Traffic Matrix Prediction Method, Device, Computer Equipment and Storage Medium

Abstract

The present invention discloses a traffic matrix prediction method, device, computer equipment and storage medium, the method includes: obtaining network traffic matrix data, performing relevant preprocessing on the obtained network traffic matrix data to obtain the corresponding traffic matrix training set; determining the basic settings related to discrete wavelet transform, constructing high and low pass filter weight matrices, and constructing a multi-layer discrete wavelet transform network; determining the parameter settings related to long short-term memory neural network, and constructing a recurrent neural network composed of several parallel long short-term memory neural networks; connecting the multi-layer discrete wavelet transform network and the long short-term memory neural network, constructing a prediction model, determining the settings related to the prediction model parameter optimization, and inputting the traffic matrix training set for training; obtaining the network traffic matrix data to be predicted, inputting the trained prediction model, and obtaining the prediction result of the future traffic matrix. The present invention has the prediction accuracy of the traffic sequence with frequent and severe high volatility characteristics.

Description

Technical Field

[0001] The present invention relates to a traffic matrix prediction method, apparatus, device and medium, and belongs to the technical field of network traffic prediction. Background Art

[0002] Most traditional prediction models adopt linear prediction methods based on Gaussian distribution and Markov process, such as Autoregressive Integrated Moving Average and Seasonal Autoregressive Integrated Moving Average. However, the traffic in modern networks has increasingly shown characteristics such as self-similarity, long-range dependence and high non-linearity. Therefore, the network traffic characteristics can no longer be simply described by a linear model, and satisfactory prediction performance can no longer be obtained through linear traffic prediction methods.

[0003] In order to obtain a model with better prediction effect for network traffic with non-linear characteristics, various traffic prediction methods based on deep learning algorithms have received extensive attention in recent years. A large number of research results show that deep learning-based traffic prediction models can obtain higher accuracy than other types of models in most cases. Models based on Recurrent Neural Network have successfully achieved good results in many time series prediction tasks, and Long Short-Term Memory (LSTM) is an improved RNN architecture, which can alleviate the problems of gradient disappearance and gradient explosion encountered in traditional RNNs.

[0004] In addition, most existing prediction methods can obtain the time domain characteristics of time series, but ignore the frequency characteristics of the series. It should be noted that network traffic always has multiple frequency variations. For example, in a certain network, the peak traffic at a certain time of day is often higher than other times, and the traffic peak on a certain day of the week is usually greater than other days. Such characteristics on a larger time scale lead to low-frequency correlation in traffic; while the traffic peak at a certain time of day is usually greater than other times, which are characteristics on a smaller time scale and lead to high-frequency correlation in traffic. Effectively extracting the time series characteristics on these different frequency scales and using them as part of the training set of deep learning algorithms can help improve the prediction effect.

[0005] Wavelet transform can effectively extract the time domain and frequency domain features of time series at the same time. At present, some hybrid prediction methods combining wavelet transform and recurrent neural network have also emerged. However, in most studies, wavelet transform is only used as a time series preprocessing method independent of neural network. Its parameters are fixed during the training process of neural network, which makes the prediction model less adaptable. In different scenarios, operators need to constantly manually test and adjust the wavelet transform parameters of the prediction model to ensure that the model maintains a certain performance. Summary of the invention

[0006] In view of this, the present invention provides a traffic matrix prediction method, device, computer equipment and storage medium, which use discrete wavelet transform to extract multi-scale time-frequency characteristics of traffic matrix sequence to improve the prediction accuracy of traffic prediction model for traffic sequences with frequent and intense high volatility characteristics; on the other hand, the discrete wavelet transform is approximated in the form of a neural network linear layer, and the wavelet decomposition part is embedded in the neural network to participate in training, thereby achieving the effect of global parameter optimization, and further improving the prediction accuracy and adaptability of the prediction model.

[0007] The first object of the present invention is to provide a flow matrix prediction method

[0008] The second object of the present invention is to provide a traffic matrix prediction device.

[0009] A third object of the present invention is to provide a computer device.

[0010] A fourth object of the present invention is to provide a storage medium.

[0011] The first object of the present invention can be achieved by adopting the following technical solutions:

[0012] A traffic matrix prediction method, the method comprising:

[0013] Acquire network traffic matrix data, and perform relevant preprocessing on the acquired network traffic matrix data to obtain a corresponding traffic matrix training set;

[0014] Determine the basic settings related to discrete wavelet transform, construct high-pass and low-pass filter weight matrices, and use the linear layers of neural networks to build a multi-layer discrete wavelet transform network;

[0015] Determine the parameter settings related to the long short-term memory neural network, and construct a recurrent neural network consisting of several parallel long short-term memory neural networks;

[0016] Connect the multi-layer discrete wavelet transform network and the long short-term memory neural network to construct a multi-layer discrete wavelet transform network-long short-term memory neural network prediction model, determine the relevant settings for optimizing the prediction model parameters, and input the traffic matrix training set for training;

[0017] Obtain the network traffic matrix data to be predicted, input it into the trained multi-layer discrete wavelet transform network-long short-term memory neural network prediction model, and obtain the prediction result of the future traffic matrix.

[0018] Further, for the obtained network traffic matrix data, perform relevant preprocessing on the obtained network traffic matrix data to obtain the corresponding traffic matrix training set, which specifically includes:

[0019] Collect the traffic data between network nodes for a period of time at a fixed period, and integrate the data collected each time into a traffic matrix;

[0020] Arrange the continuously collected traffic matrices in sequence to form a traffic matrix sequence TM;

[0021] Pass the normalized traffic matrix sequence through a sliding window of a certain length. Each time the sliding window moves, a sample is obtained. When the sliding window traverses the entire traffic matrix sequence, the entire traffic matrix sample set is obtained, and the traffic matrix training set is divided from the traffic matrix sample set;

[0022] Statistically calculate the historical maximum and minimum values of the traffic between each pair of nodes in the traffic matrix, and use the maximum and minimum traffic values to normalize all the traffic matrices in the traffic matrix sequence TM.

[0023] Further, for the basic settings related to the discrete wavelet transform, construct the high-pass and low-pass filter weight matrices, which specifically include:

[0024] Use wavelet transform to filter and decompose the traffic matrix sequence to obtain an approximate subsequence and several detail subsequences;

[0025] According to the subsequence performance characteristics under different wavelet mother functions, their orders, and decomposition levels, select the appropriate wavelet mother function settings and decomposition levels;

[0026] According to the determined settings related to the wavelet transform, obtain the corresponding high-pass and low-pass filter coefficients, and construct the high-pass and low-pass filter weight matrices.

[0027] Further, the high-pass and low-pass filter weight matrices are as follows:

[0028]

[0029] Among them, W j l and Wj h respectively correspond to the high-pass filtering weight matrix and the low-pass filtering weight matrix constructed based on the high-pass and low-pass filter coefficients. Among them, S is the length of the traffic matrix sequence of the input data, l = {l(1), l(2), l(3), …, l(K)} and h = {h(1), h(2), h(3), …, h(K)} respectively represent the high-pass and low-pass filter coefficient sequences, K is the length of the filter coefficient sequence, and ε is a decimal close to zero.

[0030] Further, the construction of the multi-layer discrete wavelet transform network by using the neural network linear layer specifically includes:

[0031] According to the determined number of wavelet transform layers i, create 2*i groups of neural network linear layers with the number of neurons decreasing by half, and set the initial weight matrices of these linear layers according to the values of the high-pass and low-pass filtering weight matrices;

[0032] Construct the 2*i groups of neural network linear layers into a multi-layer discrete wavelet transform network.

[0033] Further, the parameters related to the long short-term memory neural network include the input feature dimension of the long short-term memory neural network, the number of neurons in the hidden layer, the type of neuron activation function, the number of network stacking layers, and the dropout value.

[0034] Further, the determination of the relevant settings for optimizing the prediction model parameters includes designing a loss function to ensure the effectiveness of the wavelet transform network and using the RMSprop algorithm to optimize and adjust the prediction model parameters. The loss function is as follows:

[0035]

[0036] Among them, is the loss value, and respectively represent the initial setting values of the saved high-pass weight matrix and low-pass weight matrix. α and β are hyperparameters, and respectively represent the traffic value from node o to node d in the traffic matrix predicted by the model and the real traffic matrix at time t, and T is the number of input traffic matrix samples.

[0037] The second object of the present invention can be achieved by adopting the following technical solutions:

[0038] A traffic matrix prediction device, the device includes:

[0039] A training set acquisition module, configured to acquire network traffic matrix data, and perform relevant preprocessing on the acquired network traffic matrix data to obtain a corresponding traffic matrix training set;

[0040] The multi-layer discrete wavelet transform network construction module is used to determine the basic settings related to the discrete wavelet transform, construct the high-pass and low-pass filtering weight matrices, and construct a multi-layer discrete wavelet transform network using the neural network linear layer;

[0041] The recurrent neural network construction module is used to determine the parameter settings related to the long short-term memory neural network and construct a recurrent neural network composed of a number of parallel long short-term memory neural networks;

[0042] The prediction model construction module is used to connect the multi-layer discrete wavelet transform network and the long short-term memory neural network, construct a multi-layer discrete wavelet transform network-long short-term memory neural network prediction model, determine the relevant settings for optimizing the prediction model parameters, and input the traffic matrix training set for training;

[0043] The prediction module is used to obtain the network traffic matrix data to be predicted, input it into the trained multi-layer discrete wavelet transform network-long short-term memory neural network prediction model, and obtain the prediction result of the future traffic matrix.

[0044] The third object of the present invention can be achieved by adopting the following technical solutions:

[0045] A computer device includes a processor and a memory for storing the processor-executable program. The processor, when executing the program stored in the memory, implements the above-mentioned traffic matrix prediction method.

[0046] The fourth object of the present invention can be achieved by adopting the following technical solutions:

[0047] A storage medium stores a program, and when the program is executed by a processor, it implements the above-mentioned traffic matrix prediction method.

[0048] The present invention has the following beneficial effects compared with the prior art:

[0049] The present invention uses the discrete wavelet transform to fully extract and mine the multi-scale time-frequency features of the traffic matrix sequence, which are used as the input of the subsequent neural network, improving the prediction accuracy of the prediction model for time series with strong fluctuation characteristics. The discrete wavelet transform is approximately implemented by a neural network linear layer with a specific initial weight matrix, so that the wavelet transform module is embedded in the neural network in a tightly coupled form, and its related parameters can also participate in the training of the model parameters, achieving the effect of global parameter optimization, further improving the adaptability and prediction effect of the prediction model, and making up for the shortcoming that the previous wavelet transform-related models require personnel to frequently manually adjust parameters according to the application scenario. Description of the Drawings

[0050] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the accompanying drawings required for the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on the structures shown in these drawings.

[0051] Figure 1 It is a flowchart of the traffic matrix prediction method for Embodiment 1 of the present invention.

[0052] Figure 2 It is a schematic diagram of obtaining training samples through a sliding window for Embodiment 1 of the present invention.

[0053] Figure 3 It is a schematic diagram of the multi-layer discrete wavelet transform for Embodiment 1 of the present invention.

[0054] Figure 4 It is a schematic diagram of the output after the traffic sequence from node 1 to node 3 in the CERNET dataset of Embodiment 1 of the present invention undergoes multi-layer wavelet discrete transformation.

[0055] Figure 5 It is a schematic diagram of approximately implementing discrete wavelet transform based on the neural network linear layer for Embodiment 1 of the present invention.

[0056] Figure 6 It is a schematic diagram of the multi-layer discrete wavelet transform network for Embodiment 1 of the present invention.

[0057] Figure 7 It is a structural diagram of the neuron kernel of the long short-term memory neural network for Embodiment 1 of the present invention.

[0058] Figure 8 It is a framework diagram of the multi-layer discrete wavelet transform network - long short-term memory neural network prediction model for Embodiment 1 of the present invention.

[0059] Figure 9 It is a schematic diagram of the process of training the multi-layer discrete wavelet transform network - long short-term memory neural network prediction model once for Embodiment 1 of the present invention.

[0060] Figures 10a to 10f It is a schematic diagram of comparing the prediction results and the true values of the traffic matrix prediction method of Embodiment 1 of the present invention with other comparison models using the CERNET dataset.

[0061] Figure 11 It is a schematic diagram of comparing the medium and long-term prediction performances of the traffic matrix prediction method of Embodiment 1 of the present invention with other comparison models using the CERNET dataset.

[0062] Figure 12This is a structural block diagram of a flow matrix prediction device according to Embodiment 2 of the present invention.

[0063] Figure 13 This is a structural block diagram of a computer device according to Embodiment 3 of the present invention. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0065] Embodiment 1:

[0066] The embodiment of the present invention provides a traffic matrix prediction method, which is based on a multi-layer discrete wavelet transform network (MDWTN) and a long short-term memory neural network (LSTM), including using discrete wavelet transform to extract multi-scale time-frequency features of a traffic matrix sequence for neural network training, so as to improve the prediction accuracy of a traffic prediction model for a traffic sequence with frequent and drastic high volatility characteristics, and approximate the discrete wavelet transform in the form of a neural network linear layer, embedding the wavelet decomposition part into the neural network to participate in the training, further improving the prediction accuracy and adaptability of the prediction model, and making up for the shortcomings of previous wavelet transform related models that require personnel to frequently manually adjust parameters according to application scenarios.

[0067] like Figure 1 As shown, the traffic matrix prediction method of this embodiment includes the following steps:

[0068] S101, obtaining network traffic matrix data, and performing relevant preprocessing on the obtained network traffic matrix data to obtain a corresponding traffic matrix training set.

[0069] In one embodiment, step S101 specifically includes:

[0070] S1011. Collect traffic data between nodes in the network for a period of time at a fixed period, and integrate the data collected each time into a traffic matrix.

[0071] The traffic matrix data set used in this embodiment is the China Education and Research Network (CERNET), whose network topology consists of 14 routing nodes and 16 undirected links. The network traffic matrix data is collected every 5 minutes, and the total period is from 0:00 on May 1, 2013 to 0:00 on May 22, 2013.

[0072] S1012. Arrange the continuously collected traffic matrices in sequence to form a traffic matrix sequence TM.

[0073] In this embodiment, the traffic matrix TM collected at time t is expressed as:

[0074]

[0075] where represents the traffic demand from routing node o to routing node d, N represents the number of routing nodes in the network, and the traffic matrix sequence TM = {TM1, TM2, …, TM T-1 , TM T}, and T represents the total number of measurement samples.

[0076] S1013. Pass the normalized traffic matrix sequence through a sliding window of a certain length. Each time the sliding window moves, a sample is obtained. When the sliding window traverses the entire traffic matrix sequence, the entire traffic matrix sample set is obtained. Then, divide the traffic matrix sample set into a traffic matrix training set and a traffic matrix test set.

[0077] As Figure 2 shown, use a sliding window with a fixed length of l to slide and traverse on the traffic matrix sequence TM to obtain a sample set. The input features of each sample are expressed as {TM t-l , TM t-l+1 , …, TM t-2 , TM t-1}, and the label of the sample is TM t . The prediction model uses the sample input features to calculate the predicted value of the future traffic matrix and performs backpropagation based on the difference between the predicted value and the label TM t used as the true value to optimize the model parameters. In this embodiment, all the samples obtained by traversing through the sliding window are divided into a training set and a test set according to a ratio of 7:3. The training set is used to optimize and adjust the hyperparameters of the constructed prediction model, and the test set is used to evaluate the prediction performance of the prediction model under the optimal parameters.

[0078] S1014. Statistically calculate the historical maximum and minimum values of the traffic between each pair of nodes in the traffic matrix, and use the maximum and minimum traffic values to normalize all the traffic matrices in the traffic matrix sequence TM.

[0079] Specifically, statistically calculate the historical maximum value max o,d and the minimum value min o,d of the traffic between each pair of nodes in the traffic matrix, and normalize all the traffic matrices in the traffic matrix sequence TM according to the following formula:

[0080]

[0081] S102. Determine the basic settings related to the discrete wavelet transform, construct the high-pass and low-pass filtering weight matrices, and construct a multi-layer discrete wavelet transform network using the neural network linear layer.

[0082] In one embodiment, determining the basic settings related to the discrete wavelet transform and constructing the high-pass and low-pass filtering weight matrices specifically include:

[0083] S1021. Use the wavelet transform to filter and decompose the traffic matrix sequence to obtain an approximate subsequence and several detail subsequences.

[0084] In this embodiment, the tree structure diagram of the i-layer discrete wavelet transform is as Figure 3 shown, where X = {x(1), x(2), …, x(T - 1), x(T)} is the input sequence, l = {l(1), l(2), …, l(K - 1), l(K)} and h = {h(1), h(2), …, h(K - 1), h(K)} correspond to the high-pass filter and low-pass filter coefficients respectively; the subsequence A′ j+1 (0 ≤ j ≤ i) is obtained by convolving the approximate subsequence A j at the j-th layer with the low-pass filter coefficient l, and the subsequence D′ j+1 (0 ≤ j ≤ i) is obtained by convolving the approximate subsequence A j at the j-th layer with the high-pass filter coefficient h, as shown in the following formula:

[0085]

[0086]

[0087] Input X to the first-layer discrete transform, and output A1 and D1. The approximate subsequence A j+1 and the detail subsequence D j+1 are obtained from the outputs of the subsequences A′ j+1 and D′ j+1 after passing through a 1 / 2 downsampler; further, input X to an i-layer discrete wavelet transform, and after processing, a total of one approximate subsequence A i and i detail subsequences {D i , D i-1 , …, D2, D1} are obtained, where the approximate subsequence reflects the overall change trend of the input time series, and the detail subsequences reflect the local detail features of the input time series at different frequency scales.

[0088] S1022. Select the appropriate wavelet mother function settings and decomposition levels according to the performance characteristics of different wavelet mother functions, their orders, and the subsequences at different decomposition levels.

[0089] In this embodiment, a Daubechies wavelet with a finite duration, burst frequency, and amplitude is selected as the mother function of the wavelet transform, and the corresponding vanishing moment order is set to 4. To achieve a visualization effect, a single-pair node traffic sequence from node 1 to node 3 is extracted from the traffic matrix sequence in the CERNET dataset, and this sequence is input to a multi-layer discrete wavelet transform processor with different numbers of layers. The approximate subsequence and each detail sequence output under different-layer discrete wavelet transforms are observed. As Figure 4 shown, select the wavelet transform layer with relatively stable subsequences and a lower number of layers to achieve a balance between time-frequency extraction effects and algorithm complexity.

[0090] S1023. According to the determined wavelet transform-related settings, obtain the corresponding high-pass and low-pass filter coefficients, and construct high-pass and low-pass filtering weight matrices.

[0091] Based on the selected wavelet mother function and its order in this embodiment, the specific values of the high-pass filter and low-pass filter coefficients can be determined, and the high-pass filtering matrix W j l and the low-pass filtering weight matrix W j h are constructed as follows:

[0092]

[0093] where W j l and W j h correspond to the high-pass filtering weight matrix and the low-pass filtering weight matrix constructed based on the high-pass and low-pass filter coefficients respectively. Here, S is the length of the traffic matrix sequence of the input data, l = {l(1), l(2), l(3), …, l(K)} and h = {h(1), h(2), h(3), …, h(K)} represent the high-pass and low-pass filter coefficient sequences respectively, K is the length of the filter coefficient sequence, and ε is a decimal close to zero.

[0094] In one embodiment, a multi-layer discrete wavelet transform network is constructed using the neural network linear layer, specifically including:

[0095] S1024. According to the determined wavelet transform layer i, create 2*i groups of neural network linear layers with the number of neurons decreasing by half successively, and set the initial weight matrices of these linear layers according to the values of the high-pass and low-pass filtering weight matrices.

[0096] As Figure 5 shown, according to the weight matrices and Set the initial value of the weight matrix of the neural network linear layer, and use the feed-forward calculation mode of the neural network linear layer for the input samples to approximately implement the convolution calculation of the discrete wavelet transform, so as to achieve the effect of embedding the discrete wavelet transform into the neural network model. Specifically, the output A of the previous layer j,o,d is used as the input of the next linear layer, and A j,o,d is multiplied by the weight matrix of the linear layer or to perform matrix multiplication, and then add the neuron bias value or Finally, the output is calculated through the activation function σ(x). The mathematical expressions corresponding to the calculations of each subsequence are as follows:

[0097]

[0098] S1025. Construct a multi-layer discrete wavelet transform network with 2*i groups of neural network linear layers.

[0099] In this embodiment, by stacking the above fully connected linear layers, a multi-layer discrete wavelet transform approximately implemented by the neural network linear layer can be constructed, that is, a multi-layer discrete wavelet transform network, as Figure 6 shown.

[0100] S103. Determine the parameter settings related to the long short-term memory neural network, and construct a recurrent neural network composed of several parallel long short-term memory neural networks.

[0101] In this embodiment, determine the relevant setting information such as the input feature dimension, the number of neurons in the hidden layer, the type of neuron activation function, the number of network stacking layers, and the dropout value of the long short-term memory neural network, and construct a recurrent neural network composed of several parallel long short-term memory neural networks; specifically, the subsequences obtained by decomposing the input sequence through the discrete wavelet transform need to pass through a 1 / 2 downsampler at each layer, resulting in a decreasing length of each subsequence layer by layer. Therefore, the input feature dimension and the number of neurons in the hidden layer of each long short-term memory neural network need to be consistent with the length of the corresponding input subsequence. In this embodiment, a single-layer long short-term memory neural network and a ReLU neuron activation function are used, the dropout value is 0.1, and the neuron kernel structure of the long short-term memory neural network is as Figure 7 shown.

[0102] S104. Connect the multi-layer discrete wavelet transform network and the long short-term memory neural network to construct a multi-layer discrete wavelet transform-long short-term memory neural network prediction model, determine the relevant settings for optimizing the prediction model parameters, and input the traffic matrix training set for training.

[0103] In this embodiment, the output of the multi-layer discrete wavelet transform network is used as the input of the long short-term memory neural network to construct a multi-layer discrete wavelet transform network-long short-term memory neural network (MDWTN-LSTM) prediction model. Specifically, as Figure 8 shown, according to the lengths of the subsequences output by the discrete wavelet transform network, a long short-term memory neural network with the same number of hidden layer neurons corresponding to the lengths is selected. Further, in order to retain the features in the original input sequence, the original sequence is also used as the input of one of the long short-term memory neural networks. After the time series of each traffic matrix passes through the logical gates such as the forget gate, information enhancement gate, and output gate in the long short-term memory neuron kernel, the results are output. Subsequently, all the outputs of the parallel long short-term memory neural networks are integrated and then input to a fully connected layer to calculate the prediction result.

[0104] Furthermore, determining the relevant settings for optimizing the prediction model parameters includes designing a loss function to ensure the effectiveness of the wavelet transform network and using the RMSprop (Root Mean Square Prop Algorithm) algorithm to optimize and adjust the prediction model parameters.

[0105] In this embodiment, since the neural network model will optimize the weights of the multi-layer discrete wavelet transform network incorporated into it during training, causing it to deviate from the initial set values, in order to prevent the relevant parameters from deviating too much from the wavelet filter coefficient values and causing the wavelet transform to fail, a loss function that helps ensure the effectiveness of the wavelet transform network is designed:

[0106]

[0107] where is the loss value, and respectively represent the initial set values of the high-pass weight matrix and the low-pass weight matrix saved. α and β are hyperparameters. and respectively represent the traffic value from node o to node d in the traffic matrix predicted by the model at time t and the real traffic matrix. T is the number of input traffic matrix samples.

[0108] The process of training the prediction model of this embodiment once according to the samples is as Figure 9 shown.

[0109] S105. Obtain the network traffic matrix data to be predicted and input it into the trained multi-layer discrete wavelet transform network-long short-term memory neural network prediction model to obtain the prediction result of the future traffic matrix.

[0110] In this embodiment, the samples of the traffic matrix test set are used as the network traffic matrix data to be predicted and input into the trained multi-layer discrete wavelet transform network-long short-term memory neural network prediction model to obtain the prediction results of the future traffic matrix. By setting three prediction effect evaluation indexes, namely, root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE), and two scenarios of short-term and medium- and long-term predictions, the prediction effect of the prediction model is evaluated.

[0111] Furthermore, to confirm the prediction effect of this embodiment, the prediction performance of the traffic matrix prediction method based on MDWTN-LSTM (the method provided in this embodiment) is compared with the traffic matrix prediction methods based on MDWT-LSTM, RNN, GRU, LSTM, and TCN respectively based on the CERNET dataset; among them, MDWT-LSTM is a classic prediction model that combines wavelet transform and long short-term memory neural network, and it does not migrate the wavelet transform module into the neural network in a linear layer mode.

[0112] Among them, for visualization, the traffic prediction values from node 1 to node 3 are extracted from the predicted traffic results of each method, and a comparison chart of its trend with the true value is drawn as Figures 10a to 10f shown. It can be observed that the predicted values of the traffic matrix prediction method based on MDWTN-LSTM can better follow and fit the true values compared with other prediction methods. Furthermore, the RMSE, MAE, and MAPE error indexes of the prediction results of the traffic matrix prediction method based on MDWTN-LSTM relative to the true values are 3479.37, 3805.43, and 8.42% respectively, which are decreased by 1376.2, 1249.1, and 5.42% respectively compared with the sub-optimal traffic matrix prediction method based on MDWT-LSTM. The experimental results fully show that the traffic matrix prediction method based on MDWTN-LSTM can better mine the change trend characteristics of strongly fluctuating traffic, so as to predict more accurate traffic matrix results.

[0113] Among them, as Figure 11 shown, the performance effects of the traffic matrix prediction method based on MDWTN-LSTM and several other comparison methods in the medium- and long-term prediction scenario are shown. It can be observed that as the number of future traffic matrices predicted at one time increases, the performance gap between the traffic matrix prediction method based on MDWTN-LSTM and several other comparison methods narrows, but it still always maintains the smallest average prediction error index value, indicating that the method provided in this embodiment has strong robustness and also has good effects on medium- and long-term predictions.

[0114] It should be noted that although the method operations of the above embodiments are described in a specific order, this does not require or imply that these operations must be performed in that specific order, or that all the shown operations must be performed to achieve the desired result. On the contrary, the depicted steps can be changed in the order of execution. Additionally or alternatively, certain steps can be omitted, multiple steps can be combined into one step for execution, and / or one step can be decomposed into multiple steps for execution.

[0115] Embodiment 2:

[0116] As Figure 11 shown, this embodiment provides a traffic matrix prediction device, which includes a training set acquisition module 1201, a multi-layer discrete wavelet transform network construction module 1202, a recurrent neural network construction module 1203, a prediction model construction module 1204, and a prediction module 1205. The specific functions of each module are as follows:

[0117] The training set acquisition module 1201 is used to acquire network traffic matrix data and perform relevant preprocessing on the acquired network traffic matrix data to obtain a corresponding traffic matrix training set;

[0118] The multi-layer discrete wavelet transform network construction module 1202 is used to determine the basic settings related to discrete wavelet transform, construct high-pass and low-pass filter weight matrices, and construct a multi-layer discrete wavelet transform network using a neural network linear layer;

[0119] The recurrent neural network construction module 1203 is used to determine the parameter settings related to the long short-term memory neural network and construct a recurrent neural network composed of several parallel long short-term memory neural networks;

[0120] The prediction model construction module 1204 is used to connect the multi-layer discrete wavelet transform network and the long short-term memory neural network, construct a multi-layer discrete wavelet transform network-long short-term memory neural network prediction model, determine the relevant settings for optimizing the prediction model parameters, and input the traffic matrix training set for training;

[0121] The prediction module 1205 is used to acquire the network traffic matrix data to be predicted, input it into the trained multi-layer discrete wavelet transform network-long short-term memory neural network prediction model, and obtain the prediction result of the future traffic matrix.

[0122] For the specific implementation of each module in this embodiment, reference can be made to Embodiment 1 above, and details will not be repeated here; it should be noted that the device provided in this embodiment is only illustrated by the above division of functional modules. In practical applications, the above functions can be allocated to different functional modules according to needs, that is, the internal structure can be divided into different functional modules to complete all or part of the functions described above.

[0123] Embodiment 3:

[0124] This embodiment provides a computer device, which can be a computer. As shown in the figure, it includes a processor 1302, a memory, an input device 1303, a display 1304, and a network interface 1305 connected through a device bus 1301. The processor is used to provide computing and control capabilities. The memory includes a non-volatile storage medium 1306 and an internal memory 1307. The non-volatile storage medium 1306 stores an operating system, a computer program, and a database. The internal memory 1307 provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. When the processor 1302 executes the computer program stored in the memory, the traffic matrix prediction method of the above Embodiment 1 is implemented as follows: Figure 13 Obtain network traffic matrix data, and perform relevant preprocessing on the obtained network traffic matrix data to obtain a corresponding traffic matrix training set;

[0125] Determine the basic settings related to discrete wavelet transform, construct high-pass and low-pass filter weight matrices, and use the neural network linear layer to construct a multi-layer discrete wavelet transform network;

[0126] Determine the parameter settings related to the long short-term memory neural network, and construct a recurrent neural network composed of several parallel long short-term memory neural networks;

[0127] Connect the multi-layer discrete wavelet transform network and the long short-term memory neural network to construct a multi-layer discrete wavelet transform network-long short-term memory neural network prediction model, determine the relevant settings for optimizing the prediction model parameters, and input the traffic matrix training set for training;

[0128] Obtain the network traffic matrix data to be predicted, input it into the trained multi-layer discrete wavelet transform network-long short-term memory neural network prediction model, and obtain the prediction result of the future traffic matrix.

[0129] Embodiment 4:

[0130] This embodiment provides a storage medium, which is a computer-readable storage medium and stores a computer program. When the computer program is executed by a processor, the traffic matrix prediction method of the above Embodiment 1 is implemented as follows:

[0131] Obtain network traffic matrix data, and perform relevant preprocessing on the obtained network traffic matrix data to obtain a corresponding traffic matrix training set;

[0132] Determine the basic settings related to discrete wavelet transform, construct high-pass and low-pass filter weight matrices, and use the neural network linear layer to construct a multi-layer discrete wavelet transform network;

[0133] Determine the basic settings related to discrete wavelet transform, construct high-pass and low-pass filter weight matrices, and use the neural network linear layer to construct a multi-layer discrete wavelet transform network;

[0134] Determine the parameter settings related to the long short-term memory neural network, and construct a recurrent neural network composed of several parallel long short-term memory neural networks;

[0135] Connect the multi-layer discrete wavelet transform network and the long short-term memory neural network, construct a multi-layer discrete wavelet transform network-long short-term memory neural network prediction model, determine the relevant settings for optimizing the prediction model parameters, and input the traffic matrix training set for training;

[0136] Obtain the network traffic matrix data to be predicted, input it into the trained multi-layer discrete wavelet transform network-long short-term memory neural network prediction model, and obtain the prediction result of the future traffic matrix.

[0137] It should be noted that the computer-readable storage medium in this embodiment can be a computer-readable signal medium or a computer-readable storage medium or any combination of the two. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor device, apparatus, or device, or any combination of the above. More specific examples of the computer-readable storage medium can include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0138] In this embodiment, a computer-readable storage medium can be any tangible medium that contains or stores a program, and this program can be used by or in combination with an instruction execution device, apparatus, or device. And in this embodiment, a computer-readable signal medium can include a data signal propagated in a baseband or as part of a carrier wave, in which a computer-readable program is carried. Such a propagated data signal can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. A computer-readable signal medium can also be any computer-readable storage medium other than a computer-readable storage medium, and this computer-readable signal medium can send, propagate, or transmit a program for use by or in combination with an instruction execution device, apparatus, or device. The computer program contained on the computer-readable storage medium can be transmitted by any suitable medium, including but not limited to: wires, optical cables, RF (radio frequency), etc., or any suitable combination of the above.

[0139] The computer readable storage medium can be written in one or more programming languages ​​or a combination thereof to execute the computer program of the present embodiment, and the programming language includes an object-oriented programming language, such as Java, Python, C++, and also includes a conventional procedural programming language, such as C language or a similar programming language. The program can be executed entirely on the user's computer, partially on the user's computer, as an independent software package, partially on the user's computer and partially on a remote computer, or completely on a remote computer or server. In the case of a remote computer, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computer (e.g., using an Internet service provider to connect through the Internet).

[0140] In summary, the method of the present invention can accurately predict the future traffic matrix of the communication network, and can help network managers adjust the traffic scheduling strategy in advance, thereby reducing the possibility of link congestion and improving network operation efficiency. Discrete wavelet transform is introduced into the prediction model, and multi-scale time-frequency features are extracted. The feature information and the original traffic matrix sequence are respectively input into a recurrent neural network composed of several long-term short-term memory networks, so that the prediction model can better grasp the trend of network traffic changes, thereby improving the prediction effect of traffic sequences with frequent and drastic high volatility characteristics; the discrete wavelet transform is approximated by using a neural network linear layer with a specific initial weight matrix, so that the wavelet transform module is embedded in the neural network in a tightly coupled form, and participates in the training of model parameters, and finally achieves the effect of global parameter optimization, further improving the prediction accuracy and adaptability of the prediction model, reducing the tedious steps of manual parameter adjustment, and comparing the method of the present invention with a variety of other prediction methods under actual data sets. The experimental results show that it can obtain relatively higher prediction accuracy.

[0141] The above is only a preferred embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical solution and invention concept of the present invention within the scope disclosed by the present invention, which shall fall within the protection scope of the present invention.

Claims

1. A traffic matrix prediction method, characterized in that, The method includes: Obtain network traffic matrix data, and perform relevant preprocessing on the obtained network traffic matrix data to obtain a corresponding traffic matrix training set; Determine the basic settings related to discrete wavelet transform, construct high-pass and low-pass filtering weight matrices, and use the neural network linear layer to construct a multi-layer discrete wavelet transform network; Determine the parameter settings related to the long short-term memory neural network, and construct a recurrent neural network composed of several parallel long short-term memory neural networks; Connect the multi-layer discrete wavelet transform network and the long short-term memory neural network to construct a multi-layer discrete wavelet transform network-long short-term memory neural network prediction model, determine the relevant settings for optimizing the prediction model parameters, and input the traffic matrix training set for training; Obtain the network traffic matrix data to be predicted, input it into the trained multi-layer discrete wavelet transform network-long short-term memory neural network prediction model, and obtain the prediction result of the future traffic matrix; The determination of the basic settings related to discrete wavelet transform and the construction of high-pass and low-pass filtering weight matrices specifically include: Use wavelet transform to filter and decompose the traffic matrix sequence to obtain an approximate subsequence and several detail subsequences; According to the subsequence performance characteristics under different wavelet mother functions, their orders, and decomposition levels, select a suitable wavelet mother function setting and decomposition level; According to the determined wavelet transform related settings, obtain the corresponding high-pass and low-pass filter coefficients, and construct high-pass and low-pass filtering weight matrices; The construction of the multi-layer discrete wavelet transform network using the neural network linear layer specifically includes: According to the determined number of wavelet transform layers \(i\), create \(2\times i\) neural network linear layers with the number of neurons halved successively, and set the initial weight matrices of these linear layers according to the values of the high-pass and low-pass filter weight matrices. Specifically: According to the high-pass filter weight matrix and the low-pass filter weight matrix set the initial values of the weight matrices of the neural network linear layers, and use the feed-forward calculation mode of the neural network linear layers for the input samples to approximately implement the convolution calculation of the discrete wavelet transform. Take the output \(A\) of the previous layer j,o,d as the input of the next linear layer. \(A\) j,o,d is multiplied by the high-pass filter weight matrix of the linear layer or the low-pass filter weight matrix through matrix multiplication, and then add the neuron bias value corresponding to the high-pass filter weight matrix or the neuron bias value corresponding to the low-pass filter weight matrix or the neuron bias value corresponding to the low-pass filter weight matrix Finally, calculate the output through the activation function \(\sigma(x)\). The mathematical expressions for the calculation of each subsequence are:​ Construct a multi-layer discrete wavelet transform network by using 2*i groups of neural network linear layers.

2. The traffic matrix prediction method according to claim 1, characterized in that, The obtaining of the network traffic matrix data and the performance of relevant preprocessing on the obtained network traffic matrix data to obtain a corresponding traffic matrix training set specifically include: Collect the traffic data between network nodes for a period of time at a fixed period, and integrate the data collected each time into a traffic matrix; Arrange the continuously collected traffic matrices in order to form a traffic matrix sequence TM; Pass the normalized traffic matrix sequence through a sliding window of a certain length. Each time the sliding window moves, a sample is obtained. When the sliding window traverses the entire traffic matrix sequence, the entire traffic matrix sample set is obtained. Divide the traffic matrix sample set into a traffic matrix training set; Statistically calculate the historical maximum and minimum values of the traffic between each pair of nodes in the traffic matrix, and use the maximum and minimum traffic values to normalize all the traffic matrices in the traffic matrix sequence TM.

3. The traffic matrix prediction method according to claim 1, characterized in that, The high-pass and low-pass filtering weight matrices are as follows: Among them, W j l and W j h correspond to the high-pass filtering weight matrix and the low-pass filtering weight matrix constructed based on the high-pass and low-pass filter coefficients respectively, where S is the length of the traffic matrix sequence of the input data, l = {l(1), l(2), l(3), …, l(K)} and h = {h(1), h(2), h(3), …, h(K)} represent the high-pass and low-pass filter coefficient sequences respectively, K is the length of the filter coefficient sequence, and ε is a decimal close to zero.

4. The traffic matrix prediction method according to any one of claims 1-3, characterized in that, The parameters related to the long short-term memory neural network include the input feature dimension of the long short-term memory neural network, the number of neurons in the hidden layer, the neuron activation function type, the number of network stacking layers, and the dropout value.

5. The traffic matrix prediction method according to claim 1, characterized in that, The determination of the relevant settings for optimizing the prediction model parameters includes designing a loss function to ensure the performance of the wavelet transform network and using the RMSprop algorithm to optimize and adjust the prediction model parameters. The loss function is as follows: Among them, is the loss value, and respectively represent the initial setting values of the saved high-pass weight matrix and low-pass weight matrix. α and β are hyperparameters. and respectively represent the traffic value from the o-th node to the d-th node in the traffic matrix predicted by the model and the real traffic matrix at time t. T is the number of input traffic matrix samples.

6. A traffic matrix prediction device, characterized in that, The device includes: A training set acquisition module, which is used to acquire network traffic matrix data, perform relevant preprocessing on the acquired network traffic matrix data to obtain a corresponding traffic matrix training set; A multi-layer discrete wavelet transform network construction module, which is used to determine the basic settings related to discrete wavelet transform, construct high-pass and low-pass filtering weight matrices, and construct a multi-layer discrete wavelet transform network by using the neural network linear layer; A recurrent neural network construction module, which is used to determine the parameter settings related to the long short-term memory neural network and construct a recurrent neural network composed of several parallel long short-term memory neural networks; A prediction model construction module, which is used to connect the multi-layer discrete wavelet transform network and the long short-term memory neural network, construct a multi-layer discrete wavelet transform network-long short-term memory neural network prediction model, determine the relevant settings for optimizing the prediction model parameters, and input the traffic matrix training set for training; A prediction module, which is used to acquire the network traffic matrix data to be predicted, input it into the trained multi-layer discrete wavelet transform network-long short-term memory neural network prediction model, and obtain the prediction result of the future traffic matrix; The determination of the basic settings related to discrete wavelet transform and the construction of high-pass and low-pass filtering weight matrices specifically include: Filtering and decomposing the traffic matrix sequence by using wavelet transform to obtain an approximate subsequence and several detail subsequences; Selecting a suitable wavelet mother function setting and decomposition layer number according to the subsequence performance characteristics under different wavelet mother functions, their orders, and decomposition layer numbers; According to the determined wavelet transform related settings, obtaining the corresponding high-pass and low-pass filter coefficients and constructing high-pass and low-pass filtering weight matrices; The construction of the multi-layer discrete wavelet transform network by using the neural network linear layer specifically includes: According to the determined number of wavelet transform layers \(i\), create \(2\times i\) neural network linear layers with the number of neurons halved successively, and set the initial weight matrices of these linear layers according to the values of the high-pass and low-pass filter weight matrices. Specifically: According to the high-pass filter weight matrix and the low-pass filter weight matrix set the initial values of the weight matrices of the neural network linear layers, and use the feedforward calculation mode of the neural network linear layers for the input samples to approximately implement the convolution calculation of the discrete wavelet transform. Take the output \(A\) of the previous layer j,o,d as the input of the next linear layer. \(A\) j,o,d is multiplied by the high-pass filter weight matrix of the linear layer or the low-pass filter weight matrix in matrix multiplication, and then add the neuron bias value corresponding to the high-pass filter weight matrix or the neuron bias value corresponding to the low-pass filter weight matrix or the neuron bias value corresponding to the low-pass filter weight matrix Finally, calculate the output through the activation function \(\sigma(x)\). The mathematical expressions for the calculation of each subsequence are:​ Constructing a multi-layer discrete wavelet transform network by using 2*i groups of neural network linear layers.

7. A computer device, comprising a processor and a memory for storing processor-executable programs, characterized in that, When the processor executes the program stored in the memory, it implements the traffic matrix prediction method according to any one of claims 1-5.

8. A storage medium storing a program, characterized in that, When the program is executed by the processor, it implements the traffic matrix prediction method according to any one of claims 1-5.

Images