CN117060984B - Satellite network flow prediction method based on empirical mode decomposition and BP neural network - Google Patents

Abstract

The invention relates to the technical field of satellite communication, and particularly discloses a satellite network flow prediction method based on empirical mode decomposition and BP neural network, which comprises the following steps: s10, generating a time sequence according to an ON/OFF source model meeting Pareto distribution; s20, decomposing the time sequence into multi-order inclusion modal components with short-range correlation by adopting empirical mode decomposition; s30, respectively establishing a prediction model for each order of connotation modal components by adopting an autoregressive moving average model; s40, predicting the connotation modal components by adopting each-order prediction model, and summarizing to obtain a linear characteristic sequence; s50, subtracting the linear characteristic sequence from the original flow sequence to obtain a residual sequence of the satellite network flow; s60, optimizing the BP neural network by adopting a wolf algorithm, and constructing an optimization model; s70, predicting the residual sequence according to the optimization model to obtain a residual sequence prediction result; s80, determining a satellite network flow prediction result according to the linear characteristic sequence and the residual sequence prediction result.

Description

Satellite network traffic prediction method based on empirical mode decomposition and BP neural network

Technical field

The invention relates to the technical field of satellite communication, and specifically relates to a satellite network traffic prediction method based on empirical mode decomposition and BP neural network.

Background technique

The low-orbit satellite communication network has the characteristics of wide coverage and large communication capacity. It can break through the restrictions of geographical conditions with the help of inter-satellite networking, achieve uninterrupted signal coverage, and provide global users with large broadband, low latency, and seamless network. Access service.

In recent years, with the construction of giant low-orbit satellite constellations, satellite network traffic demand has shown a rapid upward trend, and network bandwidth resources have become increasingly tight. The on-board computing resources and storage resources of low-orbit satellites are limited by the power consumption and volume of the satellite platform, and the bandwidth they can provide is relatively limited. At the same time, with the continuous switching of inter-satellite links (ISL) between satellites, the network topology of low-orbit satellites has also changed. The traffic of satellite networks changes drastically over time, and the traffic of satellite networks shows non-uniformity in time and space. As user scale and traffic demands continue to expand, the above factors make low-orbit satellite networks more prone to network congestion, affecting network service quality. Through traffic prediction, the changing characteristics and trends of network traffic can be predicted in advance, and network traffic control can be transformed from a passive response mode to an active sensing mode.

Traditional network traffic models are generally based on the Poisson process, including Poisson models, Markov models, Markov models, etc. These models can only describe the short correlation of traffic in the time domain. For satellite network traffic that exhibits long correlation processes, it is difficult for traditional models to accurately describe network characteristics. Since the self-similarity characteristics of network traffic were discovered in 1994, various traffic forecast models based on self-similarity have been continuously proposed. One is to describe the observed data characteristics by constructing physical models, including ON/OFF models of heavy-tail distribution, Queuing models, etc.; the other type is statistical models, which try to simulate the changing trend of network data through data fitting methods. The difference between these self-similar traffic forecast models and traditional forecast models is that the self-similar forecast model is based on network characteristics, can describe the LRD and burstiness of traffic, and helps to forecast based on the inherent laws of network traffic. The accuracy has been improved, but the calculation process is complex and time-consuming.

With the development of machine learning, neural network models, fuzzy theory, chaos theory, etc. have good nonlinear mapping capabilities and can better characterize the characteristics of network traffic and improve the performance of network traffic forecasting. However, the related algorithms have problems such as a large number of optional parameters and high computational complexity, and are not suitable for satellite network platforms. Since a single model can only describe the Poisson process or self-similar characteristics of network traffic and cannot well describe satellite network business traffic, many scholars have proposed hybrid models. The EMD or wavelet model is used to decompose the network traffic, and then the forecast model is applied to each obtained component to extract the self-similar characteristics of the network traffic step by step, which effectively reduces the computational complexity and improves the forecast accuracy. However, the reliability of the prediction results using different models for each component has not been fully demonstrated.

Contents of the invention

In response to the above problems, the purpose of the present invention is to provide a satellite network traffic prediction method based on empirical mode decomposition and BP neural network. The prediction accuracy is better than the traditional satellite network traffic prediction model, and has higher solution accuracy and lower calculation time. the complexity.

The present invention provides a satellite network traffic prediction method based on empirical mode decomposition and BP neural network, including:

Step S10: Generate a time series of satellite network traffic based on an ON/OFF source model that satisfies Pareto distribution; the ON/OFF source model that satisfies Pareto distribution is generated for end-to-end connection data in the satellite network Model;

Step S20: Use empirical mode decomposition to decompose the time series into multi-order connotative modal components with short-range correlation;

Step S30: Use an autoregressive moving average model to establish prediction models for the intrinsic modal components at each stage;

Step S40: Use the prediction model at each stage to predict the connotative modal components respectively, and summarize the prediction results to obtain a linear feature sequence of satellite network traffic;

Step S50: Subtract the linear feature sequence from the original traffic sequence of the satellite network traffic to obtain a residual sequence of the satellite network traffic;

Step S60: Use the gray wolf algorithm to optimize the BP neural network and build an optimization model;

Step S70: Predict the residual sequence according to the optimization model to obtain a residual sequence prediction result;

Step S80: Determine the satellite network traffic prediction result based on the linear feature sequence and the residual sequence prediction result.

In a possible implementation, in S10, the probability distribution function of the Pareto distribution is as follows:

;

in, is a positive integer, representing a random variable The minimum value that can be taken; determines the mean of the random variable and random variable variance ;like Then the mean of the Pareto distribution existence, variance The supreme realm.

In a possible implementation, the S20 includes:

Step S21, convert the data to be analyzed in the ON/OFF source model to All extreme points of are fitted with two cubic spline curves to obtain the data to be analyzed The extreme envelope of

Step S22, let the average value of the extreme value envelope be , then the remaining signal ;like If the IMF conditions are met, then is the first IMF component, otherwise it will as ;

Step S23, after k rounds of iterations, the difference between the signal obtained and the mean value of the envelope, the difference obtained in the k -1th round of iterations is ;When the following formula is established, then As the first intensional modal component:

;

in, is the threshold, T is the number of data samples, that is, after k rounds of iterations Compared with the results of the previous iteration The root mean square difference between , then As the first intensional modal component that satisfies the condition;

Step S24, will as , repeat the above steps; when the remaining residual is a monotonic function and the amplitude is less than the threshold, stop the calculation and obtain several IMF components , then the data to be analyzed is obtained through the following formula:

;

In the formula, is the final residual amount.

In a possible implementation, the S30 includes:

The predictive model is determined according to the following formula:

;

In the formula, is the model parameter of the autoregressive moving average model, that is, the sum of squares of the fitting residuals, are the independent error term, autoregressive order, difference order, and moving average order, respectively. is the intensional modal component, To fit the standard deviation of the intrinsic modal component sequence, AIC is the Akaike information content criterion.

In a possible implementation, the S50 includes:

The residual sequence is calculated as follows according to the following formula:

;

In the formula, is the residual sequence, is the original traffic sequence, is a linear feature sequence.

In a possible implementation, the S60 includes:

Step S61, initialize the algorithm parameters of the gray wolf algorithm and the parameters of the BP neural network;

Step S62, normalize the parameters to obtain normalized parameters, and divide the normalized parameters into a training data set and a test data set;

Step S63, train the BP neural network according to the training data set to obtain multiple training models;

Step S64: Evaluate the performance of the training model based on the fitness of individuals in the test data set;

Step S65: Classify the gray wolf population according to fitness, retain the position of the individual with the best fitness, and update the positions of the remaining individuals;

Step S66: If the number of iterations reaches the preset value, the individual position with the best fitness is used as the optimal parameter of the BP neural network model to build a prediction model.

In a possible implementation, the S62 includes:

Normalization is performed according to the following formula:

;

In the formula, is the normalization parameter, It’s the first original parameters, and are respectively the maximum and minimum values corresponding to the original parameters.

In a possible implementation, the S63 includes:

Train the BP neural network according to the following formula to obtain multiple training models;

;

In the formula, respectively and distance between for current location, is a random vector, is the optimal solution, Connect weights for each layer of the BP neural network.

In a possible implementation, the S64 includes:

The fitness of an individual is calculated based on the mean square error. The mean square error is calculated according to the following formula:

;

In the formula, T is the sequence length of satellite network traffic; for The tth value of the fitted residual sequence.

In a possible implementation, the S65 includes:

The location of the individual is updated according to the following formula:

;

;

In the formula, respectively and distance between respectively current location, for current location, is a random vector, is the number of iterations, are the positions of the three types of operators respectively.

The satellite network traffic prediction method based on empirical mode decomposition and BP neural network of the present invention decomposes satellite network traffic with self-similarity into multi-order IMF components with short-range correlation through empirical mode decomposition, and adopts adaptive ordering. The ARIMA improved by the optimization operator predicts the IMF component and reduces the computational complexity. At the same time, the improved gray wolf algorithm is used to optimize the BP neural network parameters, and the optimized BP neural network is used to forecast the residuals of the EMD-ARIMA model. Finally, the satellite network traffic forecast results are obtained. The forecast accuracy is better than the traditional satellite network traffic forecast model, and has better performance. High solution accuracy and low computational complexity.

Description of the drawings

Figure 1 is a schematic flow chart of a satellite network traffic prediction method provided by an embodiment of the present invention;

Figure 2 is a schematic diagram of a satellite network traffic prediction method provided by an embodiment of the present invention;

Figure 3 is a schematic structural diagram of a satellite network system provided by an embodiment of the present invention;

Figure 4 is a schematic diagram of an ON/OFF source superposition model that obeys Pareto distribution provided by an embodiment of the present invention;

Figure 5 is a satellite network traffic time series data diagram provided by an embodiment of the present invention;

Figure 6 is a schematic diagram of IMF components of each order obtained after decomposition according to an embodiment of the present invention;

Figure 7 is a schematic diagram of the satellite network traffic prediction results provided by the embodiment of the present invention.

Detailed ways

The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The detailed description of the following embodiments and the accompanying drawings are used to illustrate the principles of the present invention, but cannot be used to limit the scope of the present invention. That is, the present invention is not limited to the described preferred embodiments. The scope of the present invention is determined by the claims. limited.

In the description of the present invention, it should be noted that, unless otherwise stated, the meaning of "plurality" is two or more; the terms "first", "second", etc. are only used for descriptive purposes and cannot be understood. To indicate or imply relative importance; for those of ordinary skill in the art, the specific meanings of the above terms in the present invention may be understood based on specific circumstances.

Satellite network traffic prediction can provide key information for satellite network routing and resource allocation, and is of great significance to the efficient operation of satellite communication networks. However, satellite network traffic has self-similarity and long-range correlation, and traditional linear or nonlinear network traffic forecast models cannot achieve sufficient forecast accuracy. The prediction accuracy of the satellite network traffic prediction method based on empirical mode decomposition and BP neural network of the present invention is better than the traditional satellite network traffic prediction model, and has higher solution accuracy and lower computational complexity.

Figure 1 is a schematic flowchart of a satellite network traffic prediction method provided by an embodiment of the present invention. Figure 2 is a schematic diagram of a satellite network traffic prediction method provided by an embodiment of the present invention. As shown in Figures 1 and 2, the present invention provides a Satellite network traffic prediction method based on empirical mode decomposition and BP neural network, including:

Step S10: Generate a time series of satellite network traffic based on the ON/OFF source model that satisfies Pareto distribution;

Figure 3 is a schematic structural diagram of a satellite network system provided by an embodiment of the present invention. As shown in Figure 3, the satellite network to which the present invention is applicable is a low-orbit satellite network, including a space network and a ground network. Each satellite of the space network is connected to neighboring satellites in the same orbit and satellites in adjacent orbits to the left and right through four inter-satellite links. The space network is connected to the gateway station of the ground network through feeder links to form a satellite network.

Among them, the ON/OFF source model that satisfies Pareto distribution is the data generation model of end-to-end connections in the satellite network. The ON period is regarded as the size of the data packet that satisfies the Pareto distribution, and the OFF period is regarded as the interval between transmission of the data packet that satisfies the Pareto distribution. Pareto distribution is a classic heavy-tailed distribution. Think of the end-to-end connection in the satellite network as an ON/OFF source that satisfies Pareto distribution. ON corresponds to the sending of messages, the source node generates data packets at a constant rate, and the source nodes remain independent of each other; while OFF corresponds to the interruption of message sending, after accumulating a sufficient number of samples, they are also independent and identically distributed.

The probability distribution function of Pareto distribution is as follows:

;

in, is a positive integer, representing a random variable The minimum value that can be taken; determines the mean of the random variable and random variable variance ;like Then the mean of the Pareto distribution existence, variance The supreme realm.

Figure 4 is a schematic diagram of an ON/OFF source superposition model that obeys Pareto distribution provided by an embodiment of the present invention. Figure 5 is a time series data diagram of satellite network traffic provided by an embodiment of the present invention.

Step S20: Use empirical mode decomposition EMD to decompose the time series into multi-order intrinsic modal components with short-range correlation;

In a possible implementation, S20 includes:

Step S21, turn ON/OFF the data to be analyzed in the source model All extreme points of are fitted with two cubic spline curves to obtain the data to be analyzed The extreme envelope of

Step S22, let the average value of the extreme value envelope be , then the remaining signal ;like If the IMF conditions are met, then is the first IMF component, otherwise it will as ;

Step S23, after k rounds of iterations, the difference between the signal obtained and the mean value of the envelope, the difference obtained in the k -1th round of iterations is ;When the following formula is established, then As the first intensional modal component:

;

in, is the threshold, T is the number of data samples, that is, after k rounds of iterations Compared with the results of the previous iteration The root mean square difference between , then As the first intensional modal component that satisfies the condition;

Step S24, will as , repeat the above steps; when the remaining residual is a monotonic function and the amplitude is less than the threshold, stop the calculation and obtain several IMF components , then the data to be analyzed is obtained through the following formula:

;

In the formula, is the final residual amount.

Figure 6 is a schematic diagram of the IMF components of each order obtained after decomposition according to the embodiment of the present invention, a, b, c, d, e, f, g, h, i, and j are IMF1, IMF2, IMF3, IMF4, IMF5, IMF6, IMF7, IMF8, IMF9, and IMF10 respectively.

Step S30: Use the autoregressive moving average model to establish prediction models for the intrinsic modal components of each order;

In a possible implementation, S30 includes:

Determine the prediction model according to the following formula:

;

In the formula, is the model parameter of the autoregressive moving average model, that is, the sum of squares of the fitting residuals, are the independent error term, autoregressive order, difference order, and moving average order, respectively. is the intensional modal component, To fit the standard deviation of the intrinsic modal component sequence, AIC is the Akaike information content criterion.

For the n-order IMF components obtained by decomposition, the ARIMA model is used to establish the prediction model ARIMA (pi, di, qi) for each order component. Among them (pi, di, qi) are ARIMA model parameters, i = 1,2,…,n. According to Figure 6, a total of 10 orders of IMF are obtained after the original traffic data is decomposed. The optimal models of each order of IMF determined by the adaptive fixed-order optimization operator are shown in Table 1:

Table 1

Step S40: Use prediction models of each order to predict the connotative modal components respectively, and summarize the prediction results to obtain a linear feature sequence of satellite network traffic;

Step S50: Subtract the linear feature sequence from the original traffic sequence of the satellite network traffic to obtain the residual sequence of the satellite network traffic;

In a possible implementation, S50 includes:

The residual sequence is calculated as follows according to the following formula:

;

In the formula, is the residual sequence, is the original traffic sequence, is a linear feature sequence.

Step S60, use the gray wolf algorithm to optimize the BP neural network and build the optimization model IGWO-BPNN;

In a possible implementation, S60 includes: steps S61-S66.

Step S61, initialize the algorithm parameters of the gray wolf algorithm IGWO and the parameters of the BP neural network;

In one possible implementation, in the gray wolf algorithm GWO, the wolves are divided into four groups from top to bottom according to the fitness value: ,in is the leadership level (optimal solution), and the candidate solutions revolve around Make location updates. The present invention determines that the gray wolf population size is set to M; the maximum number of iterations is ;The upper and lower boundaries of the jth dimension are set to . The connection weight of each layer of the neural network is , the error threshold is , the number of hidden layer neurons is .

Step S62, normalize the parameters to obtain normalized parameters, and divide the normalized parameters into a training data set and a test data set;

In order to eliminate the impact of the input flow residual sequence dimension on the model results, the data needs to be normalized. Normalize the data to , and divided into training data set and test data set.

In one possible implementation, normalization is performed according to the following formula:

;

In the formula, is the normalization parameter, It’s the first original parameters, and are respectively the maximum and minimum values corresponding to the original parameters.

Before training, determine the termination condition (the current solution is the minimum value) as: if a value remains unchanged in consecutive iterations, it is considered the minimum value. The selected sample executes the following steps S63-S66 until the termination condition is met, and then exits.

Step S63, train the BP neural network according to the training data set to obtain multiple training models;

In a possible implementation, each individual position contains BPNN parameters, and the BP neural network is trained according to the following formula to obtain multiple training models;

;

In the formula, respectively and distance between for current location, is a random vector, is the optimal solution, Connect weights for each layer of the BP neural network.

Step S64: Evaluate the performance of the training model based on the fitness of individuals in the test data set;

In a possible implementation, the individual fitness is calculated based on the mean square error, and the mean square error is calculated according to the following formula:

;

In the formula, T is the sequence length of satellite network traffic; for The tth value of the fitted residual sequence.

Step S65: Classify the gray wolf population according to fitness, retain the position of the individual with the best fitness, and update the positions of the remaining individuals;

In one possible implementation, the position of the individual is updated according to the following formula:

;

;

In the formula, respectively and distance between respectively current location, for current location, is a random vector, is the number of iterations, are the positions of the three types of operators respectively.

Step S66, if the number of iterations reaches the preset value, that is , the parameter optimization process ends and the loop is jumped out. Position the individual with the best fitness As the optimal parameters of the BP neural network model, a prediction model is constructed.

Step S70: Predict the residual sequence according to the optimization model to obtain the residual sequence prediction result;

Step S80: Determine the satellite network traffic prediction result based on the linear feature sequence and residual sequence prediction results.

In a possible implementation, the satellite network traffic prediction result is determined according to the following formula:

;

In the formula, is the satellite network traffic prediction result, is a linear feature sequence, Predict the result for the residual sequence.

The prediction result accuracy index of the EMD-ARIMA-BPNN model is shown in Table 2:

Table 2

The calculation formulas for each indicator in the table are as follows:

;

;

Among them, T is the traffic sequence length; is the flow sequence residual, is the predicted value of the flow series residual, is the mean value of the residual prediction value of the flow series.

Figure 7 is a schematic diagram of the satellite network traffic prediction results provided by the embodiment of the present invention. The method proposed by the present invention has better prediction performance on satellite network traffic data with long-range correlation characteristics, and the predicted network traffic is close to the actual traffic data.

The satellite network traffic prediction method based on empirical mode decomposition and BP neural network of the present invention quantitatively analyzes the self-similarity characteristics of satellite network traffic, and proves through theoretical analysis and experiments respectively that the satellite network traffic is obtained by EMD decomposition Multi-order IMF components have short-range correlation. The satellite network traffic with self-similarity is decomposed into multi-order IMF components with short-range correlation through EMD, and the ARIMA improved by the adaptive fixed-order optimization operator is used to predict the IMF components to reduce the computational complexity. At the same time, the IGWO algorithm is used to optimize the BPNN network weight, and the optimized BPNN is used to predict the residuals of the EMD-ARIMA model, and finally the satellite network traffic forecast results are obtained. Comparative experiments with traditional traffic forecast models and hybrid models prove that the proposed EMD-ARIMA-BPNN model has higher forecast accuracy for satellite network traffic, and can describe the phased change trend of satellite network traffic, and can meet the needs of satellite network traffic. efficient forecasting.

The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention. should be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims (2)

1. The satellite network flow prediction method based on empirical mode decomposition and BP neural network is characterized by comprising the following steps of:

step S10, generating a time sequence of satellite network traffic according to an ON/OFF source model meeting Pareto distribution; the ON/OFF source model meeting Pareto distribution is a data generation model of end-to-end connection in a satellite network;

s20, decomposing the time sequence into multi-order inclusion modal components with short-range correlation by adopting empirical mode decomposition;

step S30, respectively establishing a prediction model for the content modal components of each order by adopting ARIMA improved by the self-adaptive order-determining optimizing operator;

step S40, predicting the connotation modal components by adopting the prediction models of all orders, and summarizing the prediction results to obtain a linear characteristic sequence of the satellite network flow;

step S50, subtracting the linear characteristic sequence from the original flow sequence of the satellite network flow to obtain a residual sequence of the satellite network flow;

step S60, adopting an improved gray wolf algorithm IGWO to optimize the BP neural network, and constructing an optimized model IGWO-BPNN;

step S70, predicting the residual sequence according to the optimization model IGWO-BPNN to obtain a residual sequence prediction result;

and S80, determining a satellite network flow prediction result according to the linear characteristic sequence and the residual sequence prediction result.

2. The method according to claim 1, wherein in S10, the probability distribution function of the Pareto distribution is the following formula:

where k is a positive integer, representing the minimum value that the random variable x can take.