An edge network traffic prediction method based on redundant reserve pool calculation
By employing a redundant reserve pool calculation method, key and non-key neurons are identified and their weights are transferred, thus addressing the issue of neural network models on edge devices being susceptible to interference. This approach enables high-precision and lightweight edge network traffic prediction, making it suitable for edge prediction scenarios.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NORTH CHINA UNIVERSITY OF SCIENCE AND TECHNOLOGY
- Filing Date
- 2023-06-19
- Publication Date
- 2026-07-10
AI Technical Summary
When performing traffic prediction on edge devices, existing technologies are susceptible to hardware environment interference, which can lead to failures. Furthermore, traditional fault-tolerant methods consume computational resources and are not conducive to practical deployment.
A redundant reservoir computation method is adopted. Key and non-key neurons are identified through ordinal pattern transformation network, and the weights of key neurons are evenly transferred to non-key neurons to form a new ESN model, which improves the model's fault tolerance and allows for retraining.
It achieves high-precision and lightweight network traffic prediction on edge devices, improves the model's fault tolerance and prediction accuracy, and is suitable for edge prediction scenarios with similar requirements.
Smart Images
Figure CN116668320B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer application technology, and in particular to an edge network traffic prediction method based on redundant reserve pool calculation. Background Technology
[0002] Mobile edge computing is considered a major driving force for unlocking the next generation of mobile communication systems and realizing the vision of the Internet of Everything. With the full commercialization of 5G communication technology, edge network traffic is increasing daily, and edge devices are becoming more diverse while user mobility patterns are becoming more complex. In this context, accurate and reliable edge network traffic prediction has become a crucial means of ensuring network management and control. However, edge devices typically have limited computing and storage resources, thus requiring the development of a more efficient model to meet the needs of edge network traffic prediction.
[0003] Pooled computation is a computational method that uses large-scale sparsely connected neurons for nonlinear modeling, effectively avoiding the time-consuming and gradient-exploding drawbacks of gradient descent. Echo State Networks (ESNs) are a computationally inexpensive pooled computation model with strong nonlinear mapping capabilities. This network only needs to use linear regression to train the output weights to achieve prediction results comparable to or better than general recurrent neural networks. Based on this advantage, ESNs are widely used for nonlinear time series forecasting tasks. This provides an effective solution for traffic forecasting modeling on edge devices.
[0004] In practical deployments, the complexity of edge devices and hardware environments often leads to interference factors such as ray interference and voltage fluctuations, causing neural network model failures. Traditional research on fault tolerance theory largely focuses on models like feedforward neural networks and convolutional neural networks, addressing fault tolerance issues from three aspects: adding redundant nodes, modifying training algorithms, or transforming the training and fault tolerance processes into optimization problems solved by nonlinear optimization algorithms. However, in edge traffic prediction scenarios, modifying training algorithms or optimization methods consumes computational resources that edge devices cannot support, hindering practical model deployment. Summary of the Invention
[0005] The purpose of this application is to provide an edge network traffic prediction method based on redundant reserve pool calculation, in order to address the technical deficiencies existing in the prior art.
[0006] The technical solution adopted to achieve the purpose of this application is:
[0007] An edge network traffic prediction method based on redundant reserve pool calculation includes the following steps:
[0008] Step 1: Collect traffic data from edge base stations, and construct training and test sets based on the traffic data. Input the constructed training set into the ESN model and train it, and collect the reserve pool state matrix.
[0009] Step 2: Use the ordinal pattern transformation network modeling method to perform high-dimensional mapping of the reserve pool state matrix, and calculate the ordinal complexity of the reserve pool neurons according to the complex network metric to determine the key and non-key neurons of the reserve pool.
[0010] Step 3: Transfer the input weights and internal weights of the key neurons in the reserve pool to the non-key neurons in the reserve pool in a redundant manner to form a new ESN model, and retrain the new ESN model to obtain the optimal output weight matrix.
[0011] Step 4: Deploy the trained new ESN model on the edge network traffic management device, collect traffic data from the edge base station in real time as input to the model, and output the model as a prediction of future network traffic.
[0012] In the above technical solution, step 1 specifically includes the following steps:
[0013] S101: Collect traffic data from l edge base stations located at the same position within time T, and divide the traffic data collected within time T into training set and test set in a ratio of 8:2;
[0014] S102: Initialize the ESN model by inputting the training set into the initialized ESN model;
[0015] S103: Train the ESN model after inputting the training set, update and record the state of the reservoir neurons to obtain the reservoir state matrix, and use ridge regression to obtain the output weight matrix.
[0016] S104: Calculate the output based on the output weight matrix.
[0017] In the above technical solution, the initialization of the ESN model includes:
[0018] Select N reservoir neurons from the ESN model, randomly generate input weight matrix and internal weight matrix, and scale the internal weight matrix.
[0019] In the above technical solution, the formula for updating the reservoir neuron state is as follows:
[0020] X(t+1)=Tanh(W in ·u(t+1)+W·x(t))
[0021] In the formula, X(t+1) represents the updated reservoir neuron state, u(t+1) represents the traffic data of the edge base station, and W in represents the input weight matrix, W represents the internal weight matrix, and x(t) represents the reservoir neuron state.
[0022] In the above technical solution, the formula for calculating the output weight matrix using ridge regression is as follows:
[0023] W out =(X T X+λI) -1 X T y
[0024] In the formula, W out X represents the output weight matrix, and X represents the reservoir state matrix. T λ represents the transpose of the reservoir state, λ represents the regularization term, I represents the identity matrix, and y represents the predicted teacher value.
[0025] In the above technical solution, the calculation formula for the output is as follows:
[0026] Y(t)=Tanh(W out ·x(t))
[0027] In the formula, Y(t) represents the output, and W... out represents the output weight matrix, and x(t) represents the reservoir neuron state.
[0028] In the above technical solution, step 2 includes the following steps:
[0029] S201: For the time-varying sequence of the nth reservoir neuron in the reservoir state matrix during training time, select the embedding dimension and embedding delay to obtain the phase space representation of the time-varying sequence;
[0030] S202: Sort each reconstructed component in the phase space representation according to the numerical value of the elements, count the vector composed of the original indices after sorting as a set of ordinal patterns, count the ordinal patterns of each reconstructed component to form a set of ordinal codes, and calculate the transition probability of each set of ordinal codes.
[0031] S203: Using ordinal patterns as nodes and ordinal transition probabilities as edges, generate the ordinal pattern transition graph of the time-varying state of the nth neuron;
[0032] S204: Calculate the global node entropy and average path length of the ordinal pattern transition graph, and combine the global node entropy and average path length to measure the importance of the nth neuron;
[0033] S205: Repeat S201 to S204 to obtain the ordinal complexity of each neuron, and define the neuron type according to the proportions of 10%, 80%, and 10%.
[0034] In the above technical solution, the formula for calculating the conversion probability of each group of ordinal codes is as follows:
[0035]
[0036] In the formula, ρ i,j π represents the conversion probability from the i-th ordinal code to the j-th ordinal pattern, π represents the ordinal pattern, and O represents the number of ordinal patterns.
[0037] In the above technical solution, the formula for measuring the importance of the nth neuron is as follows:
[0038]
[0039] In the formula, c n G represents the importance of the nth neuron; α and β represent the proportionality coefficients of the global node entropy and average path length of the ordinal pattern transition graph; GNE represents the global node entropy of the ordinal pattern transition graph; APL represents the average path length of the ordinal pattern transition graph; G n The ordinal pattern transition diagram representing the time-varying state of the nth neuron, d ij p represents the shortest distance between nodes i and j. i s represents the probability of the i-th node. i O represents the Shannon entropy of the node, and O represents the number of ordinal patterns.
[0040] In the above technical solution, step 3 includes the following steps:
[0041] S301: Sort the input weights and internal weights of the key neurons and non-key neurons from largest to smallest, and select neurons to be weighted according to the principle of matching the highest complexity of the key neurons with the lowest complexity of the non-key neurons.
[0042] S302: According to the principle of halving the weights, the input weight connections, internal weight connections, and bias connections of key neurons are balanced and transferred to non-key neurons, and the connections corresponding to the original weights of key neurons are halved in turn.
[0043] S303: If there is a connection in the neurons to be balanced and migrated that is consistent with the key neuron, then add the connection corresponding to the original weight of the key neuron after halving to the connection corresponding to the original weight of the key neuron; otherwise, create a new connection and use the connection corresponding to the original weight of the key neuron after halving as the weight of the new connection.
[0044] S304: Repeat S302 and S303 until all key neurons have completed the weight balance transfer, and obtain a new input weight matrix and internal weight matrix;
[0045] S305: Input the training set into the new ESN model, update / record the reservoir neuron state to obtain the reservoir state matrix, and use ridge regression to obtain the output weight matrix. Calculate the output based on the output weight matrix.
[0046] S306: Validate the prediction accuracy and execution efficiency of the newly trained ESN model using the test set.
[0047] The beneficial effects of this invention are as follows:
[0048] This invention uses an ordinal pattern transformation network (ESN) as an indicator to determine the importance of neurons, thereby identifying key, intermediate, and non-key neurons in a reserve pool. The weights of key neurons are then evenly distributed to non-key neurons, and the output weights are retrained to improve the model's fault tolerance while maintaining accuracy. This achieves accurate and lightweight edge network traffic prediction based on a highly fault-tolerant ESN. This method and prediction model are portable to edge prediction scenarios with similar requirements, demonstrating a degree of versatility. Attached Figure Description
[0049] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0050] Figure 1 This is a flowchart of the edge network traffic prediction method based on redundant reserve pool calculation according to the present invention.
[0051] Figure 2 This is a flowchart of the invention for identifying key neurons based on ordinal complexity.
[0052] Figure 3 This is a schematic diagram of the weight allocation in the ESN model of the present invention. Detailed Implementation
[0053] To enable those skilled in the art to better understand the present invention, the technical solution of the present invention will be further described below with reference to specific embodiments.
[0054] See Figure 1 An edge network traffic prediction method based on redundant reserve pool calculation includes the following steps:
[0055] Step 1: Collect traffic data from edge base stations, and construct training and testing sets based on the traffic data. Input the constructed training set into the ESN model and train it, and collect the reserve pool state matrix.
[0056] Step 1 specifically includes the following steps:
[0057] S101: Collect traffic data from l edge base stations located at the same position within time T, and divide the traffic data collected within time T into training set and test set in a ratio of 8:2;
[0058] S102: Initialize the ESN model by inputting the training set into the initialized ESN model;
[0059] Furthermore, the initialization of the ESN model includes:
[0060] N reservoir neurons from the ESN model are selected, and the input weight matrix and internal weight matrix are randomly generated. The internal weight matrix is then scaled. The scaled internal weight matrix satisfies the spectral radius ρ < 1.
[0061] S103: Train the ESN model after inputting the training set, update and record the state of the reservoir neurons to obtain the reservoir state matrix, and use ridge regression to obtain the output weight matrix.
[0062] The formula for updating the reservoir neuron state is as follows:
[0063] X(t+1)=Tanh(W in ·u(t+1)+W·x(t))
[0064] In the formula, X(t+1) represents the updated reservoir neuron state, u(t+1) represents the traffic data of the edge base station, and W in represents the input weight matrix, W represents the internal weight matrix, and x(t) represents the reservoir neuron state.
[0065] The formula for calculating the output weight matrix using ridge regression is as follows:
[0066] W out =(X T X+λI) -1 X T y
[0067] In the formula, W out X represents the output weight matrix, and X represents the reservoir state matrix. T λ represents the transpose of the reservoir state, λ represents the regularization term, I represents the identity matrix, and y represents the predicted teacher value.
[0068] S104: Calculate the output based on the output weight matrix.
[0069] The formula for calculating the output is as follows:
[0070] Y(t)=Tanh(W out ·x(t))
[0071] In the formula, Y(t) represents the output, and W... out represents the output weight matrix, and x(t) represents the reservoir neuron state.
[0072] Step 2: The state matrix of the reserve pool is mapped to a high dimension using the ordinal pattern transformation network modeling method, and the ordinal complexity of the neurons in the reserve pool is calculated according to the complex network metric to determine the key and non-key neurons in the reserve pool.
[0073] See Figure 2 Step 2 includes the following steps:
[0074] S201: For the time-varying sequence of the nth reservoir neuron in the reservoir state matrix during training time, select the embedding dimension and embedding delay to obtain the phase space representation of the time-varying sequence;
[0075] S202: Sort each reconstructed component in the phase space representation according to the numerical value of the elements, count the vector composed of the original indices after sorting as a set of ordinal patterns, count the ordinal patterns of each reconstructed component to form a set of ordinal codes, and calculate the transition probability of each set of ordinal codes.
[0076] The formula for calculating the transition probability of each set of ordinal codes is as follows:
[0077]
[0078] In the formula, ρ i,j π represents the conversion probability from the i-th ordinal code to the j-th ordinal pattern, π represents the ordinal pattern, and O represents the number of ordinal patterns.
[0079] S203: Using ordinal patterns as nodes and ordinal transition probabilities as edges, generate the ordinal pattern transition graph of the time-varying state of the nth neuron;
[0080] S204: Calculate the global node entropy and average path length of the ordinal pattern transition graph, and combine the global node entropy and average path length to measure the importance of the nth neuron;
[0081] The formula for measuring the importance of the nth neuron is as follows:
[0082]
[0083] In the formula, c nG represents the importance of the nth neuron; α and β represent the proportionality coefficients of the global node entropy and average path length of the ordinal pattern transition graph; GNE represents the global node entropy of the ordinal pattern transition graph; APL represents the average path length of the ordinal pattern transition graph; G n The ordinal pattern transition diagram representing the time-varying state of the nth neuron, d ij p represents the shortest distance between nodes i and j. i s represents the probability of the i-th node. i O represents the Shannon entropy of the node, and O represents the number of ordinal patterns.
[0084] S205: Repeat S201 to S204 to obtain the ordinal complexity of each neuron, and define the neuron type according to the proportions of 10%, 80%, and 10% respectively. The neuron type includes key neurons, intermediate neurons, and non-key neurons.
[0085] Step 3: Transfer the input weights and internal weights of the key neurons in the reserve pool to the non-key neurons in the reserve pool in a redundant manner to form a new ESN model, and retrain the new ESN model to obtain the optimal output weight matrix.
[0086] See Figure 3 Step 3 includes the following steps:
[0087] S301: Sort the input weights and internal weights of the key neurons and non-key neurons from largest to smallest, and select neurons to be weighted according to the principle of matching the highest complexity of the key neurons with the lowest complexity of the non-key neurons.
[0088] S302: According to the principle of halving the weights, the input weight connections, internal weight connections, and bias connections of key neurons are balanced and transferred to non-key neurons, and the connections corresponding to the original weights of key neurons are halved in turn.
[0089] S303: If there is a connection in the neurons to be balanced and migrated that is consistent with the key neuron, then add the connection corresponding to the original weight of the key neuron after halving to the connection corresponding to the original weight of the key neuron; otherwise, create a new connection and use the connection corresponding to the original weight of the key neuron after halving as the weight of the new connection.
[0090] S304: Repeat S302 and S303 until all key neurons have completed the weight balance transfer, and obtain a new input weight matrix and internal weight matrix;
[0091] S305: Input the training set into the new ESN model, update / record the reservoir neuron state to obtain the reservoir state matrix, and use ridge regression to obtain the output weight matrix. Calculate the output based on the output weight matrix.
[0092] S306: Validate the prediction accuracy and execution efficiency of the newly trained ESN model using the test set.
[0093] Step 4: Deploy the trained new ESN model on the edge network traffic management device, collect real-time traffic data from the edge base stations as input to the model, and output the model as a prediction of future network traffic. This prediction enables more intelligent network traffic management.
[0094] The present invention has been described above by way of example. It should be noted that any simple modifications, alterations or other equivalent substitutions that can be made by those skilled in the art without creative effort without departing from the core of the present invention fall within the protection scope of the present invention.
Claims
1. A method for predicting edge network traffic based on redundant reserve pool calculation, characterized in that, Includes the following steps: Step 1: Collect traffic data from edge base stations, and construct training and test sets based on the traffic data. Input the constructed training set into the ESN model and train it, and collect the reserve pool state matrix. Step 2: Use the ordinal pattern transformation network modeling method to perform high-dimensional mapping of the reserve pool state matrix, and calculate the ordinal complexity of the reserve pool neurons according to the complex network metric to determine the key and non-key neurons of the reserve pool. Step 3: Transfer the input weights and internal weights of the key neurons in the reserve pool to the non-key neurons in the reserve pool in a redundant manner to form a new ESN model, and retrain the new ESN model to obtain the optimal output weight matrix. Step 3 includes the following steps: S301: Sort the input weights and internal weights of the key neurons and non-key neurons from largest to smallest, and select neurons to be weighted according to the principle of matching the highest complexity of the key neurons with the lowest complexity of the non-key neurons. S302: According to the principle of halving the weights, the input weight connections, internal weight connections, and bias connections of key neurons are balanced and transferred to non-key neurons, and the connections corresponding to the original weights of key neurons are halved in turn. S303: If there is a connection in the neurons to be balanced and migrated that is consistent with the key neuron, then add the connection corresponding to the original weight of the key neuron after halving to the connection corresponding to the original weight of the key neuron; otherwise, create a new connection and use the connection corresponding to the original weight of the key neuron after halving as the weight of the new connection. S304: Repeat S302 and S303 until all key neurons have completed the weight balance transfer, and obtain a new input weight matrix and internal weight matrix; S305: Input the training set into the new ESN model, update / record the reservoir neuron state to obtain the reservoir state matrix, and use ridge regression to obtain the output weight matrix. Calculate the output based on the output weight matrix. S306: Validate the prediction accuracy and execution efficiency of the newly trained ESN model using the test set; Step 4: Deploy the trained new ESN model on the edge network traffic management device, collect traffic data from the edge base station in real time as input to the model, and output the model as a prediction of future network traffic.
2. The edge network traffic prediction method according to claim 1, characterized in that, Step 1 specifically includes the following steps: S101: Collect data from the same location within time T. l Traffic data from each edge base station is divided into a training set and a test set in an 8:2 ratio based on the traffic data collected within time T. S102: Initialize the ESN model by inputting the training set into the initialized ESN model; S103: Train the ESN model after inputting the training set, update and record the state of the reservoir neurons to obtain the reservoir state matrix, and use ridge regression to obtain the output weight matrix. S104: Calculate the output based on the output weight matrix.
3. The edge network traffic prediction method according to claim 2, characterized in that, Initializing the ESN model includes: Select N In each ESN model, reservoir neurons randomly generate input weight matrices and internal weight matrices, and then scale the internal weight matrices.
4. The edge network traffic prediction method according to claim 2, characterized in that, The formula for updating the reservoir neuron state is as follows: In the formula, This represents the updated state of the reservoir neurons. Traffic data representing edge base stations, Represents the input weight matrix. Represents the internal weight matrix. This represents the state of reservoir neurons.
5. The edge network traffic prediction method according to claim 2, characterized in that, The formula for calculating the output weight matrix using ridge regression is as follows: In the formula, Represents the output weight matrix. Represents the state matrix of the reserve pool. The transpose matrix representing the state of the reservoir. Represents the regularization term. Represents the identity matrix. This represents the predicted teacher value.
6. The edge network traffic prediction method according to claim 2, characterized in that, The formula for calculating the output is as follows: In the formula, Represents output, Represents the output weight matrix. This represents the state of reservoir neurons.
7. The edge network traffic prediction method according to claim 1, characterized in that, Step 2 includes the following steps: S201: For the time-varying sequence of the nth reservoir neuron in the reservoir state matrix during training time, select the embedding dimension and embedding delay to obtain the phase space representation of the time-varying sequence; S202: Sort each reconstructed component in the phase space representation according to the numerical value of the elements, count the vector composed of the original indices after sorting as a set of ordinal patterns, count the ordinal patterns of each reconstructed component to form a set of ordinal codes, and calculate the transition probability of each set of ordinal codes. S203: Using ordinal patterns as nodes and ordinal transition probabilities as edges, generate the first... n Ordinal pattern transition diagram of time-varying states of neurons; S204: Calculate the global node entropy and average path length of the ordinal pattern transition graph, and combine the global node entropy and average path length to measure the first... n The importance of each neuron; S205: Repeat S201~S204 to obtain the ordinal complexity of each neuron, and define the neuron type according to the proportions of 10%, 80%, and 10% respectively.
8. The edge network traffic prediction method according to claim 7, characterized in that, The formula for calculating the transition probability of each set of ordinal codes is as follows: In the formula, Representing the i Group ordinal encoding to the th j The transition probability of an ordinal pattern Represents ordinal pattern, The number of groups representing the ordinal pattern.
9. The edge network traffic prediction method according to claim 7, characterized in that, The first n The formula for measuring the importance of a neuron is as follows: In the formula, Representing the n The importance of each neuron , The proportionality coefficient representing the global node entropy and average path length of the ordinal pattern transformation graph. The global node entropy represents the ordinal mode transition graph. This represents the average path length of the ordinal pattern transition graph. Representing the n Ordinal pattern transition diagram of time-varying states of neurons, Representative node i and j The shortest distance between them Representing the i The probability of each node. This represents the Shannon entropy of the node. The number of groups representing the ordinal pattern.