A Cloud Server Load Prediction Method Based on an Extreme Learning Machine with a Hybrid Optimization Strategy

Abstract

The present invention discloses a cloud server load prediction method based on an extreme learning machine with a hybrid optimization strategy. The steps are as follows: 1) Establish an extreme learning machine model; 2) Construct a whale optimizer; 3) Update the whale optimizer using a diverse hybrid optimization strategy; 4) Introduce a migration strategy into the whale optimizer to obtain a hybrid strategy whale optimizer; 5) Obtain the historical load data of the cloud server and train the hybrid strategy whale optimizer to obtain the optimal hyperparameters P; 6) Input the optimal hyperparameters into the extreme learning machine model to establish a cloud server load prediction model; 7) Obtain the current cloud server load data and input it into the cloud server load prediction model to predict the future load of the cloud server. The present invention provides a cloud server load prediction method based on an extreme learning machine with a hybrid optimization strategy, which can greatly improve the prediction accuracy of the cloud server load.

Description

Technical Field

[0001] The present invention relates to the field of cloud computing, and particularly to a cloud server load prediction method based on a hybrid optimization strategy extreme learning machine. Background Art

[0002] With the rapid development of communication technologies, the number of various mobile terminal devices has been increasing continuously, and a variety of software application services have emerged as the times require, which directly leads to an explosive growth in the amount of data in the Internet, making traditional data processing fall into a bottleneck. Cloud computing has powerful computing resources and can process massive amounts of data in a short time through distributed parallel computing. However, how to achieve efficient resource management has always been the most concerned issue for cloud service operators and providers. As one of the necessary measures for effectively monitoring the cloud network situation, cloud load prediction has become a hot topic in the field. The more accurate the prediction result is, the higher the resource utilization rate will be, the faster the response speed will be, and ultimately, better service quality and higher user satisfaction rate will be brought. At the same time, accurate load prediction will also lay a solid foundation for the management of cloud resources and bring huge economic benefits to cloud service operators and providers.

[0003] A large number of cloud server load prediction methods have been proposed in previous studies, and some of them are practical methods that have been successfully applied in the industry. Generally speaking, these prediction methods can be roughly divided into two categories, namely, methods based on time series models and methods based on machine learning models. Among them, the methods based on time series models convert non-stationary sequences into stationary sequences, but the time series data must be strictly continuous. At the same time, data preprocessing is required before calculation, resulting in a relatively complex calculation process and being not suitable for dealing with situations such as data loss. The methods based on machine learning models include support vector machines, neural networks, and Gaussian process regression, etc. In practice, although cloud load data has high complexity, machine learning models can effectively predict cloud load resources with non-linear changes. In addition, compared with the methods based on time series prediction models, the methods based on machine learning models have better generalization ability and mapping ability.

[0004] The extreme learning machine is a type of machine learning model constructed based on a feedforward neural network. Its characteristic is that the weights of the hidden layer nodes are randomly or artificially given and do not need to be updated, and the learning process only calculates the output weights. When the extreme learning machine is compared with other shallow machine learning models such as single-layer perceptrons and support vector machines, it has more advantages in terms of learning rate and generalization ability. However, the parameter setting of the extreme learning machine model has a great influence on the prediction effect of cloud load, and how to select the optimal parameters is still a thorny problem. Summary of the Invention

[0005] The objective of the present invention is to provide a cloud server load prediction method based on an extreme learning machine with a hybrid optimization strategy, comprising the following steps:

[0006] 1) Establish an extreme learning machine model.

[0007] The extreme learning machine model includes an input layer, a hidden layer, and an output layer. The input layer, hidden layer, and output layer are connected by neurons, and the number of neurons in each layer is u, r, and e. The weight coefficient between the hidden layer and the input layer is S, and the weight coefficient between the hidden layer and the output layer is P. θ is the set of neuron thresholds of the hidden layer,

[0008] wherein, the weight coefficient S, the weight coefficient P, and the set θ are respectively as follows:

[0009]

[0010]

[0011]

[0012] In the formula, s ru is the weight coefficient between the r-th hidden layer neuron and the u-th input layer neuron. p re is the weight coefficient between the r-th hidden layer neuron and the e-th output layer neuron. θ r is the threshold of the r-th hidden layer neuron.

[0013] The activation function of the extreme learning machine model is y(·), and the output is F, that is:

[0014] F = [f1, f2,..., f M (4)

[0015]

[0016] In the formula, the weight coefficient s i = [s i1 , s i2 ,..., s iu . a i = [a 1j , a 2j ,..., a uj T is the i-th column of the input matrix A. The transpose matrix F T of the output F = LP.

[0017] wherein, the hidden layer output matrix L is as follows:

[0018]

[0019] 2) Construct a whale optimizer,

[0020] The whale optimizer includes a foraging encircling stage, a random search stage, and a bubble net attacking stage.

[0021] In the foraging encircling stage, the position of the whale is updated as follows:

[0022]

[0023] Where pos(t + 1) is the predicted position of the individual whale. t is the current iteration number. is the position of the best individual whale. K and Q are coefficient vectors. The coefficient vector K = 2k·n - k. The parameter k = 2 - (2t) / MAX iter β. The coefficient vector Q = |J·pos(t) - pos(t)|. The parameter J = 2n. n is a random vector in [0, 1].

[0024] In the random search stage, the position of the whale is updated as follows:

[0025] pos(t + 1) = pos rand (t) - K·Q (8)

[0026] Q = |J·pos rand (t) - pos(t)|(9)

[0027] Where pos rand (t) is the position of the random whale. When |K| ≤ 1, the optimal individual is selected as the solution to achieve the local optimization of the whale optimizer. When |K| > 1, an individual is randomly selected as the solution to achieve the global optimization of the whale optimizer.

[0028] In the bubble net attacking stage, the position of the whale is updated as follows:

[0029]

[0030] Where is the distance between the individual whale and the best individual obtained so far. o is a constant defining the spiral shape. l is a number ranging from -1 to 1.

[0031] 3) Update the whale optimizer using a diverse hybrid optimization strategy. v is a random number;

[0032] The steps for updating the whale optimizer using a diverse hybrid optimization strategy include:

[0033] 3.1) Initialize the whale population using the Levy strategy. The mathematical model of the Levy strategy is shown as follows:

[0034]

[0035] In the formula, λ is the random step size.

[0036] Among them, the parameter μ and the parameter η follow the normal distribution, that is:

[0037]

[0038] Parameter δ μ and the parameter δ η are respectively as follows:

[0039]

[0040] In the formula, Γ(·) is the standard gamma function.

[0041] 3.2) Based on the Levy strategy, calculate the initial position of the individual, that is:

[0042]

[0043] In the formula, L b is the lower limit, U b is the upper limit, and Levy(v) is a random vector of the step size subject to the Levy distribution.

[0044] 3.3) Introduce the non-linear convergence factor β into the whale optimizer, that is:

[0045]

[0046] In the formula, Max iter is the maximum number of iterations.

[0047] 4) Introduce the migration strategy into the whale optimizer to obtain the hybrid strategy whale optimizer.

[0048] After introducing the migration strategy into the whale optimizer, the position update of the whale group is as follows.

[0049] pos(t + 1) = B * pos(t) + pos(t) * randn(0, σ 2 ) (16)

[0050]

[0051] In the formula, B is the convergence coefficient, and randn(0, σ 2 ) conforms to the Gaussian distribution. B max , B min are the upper and lower limits of the convergence coefficient B.

[0052] 5) Obtain the historical load data of the cloud server and train the hybrid strategy whale optimizer to obtain the optimal hyperparameters.

[0053] The load data of the cloud server has been normalized.

[0054] The load data of the cloud server includes a training set and a test set. Among them, the training set is used to train the hybrid strategy whale optimizer, and the test set is used to test the hybrid strategy whale optimizer.

[0055] The training set of the extreme learning machine model includes an input matrix A and an output matrix B, that is:

[0056]

[0057]

[0058] In the formula, M is the number of training samples of the cloud server load data.

[0059] During the training process of the hybrid strategy whale optimizer, the weight coefficient s and the neuron threshold θ are constants, and the weight coefficient P is an unknown parameter. The weight coefficient P is solved by the least squares method.

[0060] 6) Input the optimal hyperparameters into the extreme learning machine model to establish a cloud server load prediction model.

[0061] 7) Obtain the current cloud server load data and input it into the cloud server load prediction model to predict the future load of the cloud server.

[0062] It should be noted that considering the obvious non - linear and non - stationary characteristics of the cloud server load data and the relatively poor applicability of the time - series - model - based method, the present invention proposes a hybrid optimization strategy extreme learning machine to predict the cloud server load. Among them, an extreme learning machine model with strong non - linear mapping ability and generalization ability is used. Since the present invention does not use the gradient descent method to update parameters in cloud server load prediction, compared with the traditional neural network model, the model complexity and training time are greatly reduced.

[0063] In addition, since the parameter setting of the extreme learning machine model will affect the final prediction accuracy. Therefore, the present invention proposes a hybrid optimization strategy based on the whale optimizer to determine the hyperparameters of the extreme learning machine model. Specifically, the present invention uses the Levy strategy to initialize the population. The random walk mode of the Levy strategy expands the search range of the whale group, effectively avoiding the whale individuals falling into local extrema. Therefore, it helps the whale optimizer to have better convergence performance. At the same time, in the foraging and encircling stages, since the convergence speed of the whale optimizer is the same in the early and late stages, this greatly reduces the overall convergence speed of the optimization. Therefore, a non - linear convergence factor is introduced into the whale optimizer to solve this problem.

[0064] The technical effect of the present invention is beyond doubt. The present invention provides a cloud server load prediction method based on a hybrid optimization strategy extreme learning machine, which can greatly improve the prediction accuracy of cloud server load.

[0065] The present invention comprehensively considers the nonlinear and non-stationary characteristics of cloud server load data, proposes a hybrid strategy whale optimizer with better optimization performance, and combines the proposed hybrid strategy whale optimizer with the extreme learning machine model to provide effective and accurate prediction, greatly enhancing the use efficiency of cloud server computing resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 is a framework diagram of the cloud server load prediction algorithm;

[0067] Figure 2 is a flowchart of the cloud server load prediction algorithm; DETAILED DESCRIPTION OF THE EMBODIMENTS

[0068] The present invention will be further described below in conjunction with embodiments, but it should not be understood that the above-mentioned subject matter scope of the present invention is limited to the following embodiments. Without departing from the above-mentioned technical idea of the present invention, various substitutions and changes should be included in the protection scope of the present invention according to the common general knowledge and customary means in the art.

[0069] Embodiment 1:

[0070] See Figures 1 to 2 , a cloud server load prediction method based on a hybrid optimization strategy extreme learning machine, comprising the following steps:

[0071] 1) Establish an extreme learning machine model.

[0072] The extreme learning machine model includes an input layer, a hidden layer, and an output layer. The input layer, hidden layer, and output layer are connected by neurons, and the number of neurons in each layer is u, r, and e. The weight coefficient between the hidden layer and the input layer is S, and the weight coefficient between the hidden layer and the output layer is P. θ is a set of neuron thresholds of the hidden layer,

[0073] wherein, the weight coefficient S, the weight coefficient P, and the set θ are respectively as follows:

[0074]

[0075]

[0076]

[0077] In the formula, s ru is the weight coefficient between the rth hidden layer neuron and the uth input layer neuron. pre is the weight coefficient between the r-th hidden layer neuron and the e-th output layer neuron. θ r is the threshold of the r-th hidden layer neuron.

[0078] The activation function of the extreme learning machine model is y(·), and the output is F, that is:

[0079] F = [f1, f2,..., f M (4)

[0080]

[0081] In the formula, the weight coefficient s i = [s i1 , s i2 ,..., s iu . a i = [a 1j , a 2j ,..., a uj T is the i-th column of the input matrix A;. The transpose matrix F of the output F T = LP.

[0082] Among them, the hidden layer output matrix L is shown as follows:

[0083]

[0084] 2) Construct a whale optimizer,

[0085] The whale optimizer includes a foraging encircling stage, a random search stage, and a bubble net attack stage.

[0086] In the foraging encircling stage, the position of the whale is updated as follows:

[0087]

[0088] In the formula, pos(t + 1) is the predicted position of the whale individual. t is the current iteration number. is the position of the best whale individual. K and Q are coefficient vectors. The coefficient vector K = 2k·n - k. The parameter k = 2 - (2t) / MAX iter β. The coefficient vector Q = |J·pos(t) - pos(t)|. The parameter J = 2n. n is a random vector in [0, 1].

[0089] In the random search stage, the position of the whale is updated as follows:

[0090] pos(t + 1) = pos rand (t) - K·Q (8)

[0091] Q = |J·pos​rand (t)-pos(t)| (9)

[0092] where pos rand (t) is the position of the random whale. When |K| ≤ 1, the optimal individual is selected as the solution to achieve the local optimization of the whale optimizer. When |K| > 1, an individual is randomly selected as the solution to achieve the global optimization of the whale optimizer.

[0093] During the bubble net attack phase, the position of the whale is updated as follows:

[0094]

[0095] where is the distance between the whale individual and the best individual obtained so far. o is a constant defining the spiral shape. l is a number ranging between -1 and 1.

[0096] 3) Update the whale optimizer using a diverse hybrid optimization strategy. v is a random number;

[0097] The steps for updating the whale optimizer using a diverse hybrid optimization strategy include:

[0098] 3.1) Initialize the whale population using the Levy strategy. The mathematical model of the Levy strategy is shown as follows:

[0099]

[0100] where λ is the random step size.

[0101] Among them, the parameter μ and the parameter η follow the normal distribution, that is:

[0102]

[0103] The parameter δ μ and the parameter δ η are shown as follows respectively:

[0104]

[0105] where Γ(·) is the standard gamma function.

[0106] 3.2) Based on the Levy strategy, calculate the initial position of the individual, that is:

[0107]

[0108] where L b is the lower limit, U b is the upper limit, and Levy(v) is a random vector of the step size following the Levy distribution.

[0109] 3.3) Introduce the non-linear convergence factor β into the whale optimizer, i.e.:

[0110]

[0111] In the formula, Max iter is the maximum number of iterations.

[0112] 4) Introduce the migration strategy into the whale optimizer to obtain the hybrid strategy whale optimizer.

[0113] After introducing the migration strategy into the whale optimizer, the position update of the whale swarm is as follows.

[0114] pos(t + 1) = B * pos(t) + pos(t) * randn(0, σ 2 ) (16)

[0115]

[0116] In the formula, B is the convergence coefficient, and randn(0, σ 2 ) conforms to the Gaussian distribution. B max , B min are the upper and lower limits of the convergence coefficient B.

[0117] 5) Obtain the historical load data of the cloud server and train the hybrid strategy whale optimizer to obtain the optimal hyperparameters.

[0118] The load data of the cloud server has been normalized.

[0119] The load data of the cloud server includes a training set and a test set. Among them, the training set is used to train the hybrid strategy whale optimizer, and the test set is used to test the hybrid strategy whale optimizer.

[0120] The training set of the extreme learning machine model includes the input matrix A and the output matrix B, i.e.:

[0121]

[0122]

[0123] In the formula, M is the number of training samples in the cloud server load data set.

[0124] During the training process of the hybrid strategy whale optimizer, the weight coefficient s and the neuron threshold θ are constants, and the weight coefficient P is an unknown parameter. The weight coefficient P is solved by the least squares method.

[0125] 6) Input the optimal hyperparameters into the extreme learning machine model to establish a cloud server load prediction model.

[0126] 7) Obtain the current cloud server load data and input it into the cloud server load prediction model to predict the future load of the cloud server.

[0127] Embodiment 2:

[0128] A cloud server load prediction method based on an extreme learning machine with a hybrid optimization strategy, comprising the following steps:

[0129] 1) Divide the cloud server load data, determine the training set and test set of the cloud server load prediction model, and perform unified normalization processing on the data set;

[0130] 2) Construct an extreme learning machine model.

[0131] 2.1) The entire model consists of a three-layer network structure, namely an input layer, a hidden layer, and an output layer, and these three-layer network structures are connected by neurons. For the extreme learning machine model, the number of neurons in each layer is u, r, and e. The weight coefficient between the hidden layer and the input layer is S, and the weight coefficient between the hidden layer and the output layer is P. θ is the set of neuron thresholds of the hidden layer, where,

[0132]

[0133]

[0134]

[0135] 2.2) The extreme learning machine model has M cloud server load data training sample sets. A is the input matrix and B is the output matrix, where,

[0136]

[0137]

[0138] The activation function of the extreme learning machine model is y(·), and the output is F. The expressions of both are

[0139] F = [f1, f2,..., f M , (6)

[0140]

[0141] where s i = [s i1 , s i2 ,..., s iu , a i = [a 1j , a 2j , …, a uj T ​. Meanwhile, F T = LP, where

[0142] where L is the output matrix of the hidden layer, and F T is the transpose matrix of F. Specifically, when y is infinitely divisible, s and θ will be constants during the training process. Meanwhile, the only unknown parameter is P, and the parameter P can be solved by the least squares method, that is, min P ||LP - F T ||.

[0143] 3) Construct a whale optimizer to simulate the hunting strategy of humpback whales.

[0144] 3.1) Foraging encirclement stage. In this process, the whale group shares the position information of the prey and finally encircles the prey. Assume that the current best individual in the whale group is the target position, and other individuals update their positions and try to approach the best individual position. The position update formula of the whale is as follows:

[0145]

[0146] where pos(t) is the current position of the whale individual, t is the current iteration number, is the position of the best whale individual, and K and Q are coefficient vectors. K and Q can be calculated by K = 2k·n - k, J = 2n, k = 2 - (2t) / MAX iter β, Q = |J·pos(t) - pos(t)| respectively, where n is a random vector in [0, 1], and k decreases from 2 to 0 during the iteration process.

[0147] 3.2) Random search stage. In this stage, a single current whale position is randomly selected as the best solution. Contrary to the previous process, the position of the whale is updated according to the randomly selected whale rather than the best whale individual found so far. Adjusting the value of K will make other individuals move away from the selected whale, and the update formula is as follows:

[0148] pos(t + 1) = pos rand (t) - K·Q, (10)

[0149] Q = |J·pos rand (t) - pos(t)|, (11)

[0150] where pos rand (t) is the position of the random whale. When |K| ≤ 1, the optimal individual is selected as the solution to achieve the local optimization of the whale optimizer. When |K| > 1, an individual is randomly selected as the solution to achieve the global optimization of the whale optimizer.

[0151] 3.3) Foam net attack stage

[0152] In this stage, the whale has two behaviors: shrinking, surrounding and hovering. The surrounding movement of the whale is achieved by reducing k in step 2.1. The spiral function is used to simulate the spiral behavior of the whale. The choice of these two behaviors is determined by the random number v to achieve simultaneous shrinking and surrounding along the spiral path. The position update of individual whales is as follows:

[0153]

[0154] where is the distance between the individual whale and the best individual obtained so far, o is the constant defining the spiral shape, and l is a number ranging from -1 to 1.

[0155] 4) Update the whale optimizer using a diverse hybrid optimization strategy to further enhance the local and global optimization capabilities.

[0156] 4.1) Initialize the whale population using the Levy strategy. The population initialization has an important impact on the optimization process of the optimizer. At the same time, a high-quality initial state can accelerate the convergence speed of the optimizer. The Levy strategy is used to initialize the population. Thanks to its random walk pattern, it can expand the search range of the whale group and effectively avoid individual whales falling into local extrema. The mathematical model of the Levy strategy is as follows:

[0157]

[0158] where λ is the random step size, and μ and η follow the normal distribution, i.e.,

[0159] , δ μ and δ η The values of can be calculated by

[0160]

[0161] where Γ(·) is the standard gamma function. Then the initial position based on the Levy strategy can be calculated by the following formula:

[0162]

[0163] where, L b is the lower limit, U b is the upper limit, and Levy(v) is a random vector of the step size following the Levy distribution.

[0164] 4.2) Nonlinear convergence factor. In the foraging and encirclement stages, a nonlinear convergence factor is introduced into the whale optimizer to improve the convergence speed. The nonlinear convergence factor is:

[0165]

[0166] where Max iter is the maximum number of iterations.

[0167] 5) Introduce the migration strategy into the whale optimizer to increase the diversity of the whale population and enhance the ability of whales to jump out of local minima. The formula for the latest position of the whale population after migration is as follows:

[0168] pos(t + 1) = B * pos(t) + pos(t) * randn(0, σ 2 ), (18)

[0169]

[0170] where, where B(B max = 0.9, B min = 0.4) is the convergence coefficient, and randn(0, σ 2 ) follows a Gaussian distribution.

[0171] 6) Initialize the parameters of the hybrid strategy whale optimizer, input the cloud server load training set into the optimizer for model training, obtain the optimal hyperparameters and then input them into the extreme learning machine model. Finally, input the cloud server load data test set into the extreme learning machine model to predict the cloud server load.

Claims

1. A cloud server load prediction method based on an extreme learning machine with a hybrid optimization strategy, characterized in that, It includes the following steps: 1) Establish an extreme learning machine model; 2) Construct a whale optimizer; 3) Update the whale optimizer using a diverse hybrid optimization strategy; 4) Introduce a migration strategy into the whale optimizer to obtain a hybrid strategy whale optimizer; 5) Obtain the historical load data of the cloud server and train the hybrid strategy whale optimizer to obtain the optimal hyperparameter P; 6) Input the optimal hyperparameters into the extreme learning machine model to establish a cloud server load prediction model; 7) Obtain the current cloud server load data and input it into the cloud server load prediction model to predict the future load of the cloud server; The steps of updating the whale optimizer using a diverse hybrid optimization strategy include: 3.1) Initialize the whale population using the Levy strategy; the mathematical model of the Levy strategy is as follows: In the formula, λ is the random step size; v is a random number; Among them, the parameters μ and η follow a normal distribution, that is: Parameter δ μ and parameter δ η are respectively as follows: In the formula, Γ(·) is the standard gamma function; 3.2) Based on the Levy strategy, calculate the initial position of the individual, that is: where L b is the lower limit, U b is the upper limit, and Levy(v) is a random vector of step sizes following the Levy distribution; 3.3) Introduce a non-linear convergence factor β into the whale optimizer, that is: where, Max iter is the maximum number of iterations.

2. The cloud server load prediction method based on the extreme learning machine with a hybrid optimization strategy according to claim 1, wherein The extreme learning machine model includes an input layer, a hidden layer, and an output layer; the input layer, hidden layer, and output layer are connected by neurons, and the number of neurons in each layer is u, r, and e; the weight coefficient between the hidden layer and the input layer is S, and the weight coefficient between the hidden layer and the output layer is P; θ is the set of neuron thresholds of the hidden layer, Among them, the weight coefficient S, the weight coefficient P, and the set θ are as follows respectively: where s ru is the weight coefficient between the r-th hidden layer neuron and the u-th input layer neuron; p re is the weight coefficient between the r-th hidden layer neuron and the e-th output layer neuron; θ r is the threshold of the r-th hidden layer neuron.

3. A cloud server load prediction method based on a hybrid optimization strategy extreme learning machine according to claim 1, characterized in that: The cloud server load data has been normalized.

4. A cloud server load prediction method based on an extreme learning machine with a hybrid optimization strategy according to claim 3, characterized in that, The cloud server load data includes a training set and a test set; among them, the training set is used to train the hybrid strategy whale optimizer, and the test set is used to test the hybrid strategy whale optimizer; The training set of the extreme learning machine model includes an input matrix A and an output matrix B, that is: In the formula, M is the number of cloud server load data training samples.

5. A cloud server load prediction method based on an extreme learning machine with a hybrid optimization strategy according to claim 4, characterized in that, During the training process of the hybrid strategy whale optimizer, the weight coefficient s and the neuron threshold θ are constants, and the weight coefficient P is an unknown parameter; The weight coefficient P is solved using the least squares method.

6. A cloud server load prediction method based on an extreme learning machine with a hybrid optimization strategy according to claim 1, characterized in that, The activation function of the extreme learning machine model is y(·), and the output is F, that is: F = [f1, f2,..., f M (6) where the weight coefficient is s i = [s i1 , s i2 ,..., s iu ; a i = [a 1j , a 2j ,..., a uj T is the i-th column of the input matrix A; the transpose matrix F T of the output F = LP; f M is an element of the output matrix F;​ Among them, the hidden layer output matrix L is as above.

7. A cloud server load prediction method based on an extreme learning machine with a hybrid optimization strategy according to claim 1, characterized in that The whale optimizer includes a foraging encirclement stage, a random search stage, and a bubble net attack stage; In the foraging encirclement stage, the position update of the whale is as follows: In the formula, pos(t + 1) is the predicted position of the whale individual; t is the current iteration number; is the position of the best whale individual; K and Q are coefficient vectors; the coefficient vector K = 2k·n - k; the parameter k = 2 - (2t) / MAX iter β, where β is a non-linear convergence factor; the coefficient vector Q = |J·pos(t) - pos(t)|; the parameter J = 2n; n is a random vector in [0, 1]; In the random search stage, the position update of the whale is as follows: pos(t + 1)=pos rand (t)-K·Q(10) Q = |J·pos rand (t) - pos(t)| (11) where pos rand (t) is the position of the random whale; when |K| ≤ 1, the optimal individual is selected as the solution to achieve the local optimization of the whale optimizer; when |K| > 1, an individual is randomly selected as the solution to achieve the global optimization of the whale optimizer; In the bubble net attack stage, the position update of the whale is as follows: In the formula, is the distance between the whale individual and the best individual obtained so far; o is a constant defining the spiral shape; l is a number ranging between -1 and 1; v is a random number.

8. A cloud server load prediction method based on a hybrid optimization strategy extreme learning machine according to claim 1, characterized in that After introducing the migration strategy into the whale optimizer, the position update of the whale group is as follows; pos(t + 1)=B * pos(t)+pos(t)*randn(0,σ 2 ) (18) where B is the convergence coefficient, and randn(0, σ 2 ) follows a Gaussian distribution; B max , B min are the upper and lower limits of the convergence coefficient B.

Images