonlinearity so that the model can better capture the complex relationship between features. 1 To make the channel feature F c Mapped to an intermediate representation.

[0134] Specifically, channel attention and spatial attention are combined in a weighted manner to form new features so that important features are further enhanced, see formula (16):

[0135] F cout =S c ⊙F 2 …… ...

[0136] Among them, F cout Indicates new features.

[0137] Specifically, the CAM embedded in CNN is used to output the final hidden features, see formula (17):

[0138] F out =Conv next (F cout )…… ...

[0139] Among them, F out represents the final hidden features.

[0140] In a specific embodiment, the specific implementation of inputting the hidden features into the long short-term memory network to obtain the prediction value corresponding to the two-dimensional subsequence includes:

[0141] The final hidden features are converted into one-dimensional features, and the one-dimensional features are used as the input of the long short-term memory network. Specifically, the predicted value is obtained using formula (18)-formula (24):

[0142] f t =σ(W (f) IMFP (t) +U (f) h (t-1) )……………………………………(18)

[0143] i t =σ(W (i) h (t-1) +W (i)IMFP IMFP (t) +b i )………………………………(19)

[0144] o t =o(W (o) IMFP (t) +U (o) h (t-1) )……………………………………(20)

[0145]

[0146] h t =o t tanh(C t )………………………………………………(twenty two)

[0147]

[0148] IMFY k (t) = W p +h t +b p …………………………………………(twenty four)

[0149] Among them, IMFY k (t) represents the predicted value of the kth MF component at time t, h (t-1) Indicates the hidden state of the previous moment, IMFP (t) represents the one-dimensional feature corresponding to the hidden feature, W ih is the weight in the input gate weight matrix, used to connect IMFP (t) and h (t-1) The interaction, i t Indicates the state of the input gate, o t Indicates the state of the output gate, W (f) represents the weight matrix of the forget gate, U (f)represents the recursive weight matrix of the forget gate, W ( i ) represents the weight matrix of the input gate, W (o) represents the weight matrix of the output gate, U (o) represents the recursive weight matrix of the output gate, W (c) The weight matrix representing the candidate cell state, U (c) Represents the recursive weight matrix of candidate cell states, b i represents the bias term of the input gate, i t represents the activation value of the input gate, W p represents the weight matrix, b p represents the bias term, C t represents the cell state at time t, h t represents the hidden state at the current moment, W (i)IMFP represents another weight matrix of the input gate, represents the candidate cell state, C (t-1) Indicates the cell state at the previous moment.

[0150] Among them, W p and b p Corresponding to the network parameters, its function is to transform the hidden state b of the LSTM unit p Convert to predicted value.

[0151] This application discloses a LSTM cloud resource load prediction method based on random forest selection algorithm and attention mechanism, involving feature extraction technology and load prediction technology. For specific implementation, please refer to Figure 3 :

[0152] Step 301, obtaining operation data and preprocessing the data.

[0153] Step 302: Use the random forest algorithm to filter out the important feature CPU utilization.

[0154] Step 303: Decompose the CPU utilization data set into relatively stable CPU sub-sequences using a fully adaptive noise set empirical mode decomposition algorithm.

[0155] Step 304: Use a long short-term memory network with a lightweight attention mechanism to predict each CPU subsequence.

[0156] Step 305: superimpose the predicted values ​​to obtain a cloud resource load prediction result.

[0157] This application can solve the current problems of large fluctuations in cloud resource load and irregular changes, which lead to difficulty in effectively extracting features and low load prediction accuracy. It can more accurately predict cloud resource load trends and is conducive to the rational allocation of resources.

[0158] This application takes into account the nonlinearity and non-stationarity of cloud resource load data, and uses CEEMDAN to decompose the denoised data into different IMF components and residual components. CEEMDAN effectively reduces the impact of noise and improves the stability of the decomposition results by adding white noise and taking the average value. The decomposed data is predicted using multiple LSTM neural networks, and multiple prediction results are synthesized to improve the accuracy of the prediction results. At the same time, the attention mechanism is introduced. The attention mechanism improves the accuracy of hidden feature extraction through the feature extraction capabilities of the two dimensions of channel attention mechanism and spatial attention mechanism.

[0159] The embodiment of the present application also provides a cloud resource load prediction device based on data features. The specific implementation of the device can refer to the description of the cloud resource load prediction method based on data features, and the repeated parts will not be repeated. Figure 4 As shown, the device comprises:

[0160] The acquisition module 401 is used to acquire the operation data within a preset time period when the cloud platform is running, wherein the operation data includes: CPU utilization, memory utilization, incoming network packets, outgoing network packets and disk space utilization;

[0161] A key feature extraction module 402 is used to input the operation data into a pre-trained forest selection model to obtain key parameters output by the forest selection model, wherein the key parameter is any item in the operation data, and the forest selection model is trained based on the operation data sample and the key parameter sample;

[0162] A decomposition module 403 is used to perform a decomposition operation on the key parameters based on a modal decomposition algorithm to obtain multiple subsequences;

[0163] The prediction module 404 is used to construct a two-dimensional subsequence, input the obtained two-dimensional subsequence into a pre-trained load prediction model, obtain hidden features corresponding to the two-dimensional subsequence through the attention mechanism embedded in the convolutional neural network; input the hidden features into the long short-term memory network to obtain the predicted value corresponding to the two-dimensional subsequence, wherein the load prediction model is trained based on the two-dimensional subsequence samples and the predicted value samples;

[0164] The summing module 405 is used to sum each predicted value and determine the summed result as the cloud resource load prediction value.

[0165] In a specific embodiment, the prediction module 404 is used to obtain a first feature map through a first convolutional layer of a convolutional neural network, and to obtain a second feature map through a second convolutional layer of a convolutional neural network; generate an average feature corresponding to a global average pooling channel and a maximum feature corresponding to a maximum pooling channel based on the first feature map and the second feature map; and obtain hidden features using the average feature and the maximum feature based on an attention mechanism.

[0166] In a specific embodiment, the prediction module 404 is used to obtain hidden features through a feature extraction formula;

[0167] Among them, the feature extraction formula includes:

[0168] F 1 =Conv 1 (IMF(t));

[0169] F 2 =Conv 2 (F 1 );

[0170]

[0171] F c =[F avg ,F max ];

[0172] S c =σ(W 2 δ(W 1 F c ));

[0173] F cout =S c ⊙F 2 ;

[0174] F out =Conv next (F cout );

[0175] Among them, F 1 represents the first feature map, IMF(t) represents a two-dimensional subsequence, Conv 1 (IMF(t)) represents the convolution operation on the two-dimensional subsequence through the first convolution layer, F 2 Represents the second feature map, Conv 2 (F 1 ) indicates that the first feature map is convolved by the second convolutional layer, F avg represents the average feature, F max represents the maximum feature, F 2[:,i,j] represents the feature values ​​of all channels at the (i, j) position of the second feature map, F c Indicates F avg and F max The union of c represents the channel attention weight, which is used to characterize the importance of the channel, σ represents the sigmoid function, and W 2 represents the weight matrix of the second fully connected layer in the convolutional neural network, δ represents the activation function, W 1 represents the weight matrix of the first fully connected layer in the convolutional neural network, F cou t represents the new feature, F out represents the final hidden features.

[0176] In a specific embodiment, the prediction module 404 is used to obtain a prediction value through a prediction formula;

[0177] The prediction formula includes:

[0178] f t =σ(W (f) IMFP (t) +U (f) h (t-1) );

[0179] i t =σ(W ih h t-1 +W iIMFP IMFP (t) +b i );

[0180] o t =o(W (o) IMFP (t) +U (o) h (t-1) );

[0181]

[0182] h t =o t ·tanh(c t );

[0183]

[0184] IMFY k (t) = W p +h t +b p ;

[0185] Among them, IMFY k(t) represents the predicted value of the kth MF component at time t, h (t-1) Indicates the hidden state of the previous moment, IMFP (t) represents the one-dimensional feature corresponding to the hidden feature, W ih is the weight in the input gate weight matrix, used to connect IMFP (t) and h (t-1) The interaction, i t Indicates the state of the input gate, o t Indicates the state of the output gate, W (f) represents the weight matrix of the forget gate, U (f) represents the recursive weight matrix of the forget gate, W ( i ) represents the weight matrix of the input gate, W (o) represents the weight matrix of the output gate, U (o) represents the recursive weight matrix of the output gate, W (c) The weight matrix representing the candidate cell state, U (c) Represents the recursive weight matrix of candidate cell states, b i represents the bias term of the input gate, i t represents the activation value of the input gate, W p represents the weight matrix, b p represents the bias term, C t represents the cell state at time t, h t represents the hidden state at the current moment, W (i)IMFP represents another weight matrix of the input gate, represents the candidate cell state, C (t-1) Indicates the cell state at the previous moment.

[0186] In a specific embodiment, the key feature extraction module 402 is used to input the operation data into the forest selection model, determine the importance score of each parameter through the forest selection model, and determine the parameter corresponding to the maximum importance score as the key parameter;

[0187] Among them, the importance score of each parameter is determined by the scoring formula:

[0188]

[0189] Among them, MSE before represents the mean square error before splitting, N represents the number of running data in the g-th forest node, and y i represents the target value corresponding to the g-th forest node, Represents the sample mean corresponding to the g-th forest node, MSE after represents the mean square error after splitting, N Lrepresents the number of running data in the left child node of the g-th forest node, N R represents the number of running data in the right child node of the g-th forest node, MSE L Represents the mean square error corresponding to the left child node, MSE R represents the mean square error corresponding to the right child node, f i represents the importance score, T i Represents all forest nodes that undergo splitting.

[0190] In a specific embodiment, the key parameters include: CPU utilization; a decomposition module 403, which is used to add white noise to the CPU utilization to form superimposed data; decompose the superimposed data using an empirical mode decomposition method to obtain a first IMF and a residual component, and use an average value to eliminate the influence of white noise; perform the steps of adding white noise to the CPU utilization to form superimposed data on the residual component; decompose the superimposed data using an empirical mode decomposition method to obtain an IMF and a residual component, and use an average value to eliminate the influence of white noise, until the residual component cannot be decomposed.

[0191] In a specific embodiment, the decomposition module 403 is used to obtain multiple subsequences based on the decomposition formula;

[0192] The decomposition formula includes:

[0193] IMF 1 i (t)=cpu(t)+ε 0 ω i (t),i=1,2,3,...,n;

[0194]

[0195] in, represents the first IMF component obtained after adding white noise for the ith time, cpu(t) represents the CPU utilization, ε 0 represents the signal-to-noise ratio, t represents the time point, i represents white noise, ω i (t) represents the white noise sequence obtained after adding white noise for the ith time, n represents the total number of times white noise is added, It represents the average value of the first intrinsic mode function (IMF) obtained by averaging the added results after adding white noise multiple times. It means that the IMF components obtained after adding white noise each time are averaged to obtain a smooth result, EMD 1 (f i (t)) represents the process of using empirical mode decomposition after adding white noise for the i-th time, which is used to extract the first IMF. represents the average value of the kth intrinsic mode function (IMF) obtained by averaging the added results after adding white noise multiple times, K represents the total number of IMFs, r 1 (t) represents the first residual component, (r k-1 (t) k-1th residual component, r k (t) represents the kth residual component, R(t) represents the final residual component, ε k-1 represents adaptive white noise.

[0196] Figure 5 An example of a physical structure diagram of an electronic device is shown in FIG. Figure 5 As shown, the electronic device may include: a processor (processor) 501, a communication interface (Communications Interface) 502, a memory (memory) 503 and a communication bus 504, wherein the processor 501, the communication interface 502, and the memory 503 communicate with each other through the communication bus 504. The processor 501 can call the logic instructions in the memory 503 to execute a cloud resource load prediction method based on data features, which includes: obtaining operation data within a preset time period when the cloud platform is running, wherein the operation data includes: CPU utilization, memory utilization, incoming network packets, outgoing network packets and disk space utilization; inputting the operation data into a pre-trained forest selection model to obtain key parameters output by the forest selection model, wherein the key parameters are any item in the operation data, and the forest selection model is trained based on operation data samples and key parameter samples; decomposing the key parameters based on a modal decomposition algorithm to obtain multiple subsequences; constructing a two-dimensional subsequence, inputting the obtained two-dimensional subsequence into a pre-trained load prediction model, and obtaining hidden features corresponding to the two-dimensional subsequence through the attention mechanism embedded in the convolutional neural network; inputting the hidden features into a long short-term memory network to obtain a predicted value corresponding to the two-dimensional subsequence, wherein the load prediction model is trained based on two-dimensional subsequence samples and predicted value samples; summing each predicted value, and determining the obtained summation result as the cloud resource load prediction value.

[0197] In addition, the logic instructions in the above-mentioned memory 503 can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0198] On the other hand, the present invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the cloud resource load prediction method based on data features provided by the above methods, and the method includes: obtaining operation data within a preset time period when the cloud platform is running, wherein the operation data includes: CPU utilization, memory utilization, incoming network packets, outgoing network packets and disk space utilization; inputting the operation data into a pre-trained forest selection model to obtain key parameters output by the forest selection model, wherein , the key parameter is any item in the operating data, and the forest selection model is trained based on the operating data samples and the key parameter samples; the key parameters are decomposed based on the modal decomposition algorithm to obtain multiple subsequences; a two-dimensional subsequence is constructed, and the obtained two-dimensional subsequence is input into the pre-trained load prediction model, and the hidden features corresponding to the two-dimensional subsequence are obtained through the attention mechanism embedded in the convolutional neural network; the hidden features are input into the long short-term memory network to obtain the predicted value corresponding to the two-dimensional subsequence, wherein the load prediction model is trained based on the two-dimensional subsequence samples and the predicted value samples; each predicted value is summed, and the summed result is determined as the cloud resource load prediction value.

[0199] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which is implemented when executed by a processor to execute the cloud resource load prediction method based on data features provided in the above-mentioned embodiments, the method comprising: obtaining operation data within a preset time period when the cloud platform is running, wherein the operation data comprises: CPU utilization, memory utilization, incoming network packets, outgoing network packets and disk space utilization; inputting the operation data into a pre-trained forest selection model to obtain key parameters output by the forest selection model, wherein the key parameters are any item in the operation data, and the forest selection model is trained based on operation data samples and key parameter samples; performing a decomposition operation on the key parameters based on a modal decomposition algorithm to obtain multiple subsequences; constructing a two-dimensional subsequence, inputting the obtained two-dimensional subsequence into a pre-trained load prediction model, and obtaining hidden features corresponding to the two-dimensional subsequence through the attention mechanism embedded in the convolutional neural network; inputting the hidden features into a long short-term memory network to obtain a predicted value corresponding to the two-dimensional subsequence, wherein the load prediction model is trained based on two-dimensional subsequence samples and predicted value samples; summing each predicted value, and determining the obtained summed result as the cloud resource load prediction value.

[0200] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0201] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0202] Finally, it should be noted that the above is only the preferred implementation of the present application, and the present application is not limited to the above embodiments. It is understood that other improvements and changes directly derived or associated by those skilled in the art without departing from the spirit and concept of the present application should be considered to be included in the scope of protection of the present application.

Claims

1. A cloud resource load prediction method based on data features, characterized in that: The method comprises: Obtaining operation data within a preset time period when the cloud platform is running, wherein the operation data includes: CPU utilization, memory utilization, incoming network packets, outgoing network packets and disk space utilization; Inputting the operating data into a pre-trained forest selection model to obtain key parameters output by the forest selection model, wherein the key parameter is any item in the operating data, and the forest selection model is trained based on operating data samples and key parameter samples; Decomposing the key parameters based on a modal decomposition algorithm to obtain a plurality of subsequences; Constructing a two-dimensional subsequence, inputting the obtained two-dimensional subsequence into a pre-trained load prediction model, and obtaining hidden features corresponding to the two-dimensional subsequence through an attention mechanism embedded in a convolutional neural network; inputting the hidden features into a long short-term memory network to obtain a predicted value corresponding to the two-dimensional subsequence, wherein the load prediction model is trained based on two-dimensional subsequence samples and predicted value samples; Each of the predicted values ​​is summed up, and the summed result is determined as the cloud resource load prediction value.

2. The cloud resource load prediction method based on data features according to claim 1 is characterized in that: The hidden features corresponding to the two-dimensional subsequence are obtained through the attention mechanism embedded in the convolutional neural network, including: Obtaining a first feature map through a first convolutional layer of a convolutional neural network, and obtaining a second feature map through a second convolutional layer of the convolutional neural network; Generate an average feature corresponding to a global average pooling channel and a maximum feature corresponding to a maximum pooling channel based on the first feature map and the second feature map; Based on the attention mechanism, the hidden feature is obtained by using the average feature and the maximum feature.

3. The cloud resource load prediction method based on data features according to claim 2 is characterized in that: The hidden features corresponding to the two-dimensional subsequence are obtained by the attention mechanism embedded in the convolutional neural network, including: Obtaining the hidden features through a feature extraction formula; Wherein, the feature extraction formula includes: F1 = Conv1(IMF(t)); F2 = Conv2(F1); F c =[F avg ,F max ]; S c =σ(W2δ(W1F c )); F cout =S c ⊙F2; F out =Conv next (F cout ); Where F1 represents the first feature map, IMF(t) represents the two-dimensional subsequence, Conv1(IMF(t)) represents the convolution operation on the two-dimensional subsequence through the first convolution layer, F2 represents the second feature map, Conv2(F1) represents the convolution operation on the first feature map through the second convolution layer, and F avg represents the average feature, F max represents the maximum feature, F2[:,i,j] represents the feature values ​​of all channels at the (i, j) position of the second feature map, and F c Indicates F avg and F max The union of c represents the channel attention weight, which is used to characterize the importance of the channel, σ represents the sigmoid function, W2 represents the weight matrix of the second fully connected layer in the convolutional neural network, δ represents the activation function, W1 represents the weight matrix of the first fully connected layer in the convolutional neural network, and F cou t represents the new feature, F out represents the final hidden features.

4. The cloud resource load prediction method based on data features according to any one of claims 1 to 3, characterized in that: Inputting the hidden features into a long short-term memory network to obtain a prediction value corresponding to the two-dimensional subsequence includes: Obtaining the predicted value through a prediction formula; Wherein, the prediction formula includes: f t =σ(W (f) IMFP (t) +U (f) h (t-1) ); i t =σ(W ih h t-1 +W iIMFP IMFP (t) +b i ); o t =o(W (o) IMFP (t) +U (o) h (t-1) ); h t =o t ·tanh(c t ); IMFY k (t)=W p +h t +b p ; Among them, IMFY k (t) represents the predicted value of the kth MF component at time t, h (t-1) Indicates the hidden state of the previous moment, IMFP (t) represents the one-dimensional feature corresponding to the hidden feature, W ih is the weight in the input gate weight matrix, used to connect IMFP (t) and h (t-1) The interaction, i t Indicates the state of the input gate, o t Indicates the state of the output gate, W (f) represents the weight matrix of the forget gate, U (f) represents the recursive weight matrix of the forget gate, W ( i ) represents the weight matrix of the input gate, W (o) represents the weight matrix of the output gate, U (o) represents the recursive weight matrix of the output gate, W (c) The weight matrix representing the candidate cell state, U (c) Represents the recursive weight matrix of candidate cell states, b i represents the bias term of the input gate, i t represents the activation value of the input gate, W p represents the weight matrix, b p represents the bias term, C t represents the cell state at time t, h t represents the hidden state at the current moment, W (i)IMFP represents another weight matrix of the input gate, represents the candidate cell state, C (t-1) Indicates the cell state at the previous moment.

5. The cloud resource load prediction method based on data features according to any one of claims 1 to 3, characterized in that: The operating data is input into a pre-trained forest selection model to obtain key parameters output by the forest selection model, including: Inputting the operating data into the forest selection model, determining the importance score of each parameter through the forest selection model, and determining the parameter corresponding to the maximum importance score as the key parameter; Among them, the importance score of each parameter is determined by the scoring formula: Among them, MSE before represents the mean square error before splitting, N represents the number of running data in the g-th forest node, and y i represents the target value corresponding to the g-th forest node, Represents the sample mean corresponding to the g-th forest node, MSE after represents the mean square error after splitting, N L represents the number of running data in the left child node of the g-th forest node, N R represents the number of running data in the right child node of the g-th forest node, MSE L Represents the mean square error corresponding to the left child node, MSE R represents the mean square error corresponding to the right child node, f i represents the importance score, T i Represents all forest nodes that undergo splitting.

6. The cloud resource load prediction method based on data features according to any one of claims 1 to 3, characterized in that: The key parameters include: the CPU utilization; The key parameters are decomposed based on a modal decomposition algorithm to obtain multiple subsequences, including: Adding white noise to the CPU utilization to form superimposed data; decomposing the superimposed data using an empirical mode decomposition method to obtain a first IMF and a residual component, and using an average value to eliminate the influence of the white noise; The residual component is added with white noise for the CPU utilization to form the superimposed data; the superimposed data is decomposed by using the empirical mode decomposition method to obtain IMF and residual components, and the influence of the white noise is eliminated by using the average value, until the residual component cannot be decomposed.

7. The cloud resource load prediction method based on data features according to claim 6 is characterized in that: The key parameters are decomposed based on a modal decomposition algorithm to obtain multiple subsequences, including: Based on the decomposition formula, multiple subsequences are obtained; Wherein, the decomposition formula includes: in, represents the first IMF component obtained after adding white noise for the i-th time, cpu(t) represents CPU utilization, ε0 represents signal-to-noise ratio, t represents time point, i represents white noise, ω i (t) represents the white noise sequence obtained after adding white noise for the ith time, n represents the total number of times white noise is added, It represents the average value of the first intrinsic mode function (IMF) obtained by averaging the added results after adding white noise multiple times. It means that after adding white noise each time, the IMF components obtained are averaged to obtain a smooth result, EMD1(f i (t)) represents the process of using empirical mode decomposition after adding white noise for the i-th time, which is used to extract the first IMF. represents the average value of the kth intrinsic mode function (IMF) obtained by averaging the added results after multiple additions of white noise, K represents the total number of IMFs, r1(t) represents the first residual component, r k-1 (t) k-1th residual component, r k (t) represents the kth residual component, R(t) represents the final residual component, ε k-1 represents adaptive white noise.

8. A cloud resource load prediction device based on data features, characterized in that: The device comprises: An acquisition module is used to acquire operation data within a preset time period when the cloud platform is running, wherein the operation data includes: CPU utilization, memory utilization, incoming network packets, outgoing network packets and disk space utilization; A key feature extraction module, used for inputting the operation data into a pre-trained forest selection model to obtain key parameters output by the forest selection model, wherein the key parameter is any item in the operation data, and the forest selection model is trained based on the operation data sample and the key parameter sample; A decomposition module, used for performing a decomposition operation on the key parameters based on a modal decomposition algorithm to obtain a plurality of subsequences; A prediction module is used to construct a two-dimensional subsequence, input the obtained two-dimensional subsequence into a pre-trained load prediction model, obtain hidden features corresponding to the two-dimensional subsequence through an attention mechanism embedded in a convolutional neural network; input the hidden features into a long short-term memory network to obtain a predicted value corresponding to the two-dimensional subsequence, wherein the load prediction model is trained based on two-dimensional subsequence samples and predicted value samples; The summing module is used to sum each of the predicted values ​​and determine the summed result as the cloud resource load prediction value.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the cloud resource load prediction method based on data features as described in any one of claims 1 to 7 are implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the cloud resource load prediction method based on data features as described in any one of claims 1 to 7 are implemented.

Images