Satellite network self-similarity traffic prediction method based on double decomposition and neural network

Abstract

The invention provides a satellite network self-similar traffic prediction method based on double decomposition and a neural network, and relates to the technical field of satellite networks. The method comprises the following steps: firstly, performing first modal decomposition on an original self-similar traffic sequence of a satellite network by using a CEEMDAN algorithm, and decomposing the original self-similar traffic data sequence into a plurality of IMF components; dividing the IMF components into high-frequency components, intermediate-frequency components and low-frequency components by calculating the sample entropy of each IMF component after decomposition; performing second modal decomposition on the clustered high-frequency components by using a VMD decomposition model; then constructing a CNN-BiLSTM-based self-similar flow prediction model, and respectively inputting each IMF component subjected to two-time decomposition into the model for prediction; meanwhile, an improved artificial rabbit optimization algorithm is adopted to optimize hyper-parameters of the self-similar flow prediction model, and an optimal self-similar flow prediction model is obtained. The method can effectively improve prediction precision and reduce prediction errors.

Description

Technical Field

[0001] The present invention relates to the field of satellite network technology, and in particular to a satellite network self-similar traffic prediction method based on double decomposition and neural network. Background Art

[0002] Satellite networks have global coverage, high bandwidth and high-speed transmission characteristics, which play an important role in modern communications and data transmission. They are an ideal choice for communications and Internet access in special environments, supporting broadcast transmission and satellite TV services, and providing users with high-quality transmission experience. There are three types of services in satellite networks: services generated on the satellite, services from other satellites and services from ground networks. These services cover a large amount of multimedia information such as images, audio, and video. Studies have shown that self-similarity is an important attribute of multimedia services. A large number of multimedia services are transmitted in satellite networks, and their traffic has self-similarity. In addition, the service traffic of the ground Internet shows self-similarity, and this characteristic still exists after entering the satellite link. The characteristics of self-similar services in satellite networks include long-term data bursts. Data bursts will instantly occupy a large amount of bandwidth, resulting in increased network congestion and delay, and cache overflow, increasing the risk of data loss, resulting in unstable service quality, and affecting the real-time transmission performance of satellite networks.

[0003] The self-similarity of traffic describes the similarity of data sequences at different time scales. This property makes traffic sequences have persistent correlation in statistical characteristics, which can be used for prediction. Effective traffic prediction can accurately predict the trend of traffic changes and predict network traffic emergencies in advance, which helps to prevent network congestion in advance, optimize data transmission paths, and improve the stability and efficiency of network operation. Therefore, it is crucial to establish an accurate and efficient traffic prediction model.

[0004] At present, the research on network traffic prediction models can be summarized into four aspects: linear prediction models, nonlinear prediction models, neural network models, and decomposition-based prediction models.

[0005] Typical linear prediction models include the Auto Regressive Moving Average (ARMA) model, the Auto Regressive Integrated Moving Average (ARIMA) model, the Fractional Auto Regressive Integrated Moving Average (FARIMA) model, etc. The linear prediction model is based on the short-correlated Poisson process and cannot fully reflect the complex changing characteristics of satellite network traffic.

[0006] Nonlinear prediction models mainly include grey model, support vector machine (SVM) and least square support vector machine (LSSVM). Grey model has high prediction accuracy and simple model, but a slight change in the system may lead to large errors in the prediction results; SVM and LSSVM avoid the problems of local minimum and over-learning, but the key parameters are difficult to determine. In summary, nonlinear prediction models are not flexible and accurate enough to deal with the complexity of satellite network traffic.

[0007] With the development of machine learning, many intelligent neural network models have been proposed. These models can better characterize the nature of network traffic and have greatly improved prediction performance. Existing research uses quantum genetic algorithms to optimize traditional fuzzy neural networks, thereby achieving more accurate predictions of traffic flow, but the generalization ability of this method on different data sets is limited. In addition, long short-term memory neural networks show good prediction results when processing complex data, but their prediction accuracy still needs to be further improved.

[0008] In view of the shortcomings of a single model, more and more scholars have begun to explore prediction models based on decomposition. The accuracy of building demand load prediction has been improved by combining empirical mode decomposition with Bayesian optimization in a deep neural network, but the number of decomposed components has a significant impact on model performance and computational complexity. In terms of combined traffic prediction, some studies use decomposition algorithms to decompose network traffic into multiple components, make independent predictions based on the characteristics of each component, and merge the final results. However, when using decomposition algorithms, modal aliasing may occur, which affects the prediction effect. To solve this problem, researchers proposed a secondary decomposition method, which combines empirical mode decomposition, variational mode decomposition and a support vector regression model optimized by the whale algorithm to effectively handle power load prediction with nonlinear characteristics, but this also increases the complexity of the prediction. In the field of bridge health monitoring, a long short-term memory neural network model based on empirical mode decomposition, variational mode decomposition and optimization has been applied to bridge deflection prediction. Although this method performs well in decomposition accuracy and prediction accuracy, due to the large number of modal components, an independent prediction model needs to be built for each modal component, which increases the computational burden and the difficulty of data integration. Therefore, current research still faces challenges in terms of model complexity and computational efficiency. Summary of the invention

[0009] The technical problem to be solved by the present invention is to provide a satellite network self-similar traffic prediction method based on double decomposition and neural network in view of the above-mentioned deficiencies in the prior art, so as to realize the prediction of satellite network self-similar traffic.

[0010] In order to solve the above technical problems, the technical solution adopted by the present invention is: a satellite network self-similar traffic prediction method based on double decomposition and neural network, comprising the following steps:

[0011] Step 1: Use the adaptive noise complete set empirical mode decomposition (CEEMDAN) algorithm to perform the first modal decomposition on the original self-similar traffic sequence of the satellite network, and decompose the original self-similar traffic data sequence into multiple IMF components;

[0012] On the basis of the original self-similar traffic data sequence x(t), an independent Gaussian white noise sequence is added to obtain the noised sequence x i (t), as shown in formula (1):

[0013] x i (t) = x(t) + q0ε i (t) (1)

[0014] In the formula, q0 is the intensity of the noise added, ε i (t) is the i-th Gaussian white noise added, t is time, i = {1, 2, ..., N}, N is the number of Gaussian white noise added;

[0015] The CEEMDAN algorithm is used to perform noise on the sequence x i (t) Perform EMD decomposition to obtain the average IMF components of each order, among which the p-th order IMF component c decomposed by CEEMDAN is p (t) As shown in formula (2), the pth residual sequence r p (t) is shown in formula (3):

[0016]

[0017] r p (t) = x i (t)-c p (t) (3)

[0018] in, is the i′th IMF component of the pth order, and I is the number of the pth order IMF components;

[0019] When the residual sequence r p (t) When the extreme point is greater than the set value, it is considered that the conditions for continued decomposition are met and white noise needs to be added again to perform the next order decomposition until r p When the extreme value point (t) is equal to the set value, the final IMF component of the original self-similar flow data sequence x(t) is obtained;

[0020] Step 2: By calculating the sample entropy of each IMF component after decomposition, the complexity of each IMF component is evaluated, and the K-means clustering algorithm is used to divide these IMF components into three categories: high-frequency components, medium-frequency components, and low-frequency components;

[0021] First, the complexity of each IMF component is calculated by the sample entropy SE. The calculation of sample entropy SE is shown in formula (4):

[0022] SE(m,r,L)=-lnA m (r) / B m (r) (4)

[0023] Where L is the length of each IMF component sequence, m is the dimension, r is the similarity tolerance, and B m (r), A m (r) are the probability of matching m points and the probability of matching m+1 points for two consecutive subsequences in the IMF component sequence under the same similarity tolerance;

[0024] Then, according to the entropy value of each sample, all IMF components are clustered into three IMF components through K-means: high-frequency component Co-IMF1, medium-frequency component Co-IMF2 and low-frequency component Co-IMF3;

[0025] Step 3: Use the VMD decomposition model to perform a second modal decomposition on the clustered high-frequency components;

[0026] Assume that the high-frequency IMF component Co-IMF1 is represented by f(t), and decompose it into k modal components m k (t), as shown in formula (5):

[0027] m k (t) = A k (t)cos[φ k (t)] (5)

[0028] In the formula, A k (t) is the amplitude of each IMF component, φ k (t) is the phase function;

[0029] The high-frequency IMF component f(t) is calculated and analyzed by Hilbert transform, the spectrum is constructed, and its center frequency is estimated. The resulting constrained variational model expression is shown in formula (6):

[0030]

[0031] Where δ(t) is the pulse function, {m k}={m1,m2,......,m k} is a set of k modal components, {ωk}={ω1,ω2,......,ω k} is the frequency center corresponding to the k modal components, To find the sign of the partial derivative;

[0032] The alternating minimization method is used to solve equation (6), and for each modal component m k (t) Update and calculate the spectral characteristics of the modal components and adjust the center frequency; then perform iterative optimization and output the final modal components when the convergence conditions are reached. That is, the high-frequency component Co-IMF1 obtained by the CEEMDAD decomposition model is decomposed into n components - Final-IMF1, Final-IMF2, ..., Final-IMFn;

[0033] The n components Final-IMF1, Final-IMF2, ..., Final-IMFn obtained by the second modal decomposition and the intermediate frequency component Co-IMF2 and low frequency component Co-IMF3 obtained by the first modal decomposition form a new IMF component set;

[0034] Step 4: Build a self-similar traffic prediction model based on CNN-BiLSTM, input each IMF component after the two decompositions into the self-similar traffic prediction model for prediction, and generate the final self-similar predicted traffic;

[0035] Use one-dimensional CNN to learn the local trend of the IMF component data sequence obtained by the two decompositions, extract the local spatial features through CNN, and input the obtained feature vector into the BiLSTM layer for prediction;

[0036] BiLSTM consists of two LSTMs, which process forward and reverse sequences respectively. In the forward LSTM, the feature vectors extracted by the one-dimensional CNN are entered in chronological order, and in the reverse LSTM, the feature vectors extracted by the one-dimensional CNN are entered in the reverse order of chronological order. The outputs of the forward and reverse LSTMs are fused together by splicing to form the final output result.

[0037] Step 5: Use the improved artificial rabbit optimization algorithm IARO to optimize the hyperparameters of the self-similar traffic prediction model based on CNN-BiLSTM to obtain the optimal self-similar traffic prediction model;

[0038] The IARO algorithm is used to adjust the three key hyperparameters of the number of hidden units, learning rate and maximum number of training iterations of the self-similar flow prediction model, and a wider parameter space is explored to find the best combination of hyperparameters suitable for the self-similar flow prediction model. In this process, the mean square error MSE is selected as the fitness function to reflect the difference between the model prediction value and the actual value. At the same time, the IARO optimization algorithm is improved, and the concept of predators is introduced to simulate the interaction between species in the ecosystem, increase the diversity of the search for the optimal hyperparameters, and enhance the ability to avoid local optimal solutions.

[0039] Predators play a regulatory role in the algorithm, promoting the survival of the fittest mechanism of the rabbit population, enhancing the global search ability of the model, and enabling it to efficiently identify and optimize hyperparameter combinations. Under the influence of predators, the rabbit population faces two possible reactions: on the one hand, the rabbits may adopt an escape strategy to try to get rid of the threat of predators; on the other hand, the rabbits may also encounter predators and fall into a situation where they cannot escape. In the entire algorithm, the energy factor is used to simulate the changes in the ability of rabbits in the survival process. The designed energy factor E is shown in formula (7):

[0040]

[0041] Where T is the maximum number of iterations, t is the current number of iterations, α is the balance factor, which is used to adjust the dynamic changes of the rabbit between the maximum number of iterations T and the current number of iterations t, reflecting that the survival ability of the rabbit gradually decreases with the increase of iterations; P∈(0,1) is the probability of the predator rabbit escaping; β is the predation factor, which determines the impact of the threat level of the predator on the rabbit's energy; if β is large, the threat of the predator will significantly affect the rabbit's energy state, prompting it to choose to hide more frequently; if β is small, the impact of the predator is weak, and the rabbit chooses an active detour foraging strategy when it has sufficient energy.

[0042] The beneficial effect of adopting the above technical solution is that the satellite network self-similar traffic prediction method based on double decomposition and neural network provided by the present invention proposes a self-similar traffic prediction model based on double decomposition and convolutional bidirectional long short-term memory network in view of the strict requirements of satellite network for efficient transmission performance, aiming to meet the needs of real-time prediction and optimize system performance management. The model first uses the double decomposition algorithm to process the self-similar traffic data, enhances the model's ability to analyze different frequency components, and provides more accurate basic data support. Through double decomposition, high-frequency, medium-frequency and low-frequency components are effectively separated, and combined with convolutional bidirectional long short-term memory network for prediction, the model not only captures the long-term dependence in the traffic data, but also can sensitively identify complex dynamic changes. The results show that the prediction model proposed by the present invention can effectively improve the prediction accuracy and reduce the prediction error, thereby effectively supporting the requirements of self-similar and long-correlated business characteristics in satellite networks for data transmission, and meeting the space communication needs of large-scale and diversified multimedia data in different business scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 A flowchart of a satellite network self-similar traffic prediction method based on dual decomposition and neural network provided by an embodiment of the present invention;

[0044] Figure 2 A structural diagram of a self-similar traffic prediction model based on CNN-BiLSTM provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0046] Decomposing the original self-similar traffic data of the complex satellite network into simpler and easier to process parts can improve the accuracy and efficiency of traffic prediction. However, the commonly used traffic decomposition model has limited processing capabilities for the nonlinear and non-stationary characteristics of the data, making it difficult to learn all the fluctuation information in time, and the subsequent prediction results are poor. To this end, the present invention establishes a self-similar traffic double decomposition algorithm based on CEEMDAN-VMD. First, the original self-similar traffic data sequence is decomposed into multiple modal components based on the CEEMDAN algorithm; since the high-frequency components after decomposition are still highly complex and irregular, the effect of direct prediction is not ideal, so the highly complex sequences are further decomposed based on the VMD decomposition model to ensure the accuracy of subsequent predictions. The double decomposition algorithm established by the present invention effectively balances the complexity of self-similar traffic, and the secondary decomposition can more accurately capture the subtle changes and rapid fluctuations of self-similar traffic to meet the needs of sudden and rapid changes in network traffic data.

[0047] In this embodiment, a satellite network self-similar traffic prediction method based on double decomposition and neural network is used. Figure 1 As shown, the following steps are included:

[0048] Step 1: Use the adaptive noise complete set empirical mode decomposition (CEEMDAN) algorithm to perform the first modal decomposition on the original self-similar traffic sequence of the satellite network, and decompose the original self-similar traffic data sequence into multiple IMF components;

[0049] The CEEMDAN decomposition algorithm is based on empirical mode decomposition (EMD), which quantifies the relative strength between the original traffic sequence and the added Gaussian white noise, thereby adjusting the extraction order and weight of the mode during the decomposition process to maximize the retention of the original information and suppress the influence of noise. At the same time, the CEEMDAN algorithm dynamically generates modes based on the local characteristics of the signal. This adaptability makes it suitable for the decomposition of self-similar traffic. Finally, the complex time series data is decomposed into multiple intrinsic mode functions (IMF) with different time scale fluctuation information. The specific decomposition process is as follows:

[0050] On the basis of the original self-similar traffic data sequence x(t), an independent Gaussian white noise sequence is added to obtain the noised sequence x i (t), as shown in formula (1):

[0051] x i (t) = x(t) + q0ε i (t) (1)

[0052] In the formula, q0 is the intensity of the noise added, ε i (t) is the i-th Gaussian white noise added, i = {1, 2, ..., I}, t is time, I is the number of Gaussian white noise added;

[0053] The CEEMDAN algorithm is used to perform noise on the sequence x i (t) Perform EMD decomposition to obtain the average IMF (Intrinsic Mode Function) components of each order, among which the p-th order IMF component c decomposed by CEEMDAN p (t) As shown in formula (2), the pth residual sequence r p (t) is shown in formula (3):

[0054]

[0055] r p (t) = x i(t)-c p (t) (3)

[0056] in, is the i′th IMF component of the pth order, and I is the number of the pth order IMF components;

[0057] In this embodiment, when the residual sequence r p (t) When the extreme value is greater than 2, it is considered that the conditions for continued decomposition are met and white noise needs to be added again to perform the next order decomposition until r p When the extreme value of (t) is equal to 2, the final IMF component of the original self-similar traffic data sequence x(t) is obtained;

[0058] Step 2: By calculating the sample entropy of each IMF component after decomposition, the complexity of each IMF component is evaluated, and the K-means clustering algorithm is used to divide these IMF components into three categories: high-frequency components, medium-frequency components, and low-frequency components;

[0059] First, the complexity of each IMF component is calculated by using the sample entropy (SE). The calculation of sample entropy is shown in formula (4):

[0060] SE(m,r,L)=-lnA m (r) / B m (r) (4)

[0061] Where L is the length of each IMF component sequence, m is the dimension, r is the similarity tolerance, and B m (r), A m (r) are the probability of matching m points and the probability of matching m+1 points for two consecutive subsequences in the IMF component sequence under the same similarity tolerance;

[0062] Then, according to the entropy value of each sample, all IMF components are clustered into three IMF components through K-means: high-frequency component Co-IMF1, medium-frequency component Co-IMF2 and low-frequency component Co-IMF3;

[0063] In this embodiment, the relationship between the sample entropy value SE of each IMF component and the K-means clustering result is shown in Table 1.

[0064] Table 1 Sample entropy SE and K-means clustering results

[0065] SE Range Clustering results [1，2] High frequency component Co-IMF1 [0.5，1] Intermediate frequency component Co-IMF2 [0，0.5] Low frequency component Co-IMF3

[0066] Step 3: Use the VMD decomposition model (Variational mode decomposition) to perform a second modal decomposition on the clustered high-frequency components;

[0067] The high-frequency IMF component Co-IMF1 obtained by the decomposition based on the CEEMDAD algorithm is subjected to secondary modal decomposition to reduce the complexity of the high-frequency IMF component. The VMD decomposition model can determine the center frequency of each component by iteratively finding the optimal solution of the variational model. It can be used to decompose complex signals into several intrinsic mode functions, each of which has local characteristics in a specific frequency band. Based on the VMD decomposition model, the complex high-frequency component after the first decomposition is subjected to secondary decomposition, and it is accurately decomposed into multiple modal functions with specific frequency bands. This precise frequency decomposition can identify the periodicity and burstiness in self-similar traffic data, and provide more meaningful input for subsequent prediction models. The specific process of secondary decomposition based on the VMD decomposition model is as follows:

[0068] Assume that the high-frequency IMF component Co-IMF1 is represented by f(t), and decompose it into k modal components m k (t), as shown in formula (5):

[0069] m k (t) = A k (t)cos[φ k (t)] (5)

[0070] In the formula, A k (t) is the amplitude of each IMF component, φ k (t) is the phase function;

[0071] The high-frequency IMF component f(t) is calculated and analyzed by Hilbert transform, the spectrum is constructed, and its center frequency is estimated. The resulting constrained variational model expression is shown in formula (6):

[0072]

[0073] Where δ(t) is the pulse function, {m k}={m1,m2,......,m k} is a set of k modal components, {ω k}={ω1,ω2,......,ω k} is the frequency center corresponding to the k modal components, To find the sign of the partial derivative;

[0074] The alternating minimization method is used to solve equation (6), and for each modal component m k (t) Update and calculate the spectral characteristics of the modal components, adjust the center frequency to better match the spectrum; then perform iterative optimization, and when the convergence condition is reached, output the final modal components That is, the high-frequency component Co-IMF1 obtained by the CEEMDAD decomposition model is decomposed into n components - Final-IMF1, Final-IMF2, ..., Final-IMFn;

[0075] The n components Final-IMF1, Final-IMF2, ..., Final-IMFn obtained by the second modal decomposition and the intermediate frequency component Co-IMF2 and low frequency component Co-IMF3 obtained by the first modal decomposition form a new IMF component set;

[0076] The double decomposition process of step 2-step 3 of the present invention can be called a double decomposition algorithm based on CEEMDAN-VMD. This algorithm can not only overcome the problems of modal confusion and residual noise, but also double decomposition can weaken the concentration of high-frequency components, thereby improving the prediction accuracy. The double decomposition implementation process based on CEEMDAN-VMD is shown in Table 2.

[0077] Table 2 Dual decomposition algorithm based on CEEMDAN-VMD

[0078]

[0079]

[0080] Step 4: Build a self-similar traffic prediction model based on CNN-BiLSTM, input each IMF component after the two decompositions into the self-similar traffic prediction model for prediction, and generate the final self-similar predicted traffic;

[0081] The traffic data of satellite networks often have complex time dependence and self-similarity, and it is necessary to establish an accurate and efficient satellite network self-similar traffic prediction model. Convolutional Neural Network (CNN) can effectively extract spatially locally related features from time series data, while Bi-directional Long Short-term Memory (BiLSTM) can effectively capture long-term dependencies in time series data. In view of the advantages of the above two, the present invention constructs a self-similar traffic prediction model based on CNN and BiLSTM, trains and predicts each frequency component obtained by double decomposition, and can cope with complex traffic change patterns and irregular signal fluctuations, thereby providing more accurate and efficient traffic prediction and improving the overall performance of the network.

[0082] One-dimensional CNN is used to learn the local trend of the IMF component data sequence obtained by the two decompositions. The local spatial features are extracted through CNN, and the obtained feature vector is input into the following BiLSTM layer for prediction. The self-similar traffic prediction model structure based on CNN-BiLSTM is as follows: Figure 2 shown.

[0083] The three cascade layers of CNN are described as shown in equations (7), (8) and (9):

[0084]

[0085]

[0086] Formula (7) is the convolution operation in the CNN network, where and Represent the input and output of the convolutional layer respectively, is the parameter value in the channel, is the bias parameter; Formula (8) is the activation function, where and They represent the input and output of the activation layer respectively; Formula (9) is the pooling function, where and is the input and output of the pooling operation;

[0087] BiLSTM consists of two LSTMs (Long Short-Term Memory), which process forward and reverse sequences respectively. In the forward LSTM, the feature vectors extracted by the one-dimensional CNN enter in chronological order, and in the reverse LSTM, the feature vectors extracted by the one-dimensional CNN enter in the reverse order of chronological order. The outputs of the forward and reverse LSTMs are fused together by splicing to form the final output result. This bidirectional loop structure can better utilize the local and global information in the time series, improve the feature extraction ability of the model, and thus improve the accuracy of the prediction model. The mathematical model of the BiLSTM neural network is shown in equations (10), (11), and (12):

[0088] p t =f(ω1x t +ω2p t-1 ) (10)

[0089] r t =f(ω3x t +ω5r t+1 ) (11)

[0090] y t =g(ω4p t +ω6r t ) (12)

[0091] In formula (10), formula (11) and formula (12), x t represents the input vector at time t; p t and r t Respectively represent the output vectors of the forward layer and the reverse layer at time t; y t Represents the output vector of the output layer at time t; f is the activation function of the forward layer and the reverse layer; g represents the activation function of the output layer; ω1 is the weight from the input layer to the forward layer; ω3 is the weight from the input layer to the reverse layer; ω2 is the weight from the output of the forward layer at the previous moment to the current moment; ω5 is the weight from the output of the reverse layer at the previous moment to the current moment; ω4 is the weight from the output of the forward layer to the output layer; ω6 is the weight from the output of the reverse layer to the output layer.

[0092] The self-similar traffic prediction based on CNN-BiLSTM is as follows: After the self-similar traffic is double-decomposed, the correlation between traffic sequences and effective nonlinear local features are captured, the feature information is trained and learned, and the prediction results are output through the model. This algorithm can not only handle complex traffic fluctuations and nonlinear changes, but also improve the prediction accuracy and efficiency, and has strong generalization ability and robustness, so as to effectively predict the future trend of self-similar traffic. The implementation process of the traffic prediction algorithm based on CNN-BiLSTM is shown in Table 3.

[0093] Table 3 Self-similar traffic prediction based on CNN-BiLSTM

[0094]

[0095] The CEEMDAN-VMD dual decomposition / CNN-BiLSTM prediction algorithm proposed in this paper not only performs well in prediction accuracy, but also has high prediction efficiency, and can maintain a short average prediction time in the complex self-similar traffic prediction process. Therefore, the algorithm achieves a good balance between accuracy and efficiency and has a wide range of application potential.

[0096] Step 5: Use the improved artificial rabbit optimization algorithm (IARO) to optimize the hyperparameters of the self-similar traffic prediction model based on CNN-BiLSTM to obtain the optimal self-similar traffic prediction model;

[0097] By adjusting the three key hyperparameters of the number of hidden units, learning rate, and maximum number of training iterations, the IARO algorithm can explore a wider parameter space to find the best combination of hyperparameters suitable for the self-similar traffic prediction model; in this process, the mean square error (MSE) is selected as the fitness function to reflect the difference between the model prediction value and the actual value. In addition, the evaluation indicators of the model include the root mean square error (RMSE) and the goodness of fit (R-Squared, R2), both of which can not only provide an intuitive assessment of the model's prediction accuracy, but also help judge the model's explanatory power. In the improved IARO optimization algorithm, the concept of predators is introduced to simulate the interactions between species in the ecosystem, which increases the diversity of the search for the optimal hyperparameters and enhances the ability to avoid local optimal solutions.

[0098] In a relatively safe environment, the rabbit adopts a detour foraging strategy and actively explores new hyperparameter combinations. The rabbit will select potential high-quality areas for search based on the current energy level and environmental information. At this time, the rabbit shows higher flexibility and exploration ability, and can effectively find the optimal combination of hyperparameters. Its mathematical model is shown in equations (13)-(17):

[0099]

[0100] R=L·c (14)

[0101]

[0102] n1~N(0,1)(17)

[0103] in, is the candidate position of the i-th rabbit at the t+1th iteration, i=1,...,n, n is the population size, is the current position of the i-th rabbit at the t-th iteration, is the distance vector moved by the rabbit, r1, r2, r3 and r4 are random numbers between [0,1], n1 is a random number that follows the standard normal distribution, round means rounding, R means the running operator, which is used to simulate the running characteristics of the rabbit, L is the speed when foraging, T is the maximum number of iterations, c is a mapping function, which takes a value of 1 or 0, which helps the algorithm to randomly select search individual elements for mutation in foraging behavior, d is the dimension of the search space, is a rounding function, g(l) is a random selector, and one dimension is selected for key operation in each iteration. This selection method allows the algorithm to perturb in different dimensions, so that the rabbit (i.e., the solution) can flexibly explore various parts of the search space and enhance the global search capability.

[0104] When rabbits sense the threat of predators, they will randomly choose to hide according to the danger level of the current environment. At this time, rabbits will reduce their exploration activities and remain still to reduce the risk of being preyed on. The mathematical model is shown in the formula:

[0105]

[0106]

[0107] n2~N(0,1)(20)

[0108]

[0109] in, represents the hiding cave location randomly selected from d caves, H represents the hidden parameter, r4 is a random number between [0,1], n2 is a random number that obeys the standard normal distribution, and g(k) is the mapping function.

[0110] At the same time, predators play a regulatory role in the algorithm, promoting the survival of the fittest mechanism of the rabbit population, enhancing the global search ability of the model, and enabling it to efficiently identify and optimize hyperparameter combinations. Under the influence of predators, the rabbit population faces two possible reactions: on the one hand, the rabbits may adopt an escape strategy to try to get rid of the threat of predators; on the other hand, the rabbits may encounter predators and fall into a situation where they cannot escape. This complex biological interaction not only reflects the competitive relationship between species in the ecosystem, but also helps to improve the performance of the algorithm in complex search spaces. By introducing the element of predators, the algorithm forms a more dynamic balance between detour foraging and random hiding, thereby enhancing its adaptability in complex search spaces. This mechanism not only promotes the maintenance of diversity, but also improves the effectiveness and robustness of the algorithm in practical applications, enabling it to better cope with various optimization challenges.

[0111] In the whole mechanism, the energy factor is used to simulate the change of the rabbit's ability during the survival process. The designed energy factor E is shown in formula (22):

[0112]

[0113] Where T is the maximum number of iterations, t is the current number of iterations, α is the balance factor, which is used to adjust the dynamic changes of the rabbit between the maximum number of iterations T and the current number of iterations t, reflecting that the survival ability of the rabbit gradually decreases with the increase of iterations; P∈(0,1) is the probability of the predator rabbit escaping. β is the predation factor, which determines the impact of the threat level of the predator on the rabbit's energy. If β is large, the threat of the predator will significantly affect the rabbit's energy state, prompting it to choose to hide more frequently; if β is small, the impact of the predator is weak, and the rabbit chooses an active detour foraging strategy when it is full of energy.

[0114] When the value of the energy factor is larger, it means that the rabbit has more energy, has stronger physical strength and ability to forage for food, and actively explores better hyperparameter combinations; when the value of the energy factor is smaller, the rabbit has less energy and cannot actively explore, and can only adopt a random hiding strategy, which limits its search space. This mechanism enables the IARO algorithm to achieve higher efficiency and accuracy when searching for the optimal hyperparameter combination of the CNN-BiLSTM model, providing strong support for the self-similar traffic prediction task.

[0115] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. A satellite network self-similar traffic prediction method based on double decomposition and neural network, characterized by: The following steps are involved: Step 1: Use the adaptive noise complete set empirical mode decomposition (CEEMDAN) algorithm to perform the first modal decomposition on the original self-similar traffic sequence of the satellite network, and decompose the original self-similar traffic data sequence into multiple IMF components; Step 2: By calculating the sample entropy of each IMF component after decomposition, the complexity of each IMF component is evaluated, and the K-means clustering algorithm is used to divide these IMF components into three categories: high-frequency components, medium-frequency components, and low-frequency components; Step 3: Use the VMD decomposition model to perform a second modal decomposition on the clustered high-frequency components to obtain n IMF components; Step 4: Build a self-similar traffic prediction model based on CNN-BiLSTM, input each IMF component after the two decompositions into the self-similar traffic prediction model for prediction, and generate the final self-similar predicted traffic; Step 5: Use the improved artificial rabbit optimization algorithm IARO to optimize the hyperparameters of the self-similar traffic prediction model based on CNN-BiLSTM to obtain the optimal self-similar traffic prediction model.

2. The satellite network self-similar traffic prediction method based on double decomposition and neural network according to claim 1 is characterized by: The specific method of step 1 is: On the basis of the original self-similar traffic data sequence x(t), an independent Gaussian white noise sequence is added to obtain the noised sequence x i (t), as shown in formula (1): x i (t)=x(t)+q0ε i (t) (1) In the formula, q0 is the intensity of the noise added, ε i (t) is the i-th Gaussian white noise added, t is time, i = {1, 2, ..., N}, N is the number of Gaussian white noise added; The CEEMDAN algorithm is used to perform noise on the sequence x i (t) Perform EMD decomposition to obtain the average IMF components of each order, among which the p-th order IMF component c decomposed by CEEMDAN is p (t) As shown in formula (2), the pth residual sequence r p (t) is shown in formula (3): r p (t)=x i (t)-c p (t) (3) in, is the i′th IMF component of the pth order, and I is the number of the pth order IMF components; When the residual sequence r p (t) When the extreme point is greater than the set value, it is considered that the conditions for continued decomposition are met and white noise needs to be added again to perform the next order decomposition until r p When the extreme value point (t) is equal to the set value, the final IMF component of the original self-similar traffic data sequence x(t) is obtained.

3. The satellite network self-similar traffic prediction method based on double decomposition and neural network according to claim 2 is characterized by: The specific method of step 2 is: First, the complexity of each IMF component is calculated by the sample entropy SE. The calculation of sample entropy SE is shown in formula (4): SE(m,r,L)=-lnA m (r) / B m (r) (4) Where L is the length of each IMF component sequence, m is the dimension, r is the similarity tolerance, and B m (r), A m (r) are the probability of matching m points and the probability of matching m+1 points for two consecutive subsequences in the IMF component sequence under the same similarity tolerance; Then, according to the size of the entropy value of each sample, all IMF components are clustered into three IMF components through K-means: high-frequency component Co-IMF1, medium-frequency component Co-IMF2 and low-frequency component Co-IMF3.

4. The satellite network self-similar traffic prediction method based on double decomposition and neural network according to claim 3 is characterized by: The specific method of step 3 is: Assume that the high-frequency IMF component Co-IMF1 is represented by f(t), and decompose it into k modal components m k (t), as shown in formula (5): m k (t)=A k (t)cos[φ k (t)] (5) In the formula, A k (t) is the amplitude of each IMF component, φ k (t) is the phase function; The high-frequency IMF component f(t) is calculated and analyzed by Hilbert transform, the spectrum is constructed, and its center frequency is estimated. The resulting constrained variational model expression is shown in formula (6): Where δ(t) is the pulse function, {m k }={m1,m2,......,m k } is a set of k modal components, {ω k }={ω1,ω2,......,ω k } is the frequency center corresponding to the k modal components, To find the sign of the partial derivative; The alternating minimization method is used to solve equation (6), and for each modal component m k (t) updating, calculating the frequency spectrum characteristics of the modal components, and adjusting the center frequency; After that, iterative optimization is performed, and when the convergence condition is reached, the final modal component is output. That is, the high-frequency component Co-IMF1 obtained by the CEEMDAD decomposition model is decomposed into n components - Final-IMF1, Final-IMF2, ..., Final-IMFn; The n components Final-IMF1, Final-IMF2, ..., Final-IMFn obtained by the second modal decomposition and the intermediate frequency component Co-IMF2 and low frequency component Co-IMF3 obtained by the first modal decomposition form a new IMF component set.

5. The satellite network self-similar traffic prediction method based on double decomposition and neural network according to claim 4 is characterized by: In step 4, one-dimensional CNN is used to perform local trend learning on the IMF component data sequence obtained by the two decompositions, local spatial features are extracted through CNN, and the obtained feature vector is input into the BiLSTM layer for prediction; BiLSTM consists of two LSTMs, which process forward and reverse sequences respectively. In the forward LSTM, the feature vectors extracted by the one-dimensional CNN enter in chronological order, and in the reverse LSTM, the feature vectors extracted by the one-dimensional CNN enter in the reverse order of chronological order. The outputs of the forward and reverse LSTMs are fused together by splicing to form the final output result.

6. The satellite network self-similar traffic prediction method based on double decomposition and neural network according to claim 5 is characterized by: The step 5 adopts the IARO algorithm to adjust the three key hyperparameters of the number of hidden units, the learning rate and the maximum number of training iterations of the self-similar traffic prediction model, and explores a wider parameter space to find the best combination of hyperparameters suitable for the self-similar traffic prediction model; in this process, the mean square error MSE is selected as the fitness function to reflect the difference between the model prediction value and the actual value; at the same time, the IARO optimization algorithm is improved, and the concept of predators is introduced to simulate the interaction between species in the ecosystem, increase the diversity of the search for the optimal hyperparameters, and enhance the ability to avoid local optimal solutions; Predators play a regulatory role in the algorithm, promoting the survival of the fittest mechanism of the rabbit population, enhancing the global search ability of the model, and enabling it to efficiently identify and optimize hyperparameter combinations. Under the influence of predators, the rabbit population faces two possible reactions: on the one hand, the rabbits may adopt an escape strategy to try to get rid of the threat of predators; on the other hand, the rabbits may also encounter predators and fall into a situation where they cannot escape. In the entire algorithm, the energy factor is used to simulate the changes in the ability of rabbits in the survival process. The designed energy factor E is shown in formula (7): Where T is the maximum number of iterations, t is the current number of iterations, α is the balance factor, which is used to adjust the dynamic changes of the rabbit between the maximum number of iterations T and the current number of iterations t, reflecting that the survival ability of the rabbit gradually decreases with the increase of iterations; P∈(0,1) is the probability of the predator rabbit escaping; β is the predation factor, which determines the impact of the threat level of the predator on the rabbit's energy.

Images