Computing power network traffic prediction method based on wavelet decomposition and improved LSTM
Through wavelet decomposition and improved LSTM model, the complexity and uncertainty of computing power network traffic prediction are solved, high-precision and real-time traffic prediction are achieved, and resource utilization and business continuity are improved.
Patent Information
- Application Number
- CN202411191780.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-28
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2044-08-28
AI Technical Summary
The complexity and uncertainty of computing power network traffic make it difficult for the prior art to achieve accurate, real-time and highly adaptable traffic prediction, and there is a lack of special prediction methods.
Using a method based on wavelet decomposition and improved LSTM, a prediction model including wavelet decomposition algorithm, convolutional length and short-time memory network and bidirectional length and short-time memory network are constructed through data preprocessing, and the Adam optimization algorithm is used for iterative training to extract time and space Features and real-time predictions.
It realizes high-precision and real-time prediction of computing power network traffic, improves the accuracy and adaptability of prediction, and can effectively support business decision-making and resource scheduling.
Smart Images

Figure CN118714028B_ABST
Abstract
Description
Technical Field
[0001] The embodiments of the present invention relate to the field of communication technology, and in particular to a computing power network traffic prediction method based on wavelet decomposition and improved LSTM. Background Art
[0002] With the vigorous development of the digital economy, the computing network has become the core infrastructure supporting the innovative applications of new generation information technology such as cloud computing, big data, and artificial intelligence. The computing centers that make up the computing network need to be interconnected at high speed through the computing network to achieve rapid data transmission and processing, and provide strong computing support for various applications. However, with the expansion of business scale and the growth of data processing needs, the network throughput pressure faced by the computing center is increasing, and the complexity and uncertainty of network traffic have a significant impact on the overall performance of the computing network. As an important part of the integrated operation and management of the computing network, the prediction of computing network traffic is of great significance for improving the resource utilization of the computing network, ensuring business continuity, and optimizing user experience. By accurately predicting the computing network traffic, the computing network platform center can reasonably schedule and deploy computing resources to avoid data transmission congestion and idle computing resources. It also helps to discover potential security risks and abnormal traffic patterns in advance, and ensure the security of the computing network and data security.
[0003] Problems with existing technologies:
[0004] 1. The traffic characteristics of the computing power network are highly complex and uncertain, and are affected by many factors, such as user behavior, application characteristics, device performance, network topology, etc. This makes it very difficult to accurately predict the traffic of the computing power network, and there is no traffic prediction method specifically for the computing power network.
[0005] 2. The combined model often ignores the correlation between the decomposed components, which will affect the accuracy of the prediction results.
[0006] 3. The network traffic of the computing center is dynamically changing and requires real-time prediction to support business decisions. This requires the prediction model to have a high degree of dynamic adaptability and real-time response capabilities.
[0007] It can be seen that there is an urgent need for a computing network traffic prediction method based on wavelet decomposition and improved LSTM with strong real-time prediction, accuracy and adaptability. Summary of the invention
[0008] In view of this, an embodiment of the present invention provides a computing power network traffic prediction method based on wavelet decomposition and improved LSTM, which at least partially solves the problems of poor prediction real-time, accuracy and adaptability in the prior art.
[0009] The embodiment of the present invention provides a computing power network traffic prediction method based on wavelet decomposition and improved LSTM, including:
[0010] Step 1: Data preprocessing: standardize the computing power network traffic data, and then divide the computing power network traffic data into training set and test set;
[0011] Step 2, constructing an improved LSTM computing power network traffic prediction model, wherein the computing power network traffic prediction model includes a wavelet decomposition algorithm, a convolutional long short-term memory network, and a bidirectional long short-term memory network;
[0012] Step 3: Input the training set into the computing power network traffic prediction model, perform iterative training based on the Adam optimization algorithm, and obtain a trained computing power network traffic prediction model;
[0013] Step 4: Input the test set into the trained computing power network traffic prediction model to evaluate whether its accuracy meets the requirements. If yes, proceed to step 5; if no, return to step 3.
[0014] Step 5: Collect the target traffic data and input it into the trained computing network traffic prediction model. Decompose the traffic sequence through wavelet to obtain multiple components and concatenate them into a matrix. Use the parallelized convolutional long short-term memory network to extract the time and space features between the components. Then convert the feature matrix into a one-dimensional feature vector. Use the bidirectional long short-term memory network with exponential gating to calculate the one-dimensional feature vector in sequence and inverse order to obtain two different sets of hidden layer representations. Then, the final hidden representation is obtained by vector concatenation, and the real-time prediction result is output through a fully connected layer.
[0015] According to a specific implementation of the embodiment of the present invention, step 1 specifically includes:
[0016] Step 1.1, read the computing power network traffic data set to the local, wherein the computing power network traffic data set includes the network traffic values of the target computing power center at each historical moment;
[0017] Step 1.2: Get the maximum flow rate of the computing network flow data set and flow minimum;
[0018] Step 1.3, standardize the computing network traffic data set, that is ;
[0019] Among them, x is the original computing power network traffic data, is the minimum value of the historical traffic data of the computing power network, is the maximum value of the historical traffic data of the computing power network, is the standardized result;
[0020] Step 1.4, set the size t of each sample and the corresponding label size p, that is, based on the traffic data of t historical moments Predict traffic data for the next p moments , according to the set t and p, the computing network traffic data is transformed into multiple samples and corresponding labels;
[0021] Step 1.5, divide the samples and corresponding labels into training set and test set.
[0022] According to a specific implementation of an embodiment of the present invention, the expression of the convolutional long short-term memory network is:
[0023]
[0024] The current time is recorded as t, and t-1 is the previous time. is the input at the current moment, represents convolution, Representing Hadamard, is the sigmoid activation function, tanh and softmax are also activation functions. A newly designed compensation door. To compensate for the The three convolution kernels are To compensate for the three bias terms of the gate, first multiply the two convolution results and activate them, then multiply them with the third convolution result and activate them to get the result of the compensation gate. represents the hidden state, represents the input gate, represents the forget gate, represents the output gate, is the cell state, is an alternative state of the cell state, They are the input gate, forget gate, cell state and output gate pair respectively. The convolution kernel, They are the input gate, forget gate, cell state and output gate pair respectively. The convolution kernel, They are the bias items of input gate, forget gate, cell state and output gate respectively;
[0025] Each hidden state is compensated for the final output gate by adaptive weighted summation to obtain the final output. The adaptive weighted process expression is:
[0026]
[0027]
[0028]
[0029]
[0030] in, is the convolution operation, is the hidden state at time t, is a convolution kernel, is the bias term, is the scaling factor, all are learnable parameters, and the whole process is a hidden state Assign a coefficient , by adding the information in a weighted addition way, we can get , and then Update to output gate .
[0031] According to a specific implementation of an embodiment of the present invention, the expression of the bidirectional long short-term memory network is:
[0032] ;
[0033] in, is the hidden state at the previous moment, is the current input, exp is the exponential function with e as the base, is the sigmoid function, is the bias vector of the forget gate, represents the input gate, is an alternative vector, is the bias vector of the update gate, is the memory unit bias vector, is the current unit status, is the output gate, is the output gate bias vector, is the current hidden state, For a stable state, and is the result after stabilization.
[0034] According to a specific implementation of the embodiment of the present invention, step 3 specifically includes:
[0035] Step 3.1, set the batch size and divide the training set into multiple batches;
[0036] Step 3.2, input each batch of data into the computing power network traffic prediction model and calculate the loss function;
[0037] Step 3.3, perform iterative optimization based on the Adam algorithm until the set number of iterations;
[0038] Step 3.4, at the end of the iteration, save the model parameters corresponding to the trained computing power network traffic prediction model.
[0039] According to a specific implementation of the embodiment of the present invention, step 4 specifically includes:
[0040] Step 4.1, input the test set into the trained computing power network traffic prediction model to obtain the prediction results at p moments;
[0041] Step 4.2, evaluate the model accuracy using two indicators: root mean square error and absolute percentage mean error. If its accuracy meets the requirements, proceed to step 5; if not, return to step 3.
[0042] According to a specific implementation of the embodiment of the present invention, the expression of the root mean square error is:
[0043]
[0044] The expression for the absolute percentage mean error is:
[0045]
[0046] in, is the predicted value of the model, is the true value in the training set.
[0047] According to a specific implementation of an embodiment of the present invention, the model parameters include the number of hidden layers, the number of neurons in the hidden layers, the size of the convolution kernel, the learning rate, the maximum number of iterations, and the number of predicted future moments.
[0048] According to a specific implementation of the embodiment of the present invention, step 5 specifically includes:
[0049] Step 5.1, select a wavelet basis function for the target flow data Perform multi-layer decomposition to obtain multiple decomposed components, combine the components into a matrix, and fill the components with insufficient length with 0;
[0050] Step 5.2, input the component matrix into the parallelized convolutional long short-term memory network, extract the features between the components, and convert the feature matrix into a one-dimensional feature vector;
[0051] Step 5.3, input the feature vector into the bidirectional long short-term memory network with exponential gating to extract the temporal features;
[0052] Step 5.4, finally output the prediction result through a fully connected layer;
[0053] Step 5.5, add the prediction results to the sample and repeat step 5.4 until all the data for the next p moments are calculated.
[0054] The computing power network traffic prediction scheme based on wavelet decomposition and improved LSTM in the embodiment of the present invention includes: step 1, data preprocessing, standardizing the computing power network traffic data, and then dividing the computing power network traffic data into a training set and a test set; step 2, constructing an improved LSTM computing power network traffic prediction model, wherein the computing power network traffic prediction model includes a wavelet decomposition algorithm, a convolutional long short-term memory network, and a bidirectional long short-term memory network; step 3, inputting the training set into the computing power network traffic prediction model, and performing iterative training based on the Adam optimization algorithm to obtain a trained computing power network traffic prediction model; step 4, inputting the test set into the trained computing power network traffic prediction model. The network traffic prediction model is evaluated to see whether its accuracy meets the requirements. If so, it goes to step 5, if not, it returns to step 3; in step 5, the target traffic data is collected and input into the trained computing power network traffic prediction model, and the traffic sequence is decomposed by wavelet to obtain multiple components and spliced into a matrix. The time and space features between the components are extracted through a parallelized convolutional long short-term memory network, and then the feature matrix is converted into a one-dimensional feature vector. The one-dimensional feature vector is calculated sequentially and inversely through a bidirectional long short-term memory network with exponential gating to obtain two different sets of hidden layer representations, and then the final hidden representation is obtained through vector splicing, and then the real-time prediction result is output through a fully connected layer.
[0055] The beneficial effects of the embodiments of the present invention are as follows: through the scheme of the present invention, the computing power network traffic sequence is first decomposed, and the features between the decomposed components are extracted using a parallelized convolutional long short-term memory network ConvLSTM module, and then the timing features are extracted through a bidirectional long short-term memory network BiLSTM module with exponential gating, and finally output through a fully connected layer. The model is lightweight and has strong adaptability to samples, and can make high-precision predictions on changes in computing power network traffic in a specific area. BRIEF DESCRIPTION OF THE DRAWINGS
[0056] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0057] Figure 1 A flowchart of a computing power network traffic prediction method based on wavelet decomposition and improved LSTM provided in an embodiment of the present invention;
[0058] Figure 2 A schematic diagram of the structure of a computing power network traffic prediction model based on wavelet decomposition and improved LSTM provided in an embodiment of the present invention;
[0059] Figure 3 A comparison chart between the predicted results and actual results of a computing power network traffic prediction method based on wavelet decomposition and improved LSTM provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0060] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.
[0061] The following describes the embodiments of the present invention through specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work belong to the scope of protection of the present invention.
[0062] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein can be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on the present invention, it should be understood by those skilled in the art that an aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects described herein can be used to implement the device and / or practice the method. In addition, other structures and / or functionalities other than one or more of the aspects described herein can be used to implement this device and / or practice this method.
[0063] It should also be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention. The drawings only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.
[0064] Additionally, in the following description, specific details are provided to facilitate a thorough understanding of the examples. However, it will be understood by those skilled in the art that the aspects described may be practiced without these specific details.
[0065] An embodiment of the present invention provides a computing power network traffic prediction method based on wavelet decomposition and improved LSTM, which can be applied to the network traffic prediction process in big data scenarios.
[0066] See also Figure 1 , is a flow chart of a method for predicting computing network traffic based on wavelet decomposition and improved LSTM provided by an embodiment of the present invention. Figure 1 As shown, the method mainly comprises the following steps:
[0067] Step 1: Data preprocessing: standardize the computing power network traffic data, and then divide the computing power network traffic data into training set and test set;
[0068] Furthermore, the step 1 specifically includes:
[0069] Step 1.1, read the computing power network traffic data set to the local, wherein the computing power network traffic data set includes the network traffic values of the target computing power center at each historical moment;
[0070] Step 1.2: Get the maximum flow rate of the computing network flow data set and flow minimum;
[0071] Step 1.3, standardize the computing network traffic data set, that is ;
[0072] Among them, x is the original computing power network traffic data, is the minimum value of the historical traffic data of the computing power network, is the maximum value of the historical traffic data of the computing power network, is the standardized result;
[0073] Step 1.4, set the size t of each sample and the corresponding label size p, that is, based on the traffic data of t historical moments Predict traffic data for the next p moments , according to the set t and p, the computing network traffic data is transformed into multiple samples and corresponding labels;
[0074] Step 1.5, divide the samples and corresponding labels into training set and test set.
[0075] In the specific implementation, data preprocessing is performed to standardize the computing power network traffic data, and then the computing power network traffic data is divided into a training set and a test set. Specifically,
[0076] Step 1.1: Read the computing power network traffic data locally. The computing power network traffic data set contains the network traffic values of a specific computing power center at each historical moment and is saved in a csv file;
[0077] Step 1.2: Get the maximum flow rate of the computing power network flow data set and minimum flow rate ;
[0078] Step 1.3: Standardize the raw computing network traffic data by:
[0079] (1)
[0080] Where x is the original computing power network traffic data, is the minimum value of historical traffic data, is the maximum value of historical traffic data, is the standardized result;
[0081] Step 1.4: Set each sample t and the corresponding label size p, that is, sample Corresponding label , the data set is divided into multiple samples and corresponding labels;
[0082] Step 1.5: Use the first 80% of the samples as the training set and the remaining 20% of the samples as the test set.
[0083] Step 2, constructing an improved LSTM computing power network traffic prediction model, wherein the computing power network traffic prediction model includes a wavelet decomposition algorithm, a convolutional long short-term memory network, and a bidirectional long short-term memory network;
[0084] Furthermore, the expression of the convolutional long short-term memory network is
[0085] ;
[0086] The current time is recorded as t, and t-1 is the previous time. is the input at the current moment, represents convolution, Representing Hadamard, is the sigmoid activation function, tanh and softmax are also activation functions. A newly designed compensation door. To compensate for the The three convolution kernels are To compensate for the three bias terms of the gate, first multiply the two convolution results and activate them, then multiply them with the third convolution result and activate them to get the result of the compensation gate. represents the hidden state, represents the input gate, represents the forget gate, represents the output gate, is the cell state, is an alternative state of the cell state, They are the input gate, forget gate, cell state and output gate pair respectively. The convolution kernel, They are the input gate, forget gate, cell state and output gate pair respectively. The convolution kernel, They are the bias items of input gate, forget gate, cell state and output gate respectively;
[0087] Each hidden state is compensated for the final output gate by adaptive weighted summation to obtain the final output. The adaptive weighted process expression is:
[0088]
[0089]
[0090]
[0091]
[0092] in, is the convolution operation, is the hidden state at time t, is a convolution kernel, is the bias term, is the scaling factor, all are learnable parameters, and the whole process is a hidden state Assign a coefficient , by adding the information in a weighted addition way, we can get , and then Update to output gate .
[0093] Furthermore, the expression of the bidirectional long short-term memory network is
[0094] ;
[0095] in, is the hidden state at the previous moment, is the current input, exp is the exponential function with e as the base, is the sigmoid function, is the bias vector of the forget gate, represents the input gate, is an alternative vector, is the bias vector of the update gate, is the memory unit bias vector, is the current unit status, is the output gate, is the output gate bias vector, is the current hidden state, For a stable state, and is the result after stabilization.
[0096] For specific implementation, see Figure 2 First, multiple components are obtained through wavelet decomposition, and the time and space features of each component are extracted through the ConvLSTM module. Then, the feature matrix is converted into a one-dimensional feature vector. The feature vector is calculated sequentially and inversely through the BiLSTM module to obtain two different sets of hidden layer representations. Then, the final hidden representation is obtained through vector splicing, and finally, the network traffic prediction value is output through a fully connected layer. Specifically,
[0097] Step 2.1: Construct a wavelet decomposition module. Select a wavelet basis function, perform multi-layer decomposition on the sample, obtain multiple decomposed components, and concatenate the components into a matrix. Since the length of the components must be consistent, fill the components with insufficient length with 0;
[0098] Step 2.2: Construct a parallelized ConvLSTM module to extract features between components and convert the resulting feature matrix into a one-dimensional feature vector. The parallelized ConvLSTM module includes an input gate, a forget gate, an output gate, a memory element, and a hidden state. Its characteristics are that the input gate, the forget gate, and the output gate no longer depend on the hidden state at the previous moment and can be calculated in advance through parallelization. At the same time, a compensation gate is added to compensate for the loss of accuracy caused by the lack of the hidden state. The expression of the whole process is as follows:
[0099] ;
[0100] In the above formula, the current time is recorded as t, t-1 is the previous time, is the input at the current moment, represents convolution, Representing Hadamard, is the sigmoid activation function, tanh and softmax are also activation functions. A newly designed compensation door. To compensate for the The three convolution kernels are To compensate for the three bias terms of the gate, first multiply the two convolution results and activate them, then multiply them with the third convolution result and activate them to get the result of the compensation gate. represents the hidden state, represents the input gate, represents the forget gate, represents the output gate, is the cell state, is an alternative state of the cell state, They are the input gate, forget gate, cell state and output gate pair respectively. The convolution kernel, They are the input gate, forget gate, cell state and output gate pair respectively. The convolution kernel, They are the bias terms of the input gate, forget gate, cell state and output gate respectively.
[0101] Then, each hidden state is compensated for the final output gate through an adaptive weighted summation method to obtain the final output. The adaptive weighted process expression is:
[0102] (9)
[0103] (10)
[0104] (11)
[0105] (12)
[0106] In the above formula, is the convolution kernel, is the bias term, is the scaling factor, which are all learnable parameters. The whole process is hidden state Assigned a coefficient , by adding the information in this weighted addition way, we can get , and then add it to the output gate , the final output.
[0107] Step 2.3: Construct a BiLSTM module with exponential gating. Use sequential and reverse calculations on the feature vectors to obtain two different sets of hidden layer representations, and then concatenate the vectors to obtain the final hidden representation. The characteristic of the BiLSTM module with exponential gating is that its input gate and forget gate are activated by exponential functions instead of being mapped to a value less than 1 to prevent the memory from being gradually lost. Its specific expression is as follows:
[0108] ;
[0109] In the above formula, For the Gate of Forgetfulness, is an alternative vector, is the unit output, is the output gate, is a hidden state;
[0110] At the same time, in order to prevent the sudden increase of forget gates and input gates, stabilization processing is added later, specifically:
[0111] (19)
[0112] (20)
[0113] (twenty one)
[0114] In the above formula, For a stable state, and is the result after stabilization.
[0115] Step 2.4: Construct the prediction module, which is completed by a fully connected layer and outputs the final hidden representation as the prediction result;
[0116] Step 2.5: Add the prediction results to the sample and repeat the prediction steps to recursively predict the computing power network traffic data for the next p moments.
[0117] The number of hidden layers and neurons in the parallelized ConvLSTM module and the BiLSTM module with exponential gating can be set freely. To ensure the lightweight of the model, two hidden layers are set in the parallelized ConvLSTM module, and the number of neurons is set to 16 and 1. Two hidden layers are set in the BiLSTM module with exponential gating, and the number of neurons is set to 64 and 16.
[0118] Experimental environment: operating system Ubuntu 16.04.6 LTS, graphics card Tesla V100, algorithm framework PaddlePaddle 2.4.0.
[0119] Step 3: Input the training set into the computing power network traffic prediction model, perform iterative training based on the Adam optimization algorithm, and obtain a trained computing power network traffic prediction model;
[0120] Based on the above embodiment, step 3 specifically includes:
[0121] Step 3.1, set the batch size and divide the training set into multiple batches;
[0122] Step 3.2, input each batch of data into the computing power network traffic prediction model and calculate the loss function;
[0123] Step 3.3, perform iterative optimization based on the Adam algorithm until the set number of iterations;
[0124] Step 3.4, at the end of the iteration, save the model parameters corresponding to the trained computing power network traffic prediction model.
[0125] Furthermore, the model parameters include the number of hidden layers, the number of neurons in the hidden layers, the size of the convolution kernel, the learning rate, the maximum number of iterations, and the number of predicted future moments.
[0126] In specific implementation, the training data can be input into the computing power network traffic prediction model based on wavelet decomposition and improved LSTM, and the trained model can be obtained after iterative training.
[0127] Step 3.1: Set the batch size and divide the step training set into several batches;
[0128] Step 3.2: Each batch of data is fed into the computing power network traffic prediction model based on wavelet decomposition and improved LSTM to calculate the loss function;
[0129] Step 3.3: Perform iterative optimization based on the Adam algorithm until the set number of iterations.
[0130] Step 3.4: After the iteration, save the trained model parameters.
[0131] Step 4: Input the test set into the trained computing power network traffic prediction model to evaluate whether its accuracy meets the requirements. If yes, proceed to step 5; if no, return to step 3.
[0132] Based on the above embodiment, step 4 specifically includes:
[0133] Step 4.1, input the test set into the trained computing power network traffic prediction model to obtain the prediction results at p moments;
[0134] Step 4.2, evaluate the model accuracy using two indicators: root mean square error and absolute percentage mean error. If its accuracy meets the requirements, proceed to step 5; if not, return to step 3.
[0135] Furthermore, the expression of the root mean square error is:
[0136]
[0137] The expression for the absolute percentage mean error is:
[0138]
[0139] in, is the predicted value of the model, is the true value in the training set.
[0140] During the specific implementation, the model is tested to evaluate the prediction accuracy of the model.
[0141] Step 4.1: Input the test data into the trained computing power network traffic prediction model based on wavelet decomposition and improved LSTM to obtain the computing power network traffic prediction results at p moments;
[0142] Step 4.2: Evaluate the model using the root mean square error (RMSE) and mean absolute percentage error (MAPE). If the accuracy meets the requirements, proceed to step 5. If not, return to step 3.
[0143] Specifically, the root mean square error (RMSE) and mean absolute percentage error (MAPE) are calculated as follows:
[0144] (twenty two)
[0145] (twenty three)
[0146] In the above formula, is the predicted value of the model, is the true value in the training set.
[0147] Step 5: Collect the target traffic data and input it into the trained computing network traffic prediction model. Decompose the traffic sequence through wavelet to obtain multiple components and concatenate them into a matrix. Use the parallelized convolutional long short-term memory network to extract the time and space features between the components. Then convert the feature matrix into a one-dimensional feature vector. Use the bidirectional long short-term memory network with exponential gating to calculate the one-dimensional feature vector in sequence and inverse order to obtain two different sets of hidden layer representations. Then, the final hidden representation is obtained by vector concatenation, and the real-time prediction result is output through a fully connected layer.
[0148] Based on the above embodiment, step 5 specifically includes:
[0149] Step 5.1, select a wavelet basis function for the target flow data Perform multi-layer decomposition to obtain multiple decomposed components, combine the components into a matrix, and fill the components with insufficient length with 0;
[0150] Step 5.2, input the component matrix into the parallelized convolutional long short-term memory network, extract the features between the components, and convert the feature matrix into a one-dimensional feature vector;
[0151] Step 5.3, input the feature vector into the bidirectional long short-term memory network with exponential gating to extract the temporal features;
[0152] Step 5.4, finally output the prediction result through a fully connected layer;
[0153] Step 5.5, add the prediction results to the sample and repeat step 5.4 until all the data for the next p moments are calculated.
[0154] During specific implementation, the collected computing power network traffic data is input into the computing power network traffic prediction model based on wavelet decomposition and computing power network traffic LSTM to obtain real-time prediction results.
[0155] Through the technical solution of the present invention, the network export traffic data of a computing power node is collected and experimented, with an interval of one hour, 24 groups per day, a total of 29 days of data, the first 25 days of data are used for training, and the last four days of data are used for testing. The sample size is set to 24, and samples are continuously generated in the form of a sliding window. In order to verify the stability of the model, the number of selected prediction moments is experimented, and the indicators are calculated. The results are shown in Table 1.
[0156] Table 1
[0157]
[0158] In order to verify the role of each part of the model, an ablation experiment was also carried out, and the comparison results are shown in Table 2.
[0159] Table 2
[0160]
[0161] The results of the fusion experiment show that the characteristics and relationships between components have a great impact on the final prediction results. The parallelized ConvLSTM module is effective in extracting the characteristics and relationships between components, and this feature needs to be fully utilized through the BiLSTM module with exponential gating.
[0162] See also Figure 3 , use the trained network traffic prediction model to predict the traffic situation, and compare the predicted results with the actual traffic value, where " "Corresponding to the prediction curve, add " The curve represents actual data. It can be seen that the network traffic prediction results of the technical solution of the present invention are basically consistent with the actual results, and the changes in network traffic can be better predicted.
[0163] The computing power network traffic prediction method based on wavelet decomposition and improved LSTM provided in this embodiment first decomposes the computing power network traffic sequence, and uses a parallelized convolutional long short-term memory network ConvLSTM module to extract the features between the decomposed components, and then uses a bidirectional long short-term memory network BiLSTM module with exponential gating to extract the time series features, and finally outputs them through a fully connected layer. The model is lightweight and has strong adaptability to samples, and can make high-precision predictions on the changes in computing power network traffic in a specific area.
[0164] It should be understood that various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof.
[0165] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.
Claims
1. A computing network traffic prediction method based on wavelet decomposition and improved LSTM, characterized in that: include: Step 1: Data preprocessing: standardize the computing power network traffic data, and then divide the computing power network traffic data into training set and test set; Step 2, constructing an improved LSTM computing power network traffic prediction model, wherein the computing power network traffic prediction model includes a wavelet decomposition algorithm, a convolutional long short-term memory network, and a bidirectional long short-term memory network; Step 3: Input the training set into the computing power network traffic prediction model, perform iterative training based on the Adam optimization algorithm, and obtain a trained computing power network traffic prediction model; Step 4: Input the test set into the trained computing power network traffic prediction model to evaluate whether its accuracy meets the requirements. If yes, proceed to step 5; if no, return to step 3. Step 5: Collect the target traffic data and input it into the trained computing network traffic prediction model. Decompose the traffic sequence through wavelet to obtain multiple components and concatenate them into a matrix. Use the parallelized convolutional long short-term memory network to extract the time and space features between the components. Then convert the feature matrix into a one-dimensional feature vector. Use the bidirectional long short-term memory network with exponential gating to calculate the one-dimensional feature vector in sequence and inverse order to obtain two different sets of hidden layer representations. Then, the final hidden representation is obtained by vector concatenation, and the real-time prediction result is output through a fully connected layer.
2. The method according to claim 1, characterized in that , the step 1 specifically includes: Step 1.1, read the computing power network traffic data set to the local, wherein the computing power network traffic data set includes the network traffic values of the target computing power center at each historical moment; Step 1.2: Get the maximum flow rate of the computing network flow data set and flow minimum; Step 1.3, standardize the computing network traffic data set, that is ; Among them, x is the original computing power network traffic data, is the minimum value of the historical traffic data of the computing power network, is the maximum value of the historical traffic data of the computing power network, is the standardized result; Step 1.4, set the size t of each sample and the corresponding label size p, that is, based on the traffic data of t historical moments Predict traffic data for the next p moments , according to the set t and p, the computing network traffic data is transformed into multiple samples and corresponding labels; Step 1.5, divide the samples and corresponding labels into training set and test set.
3. The method according to claim 2, characterized in that , the expression of the convolutional long short-term memory network is ; The current time is recorded as t, and t-1 is the previous time. is the input at the current moment, represents convolution, Representing Hadamard, is the sigmoid activation function, tanh and softmax are also activation functions. A newly designed compensation door. To compensate for the The three convolution kernels are To compensate for the three bias terms of the gate, first multiply the two convolution results and activate them, then multiply them with the third convolution result and activate them to get the result of the compensation gate. represents the hidden state, represents the input gate, represents the forget gate, represents the output gate, is the cell state, is an alternative state of the cell state, They are the input gate, forget gate, cell state and output gate pair respectively. The convolution kernel, They are the input gate, forget gate, cell state and output gate pair respectively. The convolution kernel, They are the bias items of input gate, forget gate, cell state and output gate respectively; Each hidden state is compensated for the final output gate by adaptive weighted summation to obtain the final output. The adaptive weighted process expression is: in, is the convolution operation, is the hidden state at time t, is a convolution kernel, is the bias term, is the scaling factor, all are learnable parameters, and the whole process is a hidden state Assign a coefficient , by adding the information in a weighted addition way, we can get , and then Update to output gate .
4. The method according to claim 3, characterized in that , the expression of the bidirectional long short-term memory network is ; in, is the hidden state at the previous moment, is the current input, exp is the exponential function with e as the base, is the sigmoid function, is the bias vector of the forget gate, represents the input gate, is an alternative vector, is the bias vector of the update gate, is the memory unit bias vector, is the current unit status, is the output gate, is the output gate bias vector, is the current hidden state, For a stable state, and is the result after stabilization.
5. The method according to claim 4, characterized in that , the step 3 specifically includes: Step 3.1, set the batch size and divide the training set into multiple batches; Step 3.2, input each batch of data into the computing power network traffic prediction model and calculate the loss function; Step 3.3, perform iterative optimization based on the Adam algorithm until the set number of iterations; Step 3.4, at the end of the iteration, save the model parameters corresponding to the trained computing power network traffic prediction model.
6. The method according to claim 5, characterized in that , the step 4 specifically includes: Step 4.1, input the test set into the trained computing power network traffic prediction model to obtain the prediction results at p moments; Step 4.2, evaluate the model accuracy using two indicators: root mean square error and absolute percentage mean error. If its accuracy meets the requirements, proceed to step 5; if not, return to step 3.
7. The method according to claim 6, characterized in that , the expression of the root mean square error is The expression for the absolute percentage mean error is in, is the predicted value of the model, is the true value in the training set.
8. The method according to claim 5, characterized in that ,The model parameters include the number of hidden layers, the number of neurons in the ,hidden layers, the size of the convolution kernel, the learning rate, the maximum ,number of iterations and the number of predicted future moments.
9. The method according to claim 7, characterized in that , the step 5 specifically includes: Step 5.1, select a wavelet basis function for the target flow data Perform multi-layer decomposition to obtain multiple decomposed components, combine the components into a matrix, and fill the components with insufficient length with 0; Step 5.2, input the component matrix into the parallelized convolutional long short-term memory network, extract the features between the components, and convert the feature matrix into a one-dimensional feature vector; Step 5.3, input the feature vector into the bidirectional long short-term memory network with exponential gating to extract the temporal features; Step 5.4, finally output the prediction result through a fully connected layer; Step 5.5, add the prediction results to the sample and repeat step 5.4 until all the data for the next p moments are calculated.
