Cloud system load prediction method based on multi-model optimization

Abstract

The application provides a cloud system load prediction method and system based on multi-model optimization, a storage medium and an electronic device, and relates to the technical field of system load prediction.The application overcomes the defects that the model prediction process is easy to fall into a local optimal solution, slow and unstable convergence speed by using wavelet transform to carry out data mining on cloud system load time series data and using sparrow search algorithm to optimize neural network hyperparameters.The prediction method can extract the characteristic changes of the cloud system load data, and finally realizes high-accuracy prediction and analysis of the cloud system load time series data.

Description

Technical Field

[0001] This invention relates to the field of system load prediction technology, and specifically to a cloud system load prediction method, system, storage medium, and electronic device based on multi-model optimization. Background Technology

[0002] In recent years, cloud computing, as a new computing model, has developed rapidly and been applied to various fields, attracting increasing attention. Large enterprises have built their own cloud computing data systems to provide users with more secure and efficient computing, storage, network resources, and application services. The increasing number of users deploying applications on cloud systems often leads to a shortage of cloud resources.

[0003] With the development of cloud computing and cloud-native technologies, how to intelligently predict the changes in resource load in future periods based on the resource and application characteristics of cloud system deployment, and rationally schedule resources to improve the resource utilization of cloud systems and reduce response latency, is of significant research and economic value.

[0004] In related technologies, many scholars have conducted extensive research on cloud system load prediction:

[0005] For example, the paper (Cloud computing resource load prediction based on improved Informer[1] Li Haoyang, He Xiaowei, Wang Bin, et al. Cloud computing resource load prediction based on improved Informer[J]. Computer Engineering, 2024, 50(02):43-50.DOI:10.19678 / j.issn.1000-3428.0066399.) proposes a multi-step load prediction model based on improved Informer, Informer-DCR.

[0006] For example, the paper (Cloud Platform Load Prediction Method Based on Temporal Convolutional Network [1] Li Yinghao, Guo Haogong, Liu Panpan, et al. Cloud Platform Load Prediction Method Based on Temporal Convolutional Network [J]. Computer Science, 2023, 50(07): 254-260.) proposes a cloud platform resource load prediction method.

[0007] For example, the paper (BiGRU-LGB cloud load prediction model integrating Stacking framework [1] Liu Hui, Dong Xiyao, Yang Zhihan. BiGRU-LGB cloud load prediction model integrating Stacking framework [J]. Journal of Xi'an University of Electronic Science and Technology, 2023, 50(03): 83-94+104. DOI: 10.19665 / j.issn1001-2400.2023.03.008.) proposes a cloud load prediction model integrating Stacking framework, multi-layer BiGRU network and LightGBM algorithm.

[0008] However, the aforementioned prediction methods do not extract features from the original data, and rely solely on experience to set model parameters, making it difficult to obtain a high-accuracy and stable prediction model. Summary of the Invention

[0009] (a) Technical problems to be solved

[0010] To address the shortcomings of existing technologies, this invention provides a cloud system load prediction method, system, storage medium, and electronic device based on multi-model optimization, which solves the technical problem of low prediction accuracy of traditional prediction methods for cloud system load data with large fluctuations.

[0011] (II) Technical Solution

[0012] To achieve the above objectives, the present invention provides the following technical solution:

[0013] A cloud system load prediction method based on multi-model optimization includes:

[0014] Collect raw cloud system load time series data;

[0015] The cloud system load time series data is preprocessed, including obtaining several subsequences using wavelet transform decomposition and normalizing the data of each subsequence.

[0016] Based on each normalized subsequence, a corresponding initial CNN-DLSTM subsequence prediction model is established to obtain the prediction results of each normalized subsequence.

[0017] Fitness functions are established for each normalized subsequence and its prediction results, and the hyperparameters of each initial CNN-DLSTM subsequence prediction model are optimized using the sparrow search algorithm to obtain the optimal parameter combination.

[0018] Based on each normalized subsequence and the optimal parameter combination, a CNN-DLSTM subsequence prediction model is established to obtain the prediction results of each subsequence before inverse normalization.

[0019] Inverse normalization is performed to obtain the actual prediction results of each subsequence, and wavelet reconstruction is used to accumulate each actual prediction result to obtain the final cloud system load prediction result.

[0020] Preferably, the step of obtaining several sub-sequences using wavelet transform decomposition is represented as follows:

[0021]

[0022] Let the original cloud system load time series data be X={x1,x2,x3,...,x i ,...,x n}, where n represents the total amount of load data, x i This represents the load value of the i-th cloud system;

[0023] h(i) is a low-pass filter, corresponding to the wavelet scaling function;

[0024] g(i) is a high-pass filter, corresponding to the mother wavelet function of the wavelet;

[0025] d i Indicates by x i Decomposed load values;

[0026] If the original cloud system load time series data is decomposed m times, then X = {x1, x2, x3, ..., x...} i ,...,x n Decompose it into m subsequences D = {D1, D2, ..., D} t ,...,D m The subsequence contains one approximate component D and m-1 detail components D2,...,D t ,...,D m Each component consists of n load values ​​D t ={d1,d2,...,d n}

[0027] Preferably, the data normalization of each subsequence includes:

[0028] Map each decomposed subsequence to the interval [0,1];

[0029] Calculate the maximum and minimum values ​​of each subsequence, and denote them as d. max and d min ;

[0030] Take each subsequence D from each subsequence t Subtract d min Then divide by d max -d min Obtain D1', D'2, ..., D respectively. t ',,D' m , , respectively represent the normalized results of each subsequence.

[0031] Preferably, the step of establishing corresponding initial CNN-DLSTM subsequence prediction models based on each normalized subsequence to obtain prediction results for each normalized subsequence includes:

[0032] The normalized subsequence D is processed using a sliding window method. t 'Format the matrix L as the input matrix of the CNN, where each element of L is an a-dimensional feature vector, and a is the size of the time window;

[0033] A CNN is used to extract local features from the input matrix L. A convolution kernel matrix K is defined, consisting of k a-dimensional vectors, where k is the size of the convolution window, to perform a one-dimensional convolution operation on L, as shown in the following equation:

[0034] o q =ReLU(K·L) r:r+k-1 +b) (2)

[0035] Among them, L r:r+k-1 Let L represent the submatrix from position r to r+k-1, b be the bias term, ReLU be the mapping function, and o be the offset term. q Let O be the q-th local feature of the feature map O;

[0036] Max pooling is used to reduce the dimensionality of the feature map, as shown in the following equation:

[0037] o q `=max(o q:q+p (3)

[0038] Where p is the pooling window size, o q ` represents the q-th local feature of the pooled feature map O`;

[0039] The pooled feature map O' is processed using DLSTM, and the output hidden state matrix is ​​H. DLSTM ;

[0040] The output of the DLSTM is mapped to the target dimension through a fully connected layer and an activation function, as follows:

[0041] Y t = softmax(W·H) DLSTM +b) (4)

[0042] Among them, Y t ` is the normalized subsequence D` t The prediction results are given by W, which is the fully connected weight matrix, and softmax(·) is the probability distribution function.

[0043] Preferably, the mean squared error function (MSE) is established as the fitness function based on each normalized subsequence and its prediction results.

[0044] Preferably, the step of using the Sparrow Search algorithm to optimize the hyperparameters of each initial CNN-DLSTM subsequence prediction model to obtain the optimal parameter combination includes:

[0045] The hyperparameters to be optimized are time window, number of hidden layers, and number of hidden layer nodes; and the parameters are initialized, including updating the population size, maximum number of iterations, search radius, range of hyperparameters, and learning rate in sequence.

[0046] Population initialization is performed by generating parameter values ​​for initial individuals using random numbers and perturbing them based on a defined search radius to generate population individuals. CNN-DLSTM subsequence prediction models are then built and trained using the parameter values ​​generated for each individual, and the prediction error of the model is calculated as the individual fitness value.

[0047] After the algorithm iteration begins, the population is continuously updated, including the additions, discoverers, watchers, and the optimal sparrow position. The update strategy is improved by introducing dynamic weights to balance the global search capability.

[0048] After the algorithm iteration is completed, the sparrow with the best fitness value is obtained, and its corresponding hyperparameters are decoded as the optimal parameter combination for the CNN-DLSTM subsequence prediction model.

[0049] Preferably, the step of performing inverse normalization to obtain the actual prediction results of each subsequence refers to:

[0050] Based on the maximum value d of each subsequence after decomposition max and minimum value d min Multiply the unnormalized prediction results of each subsequence by d. max -d min Then, add d min .

[0051] A cloud system load prediction system based on multi-model optimization includes:

[0052] The acquisition module is used to collect raw cloud system load time series data;

[0053] The preprocessing module is used to preprocess the cloud system load time series data, including obtaining several subsequences by wavelet transform decomposition and normalizing the data of each subsequence.

[0054] A module is established to build corresponding initial CNN-DLSTM subsequence prediction models based on each normalized subsequence, so as to obtain the prediction results of each normalized subsequence.

[0055] The optimization module is used to establish fitness functions for each normalized subsequence and its prediction results, and to optimize the hyperparameters of each initial CNN-DLSTM subsequence prediction model using the sparrow search algorithm to obtain the best parameter combination.

[0056] The prediction module is used to build a CNN-DLSTM subsequence prediction model based on each normalized subsequence and the optimal parameter combination, so as to obtain the prediction results of each subsequence before inverse normalization.

[0057] The acquisition module is used to perform inverse normalization to obtain the actual prediction results of each subsequence, and to accumulate each actual prediction result using wavelet reconstruction to obtain the final cloud system load prediction result.

[0058] A storage medium storing a computer program for cloud system load prediction based on multi-model optimization, wherein the computer program causes a computer to perform the cloud system load prediction method as described above.

[0059] An electronic device, comprising:

[0060] One or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing cloud system load prediction as described above.

[0061] (III) Beneficial Effects

[0062] This invention provides a cloud system load prediction method, system, storage medium, and electronic device based on multi-model optimization. Compared with existing technologies, it has the following advantages:

[0063] This invention overcomes the shortcomings of traditional models, such as being prone to getting trapped in local optima, slow convergence, and instability, by utilizing wavelet transform for data mining of cloud system load time series data and employing the Sparrow Search algorithm to optimize neural network hyperparameters. This prediction method can extract the characteristic changes in cloud system load data, ultimately achieving high-accuracy prediction and analysis of cloud system load time series data. Attached Figure Description

[0064] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0065] Figure 1 A block diagram illustrating a cloud system load prediction method based on multi-model optimization, provided in an embodiment of the present invention.

[0066] Figure 2 A wavelet decomposition process diagram is provided in a cloud system load prediction method based on multi-model optimization according to an embodiment of the present invention.

[0067] Figure 3 The DLSTM network structure diagram in a cloud system load prediction method based on multi-model optimization provided in an embodiment of the present invention;

[0068] Figure 4 An LSTM cell structure diagram in a cloud system load prediction method based on multi-model optimization provided in an embodiment of the present invention;

[0069] Figure 5 The flowchart illustrates the optimization of the CNN-DLSTM model using the Sparrow Search Algorithm (SSA) in a cloud system load prediction method based on multi-model optimization, as provided in this embodiment of the invention.

[0070] Figure 6 A cloud system load data graph in a cloud system load prediction method based on multi-model optimization provided in an embodiment of the present invention;

[0071] Figure 7 The wavelet transform result is shown in the cloud system load prediction method based on multi-model optimization provided in an embodiment of the present invention.

[0072] Figure 8 The cloud system load prediction result is provided in a cloud system load prediction method based on multi-model optimization in an embodiment of the present invention. Detailed Implementation

[0073] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0074] This application provides a cloud system load prediction method, system, storage medium, and electronic device based on multi-model optimization, which solves the technical problem of low prediction accuracy of traditional prediction methods for cloud system load data with large fluctuations.

[0075] The technical solution in this application is to solve the above-mentioned technical problems, and the general idea is as follows:

[0076] Cloud system load data exhibits nonlinearity, randomness, and burstiness. Wavelet transform (WT) refines signals across multiple scales through scaling and translation operations, enabling localized analysis of time and spatial frequencies. This achieves time-level subdivision at high frequencies and frequency-level subdivision at low frequencies, automatically adapting to the requirements of time-frequency signal analysis. Deep long short-term memory recurrent neural networks (DLSTM) can detect nonlinear characteristics of sequential data based on time dependencies and establish predictive models for nonlinear data under different characteristics. Convolutional neural networks (CNNs) can extract local features, capture local time dependencies in data, handle short-term trends and fluctuations, and remove noise, allowing DLSTMs to focus more on long-term dependencies and enhancing their performance. Sparrow search algorithms optimize the hyperparameters of neural network models, replacing manual parameter tuning and overcoming the problems of models easily getting trapped in local optima and slow convergence during prediction.

[0077] In summary, this paper proposes a method based on wavelet transform (WT), convolutional neural network (CNN), deep long short-term memory (LSTM) neural network, and sparrow search algorithm to predict the changing trend of cloud system load data. Specifically, it provides a cloud system load prediction method based on wavelet transform (WT), sparrow search algorithm (SSA), convolutional neural network (CNN), and deep long short-term memory (LSTM) neural network.

[0078] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0079] Example 1:

[0080] like Figure 1 As shown, this embodiment of the invention provides a cloud system load prediction method based on multi-model optimization, including:

[0081] S1. Collect raw cloud system load time series data;

[0082] S2. Preprocess the cloud system load time series data, including using wavelet transform decomposition to obtain several sub-sequences, and normalizing the data of each sub-sequence;

[0083] S3. Based on each normalized subsequence, establish corresponding initial CNN-DLSTM subsequence prediction models to obtain the prediction results of each normalized subsequence.

[0084] S4. Establish fitness functions for each normalized subsequence and its prediction results, and use the sparrow search algorithm to optimize the hyperparameters of each initial CNN-DLSTM subsequence prediction model to obtain the best parameter combination.

[0085] S5. Based on each normalized subsequence and the optimal parameter combination, establish a CNN-DLSTM subsequence prediction model to obtain the prediction results of each subsequence before inverse normalization.

[0086] S6. Perform inverse normalization to obtain the actual prediction results of each subsequence, and use wavelet reconstruction to accumulate each actual prediction result to obtain the final cloud system load prediction result.

[0087] This invention overcomes the shortcomings of traditional models, such as being prone to getting trapped in local optima, slow convergence, and instability, by utilizing wavelet transform for data mining of cloud system load time series data and employing the Sparrow Search algorithm to optimize neural network hyperparameters. This prediction method can extract the characteristic changes in cloud system load data, ultimately achieving high-accuracy prediction and analysis of cloud system load time series data.

[0088] The following will detail each step of the above solution:

[0089] In step S1, raw cloud system load time series data is collected.

[0090] This step collects the raw cloud system load time series data, denoted as X={x1,x2,x3,...,x i ,...,x n}, where n represents the total amount of load data, x i This represents the load value of the i-th cloud system.

[0091] In step S2, the cloud system load time series data is preprocessed, including obtaining several sub-sequences using wavelet transform decomposition and normalizing the data of each sub-sequence; including:

[0092] S21. Several subsequences are obtained using wavelet transform decomposition, represented as follows:

[0093]

[0094] Where h(i) is a low-pass filter, corresponding to the wavelet scaling function;

[0095] g(i) is a high-pass filter, corresponding to the mother wavelet function of the wavelet;

[0096] d i Indicates by x i Decomposed load values;

[0097] If the original cloud system load time series data is decomposed m times, then X = {x1, x2, x3, ..., x...} i ,...,x n Decompose it into m subsequences D = {D1, D2, ..., D} t ,...,D m The subsequence contains one approximate component D and m-1 detail components D2,...,D t ,...,D m Each component consists of n load values ​​D t ={d1,d2,...,d n}

[0098] For example, in this embodiment of the invention, the Daubechies wavelet function is selected as the discrete wavelet function to perform signal decomposition on the cloud system load time series data. The decomposition depth is 3. The Mallat pyramid algorithm is used to decompose and reconstruct the cloud system load time series data. The decomposition process is as follows: Figure 2 As shown, three detail components and one approximation component are obtained.

[0099] S22. Normalize the data for each subsequence, including:

[0100] Map each decomposed subsequence to the interval [0,1];

[0101] Calculate the maximum and minimum values ​​of each subsequence, and denote them as d. max and d min ;

[0102] Take each subsequence D from each subsequence t Subtract d min Then divide by d max -d min Obtain D1', D'2, ..., D respectively. t ',,D' m , , respectively represent the normalized results of each subsequence.

[0103] In step S3, based on each normalized subsequence, a corresponding initial CNN-DLSTM subsequence prediction model is established to obtain the prediction results of each normalized subsequence.

[0104] This step first divides the subsequence into multiple shorter time windows using a sliding window approach. Each window contains several consecutive load values, representing the prediction of the load change at the next time step based on these consecutive load values. The data from each time window is then used as an input sample and fed into the CNN for local feature extraction, specifically:

[0105] S31. Use a sliding window to process the normalized subsequence D. t 'Format the matrix L as the input matrix of the CNN, where each element of L is an a-dimensional feature vector, and a is the size of the time window;

[0106] S32. A CNN is used to extract local features on the input matrix L. A convolution kernel matrix K is defined, which consists of k a-dimensional vectors, where k is the size of the convolution window, to perform a one-dimensional convolution operation on L, as shown in the following formula:

[0107] o q =ReLU(K·L) r:r+k-1 +b) (2)

[0108] Among them, L r:r+k-1 Let L represent the submatrix from position r to r+k-1, b be the bias term, ReLU be the mapping function, and o be the offset term. q Let O be the q-th local feature of the feature map O;

[0109] S33. Use max pooling to reduce the dimensionality of the feature map, as shown in the following formula:

[0110] o q `=max(o q:q+p (3)

[0111] Where p is the pooling window size, o q ` represents the q-th local feature of the pooled feature map O`;

[0112] S34. Use DLSTM to process the pooled feature maps O', and output the hidden state matrix H. DLSTM ;

[0113] S35. Map the output of the DLSTM to the target dimension through a fully connected layer and an activation function, as follows:

[0114] Y t = softmax(W·H) DLSTM +b) (4)

[0115] Among them, Y t ` is the normalized subsequence D t The prediction result is given by W, where W is the fully connected weight matrix and softmax(·) is the probability distribution function.

[0116] In other words, this embodiment of the invention sets the convolutional kernel size and slides the kernel along the time axis with a specified stride to extract local features within the kernel's range. The feature maps output by the convolutional layers are processed by a non-linear activation function (ReLU) to help the CNN model distinguish abnormal noise from important features. Pooling operations are used to retain salient features and remove insignificant features, reducing the computational cost of the model. The feature maps processed by the CNN form a feature sequence, which is then input into a DLSTM for training. The DLSTM network structure is as follows: Figure 3 As shown.

[0117] Each LSTM cell stores data through control units for the input gate, forget gate, and output gate. The LSTM cell structure is as follows: Figure 4 As shown, its specific workflow is as follows:

[0118] At each time step, the LSTM unit receives the current state x through three gates. t The hidden state h of the LSTM at the previous time step t-1 These are the two types of external information inputs. In addition, each gate also receives an internal information input, namely the state c of the memory cell. t-1 After receiving input information, each gate will perform operations on inputs from different sources, and its activation is determined by its logic function. The input to the input gate, after being transformed by a nonlinear function, is superimposed with the memory cell state processed by the forget gate to form a new memory cell state c. t Ultimately, the memory cell state c t The output h of the LSTM unit is formed through nonlinear function operations and dynamic control of the output gate. t .

[0119] The calculation formulas for the variables are as follows, where W xc W xi W xf W xo To connect the input signal x t The weight matrix W hc W ih W hf W ho To connect the hidden layer output signal h t The weight matrix W ci W cf W co The output vector c of the activation function connecting neurons t The diagonal matrix of the sum gate function; b i b c b f b o σ is the bias vector; σ is the activation function.

[0120] it =σ(W xi x t +W hi h t-1 +W ci c t-1 +b i (5)

[0121] f t =σ(W xf x t +W hf h t-1 +W cf c t-1 +b f (6)

[0122] c t =f t c t-1 +i t tanh(W xc x t +W hc h t-1 +b c (7)

[0123] o t =σ(W xo x t +W ho h t-1 +W co c t +b o (8)

[0124] h t =o t tanh(c t (9)

[0125] In step S4, fitness functions are established for each normalized subsequence and its prediction results, and the sparrow search algorithm is used to optimize the hyperparameters of each initial CNN-DLSTM subsequence prediction model to obtain the optimal parameter combination.

[0126] For example, in this step, the mean squared error function (MSE) is established as the fitness function for each normalized subsequence and its prediction result.

[0127] Next, the process of optimizing the hyperparameters of the CNN-DLSTM model using the Sparrow Search algorithm in this step is as follows: Figure 5 As shown, it specifically includes:

[0128] S100. Determine the hyperparameters to be optimized as time window, number of hidden layers, and number of hidden layer nodes; and initialize the parameters, including updating the population size, maximum number of iterations, search radius, range of hyperparameters, and learning rate in sequence.

[0129] S200. Initialize the population by generating parameter values ​​for initial individuals using random numbers and perturbing them based on the defined search radius to generate population individuals. Use the parameter values ​​generated by the individuals to build and train the CNN-DLSTM subsequence prediction model, and calculate the prediction error of the model as the individual fitness value.

[0130] S300: After the algorithm iteration begins, the population is continuously updated, including the joiners, discoverers, watchers and the optimal sparrow position. Dynamic weights are introduced to improve the update strategy and balance the global search capability.

[0131] S400. After the algorithm iteration is completed, the sparrow with the best fitness value is obtained, and its corresponding hyperparameters are decoded as the best parameter combination for the CNN-DLSTM subsequence prediction model.

[0132] In step S5, a CNN-DLSTM subsequence prediction model is established based on each normalized subsequence and the optimal parameter combination to obtain the prediction results of each subsequence before inverse normalization.

[0133] This step uses the CNN-DLSTM subsequence prediction model corresponding to the optimal parameter combination to predict each normalized subsequence and obtain the prediction results before inverse normalization.

[0134] In step S6, inverse normalization is performed to obtain the actual prediction results of each subsequence, and wavelet reconstruction is used to accumulate each actual prediction result to obtain the final cloud system load prediction result.

[0135] In this step, inverse normalization is performed to obtain the actual prediction results for each subsequence; this means:

[0136] Based on the maximum value d of each subsequence after decomposition max and minimum value d min Multiply the unnormalized prediction results of each subsequence by d. max -d min Then, add d min .

[0137] To help better understand the advantages of the embodiments of the present invention, the following specific examples are provided:

[0138] This example uses a self-built OpenStack cloud platform and takes idle memory as the experimental object. Data is collected at five-minute intervals, with the unit being Mb.

[0139] Data such as Figure 6 As shown, the wavelet decomposition results are as follows: Figure 7 As shown, the prediction results of this method in the above cloud system load data are as follows: Figure 8 As shown.

[0140] Furthermore, the mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) were used as evaluation indicators, as shown in formulas (10), (11), and (12).

[0141]

[0142] Where n is the number of data samples, y predictive For the predicted value, y true This is the actual value.

[0143] Specifically, the comparison models in this example use Support Vector Machine (SVM) and Long Short-Term Memory (LSTM) networks. The ratio of training set to test set is also 8:2. The time window for SVM is set to 10, the kernel function is RBF, the time window for LSTM is set to 10, the learning rate is 0.05, and the number of hidden layer nodes is 256. The error comparison of each model is shown in Table 1.

[0144] Table 1 Comparison of Predictive Performance of Various Models

[0145] Predictive Model RMSE MAE MAPE SVM 45.341 43.044 10.573％ LSTM 33.217 30.674 8.253％ This method 14.702 10.914 3.147％

[0146] As shown in Table 1, compared with SVM and LSTM, the cloud system load prediction method provided by this embodiment of the invention has lower errors, higher prediction accuracy, and stronger stability in various indicators, indicating that this embodiment of the invention can accurately predict the changing trend of cloud system load. Thus, this embodiment of the invention completes the entire process of the cloud system load prediction method based on multi-model optimization.

[0147] Example 2:

[0148] This invention provides a cloud system load prediction system based on multi-model optimization, comprising:

[0149] The acquisition module is used to collect raw cloud system load time series data;

[0150] The preprocessing module is used to preprocess the cloud system load time series data, including obtaining several subsequences by wavelet transform decomposition and normalizing the data of each subsequence.

[0151] A module is established to build corresponding initial CNN-DLSTM subsequence prediction models based on each normalized subsequence, so as to obtain the prediction results of each normalized subsequence.

[0152] The optimization module is used to establish fitness functions for each normalized subsequence and its prediction results, and to optimize the hyperparameters of each initial CNN-DLSTM subsequence prediction model using the sparrow search algorithm to obtain the best parameter combination.

[0153] The prediction module is used to build a CNN-DLSTM subsequence prediction model based on each normalized subsequence and the optimal parameter combination, so as to obtain the prediction results of each subsequence before inverse normalization.

[0154] The acquisition module is used to perform inverse normalization to obtain the actual prediction results of each subsequence, and to accumulate each actual prediction result using wavelet reconstruction to obtain the final cloud system load prediction result.

[0155] Example 3:

[0156] This invention provides a storage medium storing a computer program for cloud system load prediction based on multi-model optimization, wherein the computer program causes a computer to execute the cloud system load prediction method as described in Embodiment 1.

[0157] Example 4:

[0158] This invention provides an electronic device, comprising:

[0159] One or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing cloud system load prediction as described in Example 1.

[0160] It is understood that the cloud system load prediction system, storage medium and electronic device based on multi-model optimization provided in the embodiments of the present invention correspond to the cloud system load prediction method based on multi-model optimization provided in the embodiments of the present invention. The explanation, examples and beneficial effects of the relevant contents can be referred to the corresponding parts of the cloud system load prediction method, and will not be repeated here.

[0161] In summary, compared with existing technologies, it has the following beneficial effects:

[0162] This invention overcomes the shortcomings of traditional models, such as being prone to getting trapped in local optima, slow convergence, and instability, by utilizing wavelet transform for data mining of cloud system load time series data and employing the Sparrow Search algorithm to optimize neural network hyperparameters. This prediction method can extract the characteristic changes in cloud system load data, ultimately achieving high-accuracy prediction and analysis of cloud system load time series data.

[0163] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0164] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A cloud system load prediction method based on multi-model optimization, characterized in that, include: Collect raw cloud system load time series data; The cloud system load time series data is preprocessed, including obtaining several subsequences using wavelet transform decomposition and normalizing the data of each subsequence. Based on each normalized subsequence, a corresponding initial CNN-DLSTM subsequence prediction model is established to obtain the prediction results of each normalized subsequence. Fitness functions were established for each normalized subsequence and its prediction results. The sparrow search algorithm was used to optimize the hyperparameters of each initial CNN-DLSTM subsequence prediction model to obtain the optimal parameter combination. The hyperparameters to be optimized are time window, number of hidden layers, and number of hidden layer nodes. Based on each normalized subsequence and the optimal parameter combination, a CNN-DLSTM subsequence prediction model is established to obtain the prediction results of each subsequence before inverse normalization. Inverse normalization is performed to obtain the actual prediction results of each subsequence, and wavelet reconstruction is used to accumulate each actual prediction result to obtain the final cloud system load prediction result. The step involves establishing corresponding initial CNN-DLSTM subsequence prediction models based on each normalized subsequence to obtain prediction results for each normalized subsequence; including: The normalized subsequence D is processed using a sliding window method. t 'Format the matrix L as the input matrix of the CNN, where each element of L is an a-dimensional feature vector, and a is the size of the time window; A CNN is used to extract local features from the input matrix L. A convolution kernel matrix K is defined, consisting of k a-dimensional vectors, where k is the size of the convolution window, to perform a one-dimensional convolution operation on L, as shown in the following equation: o q =ReLU(K·L r:r+k-1 +b) (1) Among them, L r:r+k-1 Let L represent the submatrix from position r to r+k-1, b be the bias term, ReLU be the mapping function, and o be the offset term. q Let O be the q-th local feature of the feature map O; Max pooling is used to reduce the dimensionality of the feature map, as shown in the following equation: o q `=max(o q:q+p ) (2) Where p is the pooling window size, o q ` represents the q-th local feature of the pooled feature map O`; The pooled feature map O' is processed using DLSTM, and the output hidden state matrix is ​​H. DLSTM ; The output of the DLSTM is mapped to the target dimension through a fully connected layer and an activation function, as follows: Y t `=softmax(W·H DLSTM +b) (3) Among them, Y t ` is the normalized subsequence D t The prediction result is given by W, where W is the fully connected weight matrix and softmax(·) is the probability distribution function.

2. The cloud system load prediction method as described in claim 1, characterized in that, The method of obtaining several sub-sequences using wavelet transform decomposition is represented as follows: Let the original cloud system load time series data be X={x1,x2,x3,...,x i ,...,x n }, where n represents the total amount of load data, x i This represents the load value of the i-th cloud system; h(i) is a low-pass filter, corresponding to the wavelet scaling function; g(i) is a high-pass filter, corresponding to the mother wavelet function of the wavelet; d i Indicates by x i Decomposed load values; If the original cloud system load time series data is decomposed m times, then X = {x1, x2, x3, ..., x...} i ,...,x n Decompose it into m subsequences D = {D1, D2, ..., D} t ,...,D m The subsequence contains one approximate component D and m-1 detail components D2,...,D t ,...,D m Each component consists of n load values ​​D t ={d1,d2,...,d n } 3. The cloud system load prediction method as described in claim 2, characterized in that, The data normalization of each subsequence includes: Map each decomposed subsequence to the interval [0,1]; Calculate the maximum and minimum values ​​of each subsequence, and denote them as d. max and d min ; Take each subsequence D from each subsequence t Subtract d min Then divide by d max -d min Obtain D′1, D′2, ..., D′ respectively. t ,...,D' m , , respectively represent the normalized results of each subsequence.

4. The cloud system load prediction method as described in claim 1, characterized in that, The mean squared error function (MSE) is established as the fitness function based on each normalized subsequence and its prediction results.

5. The cloud system load prediction method as described in claim 1, characterized in that, The method of using the sparrow search algorithm to optimize the hyperparameters of each initial CNN-DLSTM subsequence prediction model to obtain the optimal parameter combination includes: Perform parameter initialization, including sequentially updating the population size, maximum number of iterations, search radius, range of hyperparameters, and learning rate; Population initialization is performed by generating parameter values ​​for initial individuals using random numbers and perturbing them based on a defined search radius to generate population individuals. CNN-DLSTM subsequence prediction models are then built and trained using the parameter values ​​generated for each individual, and the prediction error of the model is calculated as the individual fitness value. After the algorithm iteration begins, the population is continuously updated, including the additions, discoverers, watchers, and the optimal sparrow position. The update strategy is improved by introducing dynamic weights to balance the global search capability. After the algorithm iteration is completed, the sparrow with the best fitness value is obtained, and its corresponding hyperparameters are decoded as the optimal parameter combination for the CNN-DLSTM subsequence prediction model.

6. The cloud system load prediction method as described in claim 1, characterized in that, The step of performing inverse normalization to obtain the actual prediction results for each subsequence refers to: Based on the maximum value d of each subsequence after decomposition max and minimum value d min Multiply the unnormalized prediction results of each subsequence by d. max -d min Then, add d min .

7. A cloud system load prediction system based on multi-model optimization, characterized in that, For performing the cloud system load prediction method as described in claim 1, including: The acquisition module is used to collect raw cloud system load time series data; The preprocessing module is used to preprocess the cloud system load time series data, including obtaining several subsequences by wavelet transform decomposition and normalizing the data of each subsequence. A module is established to build corresponding initial CNN-DLSTM subsequence prediction models based on each normalized subsequence, so as to obtain the prediction results of each normalized subsequence. The optimization module is used to establish fitness functions for each normalized subsequence and its prediction results, and to optimize the hyperparameters of each initial CNN-DLSTM subsequence prediction model using the sparrow search algorithm to obtain the best parameter combination. The prediction module is used to build a CNN-DLSTM subsequence prediction model based on each normalized subsequence and the optimal parameter combination, so as to obtain the prediction results of each subsequence before inverse normalization. The acquisition module is used to perform inverse normalization to obtain the actual prediction results of each subsequence, and to accumulate each actual prediction result using wavelet reconstruction to obtain the final cloud system load prediction result.

8. A storage medium, characterized in that, It stores a computer program for cloud system load prediction based on multi-model optimization, wherein the computer program causes a computer to perform the cloud system load prediction method as described in any one of claims 1 to 6.

9. An electronic device, characterized in that, include: One or more processors; Memory; And one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing cloud system load prediction as described in any one of claims 1 to 6.

Images