Network modeling method based on GNN and Transform model

Abstract

The invention relates to a network modeling method based on GNN and Transform models, and the method comprises the steps: firstly carrying out the generation and preprocessing of a data set, and constructing a message passing neural network; the message passing neural network comprises three parts, namely a GNN cascade module, a Transform cascade module and a specific index reading module; and then training the network, finally inputting the preprocessed test set image into a trained network model to obtain a prediction map, comparing the prediction map with a truth value, and calculating each evaluation index. According to the method, the global dependency relationship can be better captured, so that more comprehensive data representation is obtained, and meanwhile, the model generalization ability and the calculation efficiency are improved.

Description

Technical Field

[0001] The present invention relates to a network modeling method based on GNN and Transformer models, and belongs to the technical field of computer networks. Background Art

[0002] Network models are one of the key tools for designing, building, and evaluating computer networks. By simulating network behavior, they help designers evaluate protocol performance, predict protocol behavior, and perform network performance evaluation, network topology design, network resource allocation, and network optimization. For example, designers can use network models to evaluate the latency, throughput, and packet loss rate of different protocols in the network, predict how routing protocols select paths, evaluate the performance of different network topologies, select the optimal resource allocation solution, and optimize network configuration to improve network performance.

[0003] Traditional network models, such as queuing theory and packet-level models, are widely used but have some limitations. Queuing theory models make too strict assumptions about traffic models and are difficult to simulate the complex traffic characteristics in real networks; while packet-level models can accurately simulate network behavior, they are computationally expensive and difficult to meet the needs of real-time network optimization. For example, on a single 1Gbps link, the number of packets may be in the millions per second. Therefore, this approach is slow and impractical when considering large networks with real traffic. Given the limitations of processing time of such models, their applicability to online network optimization is severely restricted. Summary of the invention

[0004] In order to overcome the problem of how to invent a more advanced technical problem for certain specific scenarios through neural networks, the present invention provides a network modeling method based on GNN and Transformer models that can accurately model complex traffic characteristics and expand to large real-world networks.

[0005] A network modeling method based on GNN and Transformer model, comprising the following steps:

[0006] Step 1: Data collection: Collect network samples generated in simulation or in actual work with different network topologies, traffic models, routing configurations, and queue scheduling policies, and record the average delay, jitter, and packet loss rate of all flows in each network sample, as well as queue-level statistics;

[0007] Step 2: Data preprocessing: The network samples are converted into a graph structure, where nodes represent network devices and edges represent links between connected devices. At the same time, features are extracted for each node and edge, such as traffic characteristics, link capacity, queue size, and queue priority.

[0008] Step 3: Construct a message passing neural network: The message passing neural network consists of three parts: a GNN cascade module, a Transformer cascade module, and a specific indicator readout module. Different functions are used to initialize the hidden states of flows, queues, and links. The flow, queue, and link features are encoded into fixed-size vectors to represent feature embedding. The GNN network is used to perform T message passing iterations. Each iteration includes three stages: hidden state message passing and updating of flows, hidden state message passing and updating of queues, and hidden state message passing and updating of links. These features are then input into the Transformer network, and its self-attention mechanism is used to further process the dependencies between features or learning graphs. The readout function is used to calculate the performance indicators of each flow.

[0009] Step 4: Training the network: Divide the preprocessed training set into a training set and a validation set for training the model and evaluating the model performance. Input the training set images into the message passing neural network, and generate predictions of the performance indicators of each flow through the GNN cascade module, the Transformer cascade module, and the specific indicator readout module. Then use the validation set to calculate the mean square error and coefficient to determine the loss calculation, back propagate, and optimize the weights through the selected optimizer and corresponding parameters. After multiple trainings, the network model is obtained.

[0010] Step 5: Model evaluation: Use unseen network samples to evaluate the generalization ability of the model to predict network performance, including different topologies, traffic models, and routing configurations, and compare them with the true values ​​to calculate various evaluation indicators.

[0011] The step 2 specifically includes:

[0012] The network graph is constructed, with the devices in the network as nodes in the graph and the connections between the devices as edges in the graph. The relevant feature information is extracted from each node and edge, including device type, function, traffic characteristics, queue information, link capacity, queue information, etc. The numerical features are then normalized, and the extracted features are combined with the connection relationship between nodes and edges to construct a network graph dataset.

[0013] The step three specifically includes:

[0014] The GNN cascade module is constructed by multiple recurrent neural networks to form a flow queue, and uses an autoregressive algorithm to perform multiple cycles inside and outside the network and then outputs it to the Transformer cascade module.

[0015] Each input feature of the recurrent neural network is first input into the network one by one. There is a trainable hidden state inside the network, which is updated when processing each element in the sequence. In the recurrent neural network, a specific hidden state plays the role of a potential embedding space. Each input feature is fed into the network one by one and passes through a trainable linear layer together with the hidden state of the previous moment. The linear layer maps the input feature and the hidden state of the previous moment to a new hidden state. The hidden state contains the information of the current input and the previous input. Recurrent neural networks usually do not explicitly learn position information, and implicitly capture the order information in the sequence through recursive structures. In recurrent neural networks, there is no explicit query vector matrix, key vector matrix, and value vector matrix. The hidden state itself is updated in each time step to reflect the information of the current position in the sequence. This process can be described as:

[0016] H t =f(W*X t +U*H t-1 +b)

[0017] Among them, H t is the hidden state at the current time step, X t is the current input feature, W is the weight matrix of the input feature, U is the weight matrix of the hidden state, b is the bias term, and f is a nonlinear activation function;

[0018] Then, the graph features extracted by the GNN cascade module are input into the Transformer cascade module. The cascade Transformer module consists of several Transformer blocks with shared parameters. The Transformer block consists of a multi-head self-attention mechanism network and a feedforward network. The long-distance dependencies of the features are calculated through the self-attention mechanism. Each input feature is first expanded into a one-dimensional sequence. Each sequence is mapped to a potential D-dimensional embedding space through a trainable linear projection layer. Specific position embeddings are learned and added to the patch embedding to retain the position information. The patch with added position information passes through the linear layer to generate three matrices: query vector matrix, key vector matrix and value vector matrix. The three matrices are then sent to the multi-head attention mechanism operation to calculate the attention between each other:

[0019] Att(Q,K,V)=ω(QK T )V

[0020] Where Att(Q,K,V) is the obtained attention value, Q, K, and V are the query vector matrix, key vector matrix, and value vector matrix respectively, and each row in these three matrices represents a corresponding vector;

[0021] The information containing more long-distance dependencies in long sequences after Transformer processing is sent to the specific indicator readout module, and parameters such as link load, end-to-end flow delay, and the sum of all link delays are calculated.

[0022]

[0023]

[0024]

[0025]

[0026] in, is the link load, is the flow f on link l j The load on f is the average flow rate of flow f, is the queue delay, is the transmission delay, is the readout function, x fps is the average packet size, x lc is the link capacity, d link is the link delay.

[0027] The step 4 specifically includes:

[0028] The network samples in the training set are input into the message passing neural network, and the performance indicators of the corresponding network samples are generated through the GNN cascade module, the Transformer cascade module and the specific indicator readout module. Then the mean square error and coefficient are calculated using the validation set to determine the loss calculation:

[0029]

[0030] Mean square error is a method to measure the difference between the model's predicted value and the actual value. It calculates the average of the squares of the difference between the predicted value and the actual value. MSE gives higher weight to larger errors and is very sensitive to outliers.

[0031]

[0032] The coefficient of determination, often called R2 or coefficient of determination, measures how well the model explains the total variation;

[0033]

[0034] Among them, SS res is the residual sum of squares, SS tot is the total sum of squares, and the value of R2 ranges from 0 to 1. The closer the value is to 1, the better the model fits the data;

[0035] After the loss calculation is completed, back propagation is performed, and the connection weights are optimized through the selected optimizer and corresponding parameters. The final network model is obtained after 150 rounds of training.

[0036] The step five specifically includes:

[0037] The evaluation process criteria are mean square error, mean absolute error, mean absolute percentage error, and coefficient determination;

[0038] The mean absolute error is the average of the absolute values ​​of the differences between the predicted values ​​and the actual values. The mathematical formula is:

[0039]

[0040] The mean absolute percentage error is another indicator to measure the accuracy of the forecast. It calculates the average of the ratio of the absolute value of the difference between each predicted value and the actual value to its actual value. MAPE is used to measure the accuracy of the forecast. When the actual value range is large, when the actual value is zero or close to zero, MAPE becomes unstable. The mathematical formula is:

[0041]

[0042] According to the above indicators, the errors compared with the simulation results are calculated when the input network samples are Poisson, Deterministic, On-Off, A.exponentials, M.exponentials, and Mixed.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] The present invention relates to a network modeling method based on GNN and Transformer model, which uses a cascaded Transformer module and embeds it into a GNN-based network modeling model to enhance feature representation. It is composed of several Transformer blocks with shared parameters, which are serially fused with the GNN cascade network and use the Transformer's multi-head attention mechanism to better capture long-distance dependencies in the network model. Serial fusion allows the model to first capture local graph structure information through GNN, and then the Transformer further processes this information to capture global dependencies, thereby obtaining a more comprehensive data representation. The serial structure can be regarded as a process of feature selection and combination, which helps to reduce the risk of overfitting and improve the generalization ability of the model. Compared with directly applying the Transformer on the graph data, the serial structure can first reduce the feature dimension through the GNN, thereby reducing the number of parameters in the Transformer and improving the computational efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art 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.

[0046] Figure 1 This is a message passing neural network structure diagram of a network modeling method based on GNN and Transformer models in the present invention. DETAILED DESCRIPTION

[0047] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0048] Reference Figure 1 , a network modeling method based on GNN and Transformer model, comprising the following steps:

[0049] Step 1: Data collection: Collect network samples generated in simulation or in actual work with different network topologies, traffic models, routing configurations, and queue scheduling policies, and record the average delay, jitter, and packet loss rate of all flows in each network sample, as well as queue-level statistics (such as average occupancy, average packet loss rate, or average packet size).

[0050] In order to build a sample set containing different network scenarios for training and testing network models, data generated randomly based on network simulators and real generated based on physical test beds were used. The omnet++ network simulator was used for data generation in the generation based on network simulators. During the simulation, various characteristics of the network were input, including parameters such as topology, routing configuration, traffic model and its intensity, and the delay of each path was used as the output result. The article used a variety of traffic models to simulate network traffic, including Poisson, On-Off, Autocorrelated Exponential, Modulated Exponential, etc., and used queue scheduling strategies such as FIFO, WFQ, DRR, SP, etc. Random sampling was performed from the possible values ​​of all input variables to generate the configuration parameters of the network scenario, such as the number of nodes, connection mode, traffic model type, parameters, traffic intensity, queue size, scheduling strategy, etc.

[0051] In order to further evaluate the performance of the model in a real environment, a data generation method based on a physical test bed was also used in the real generated data based on the physical test bed. A physical test bed containing Huawei equipment (routers, switches, servers) was built, and the network traffic was generated using the TRex traffic generator. After the traffic was generated, the PF_RING software was used to capture and analyze the network traffic, and the average delay, jitter, packet loss rate and other indicators of each flow, as well as the average occupancy, average packet loss rate, and average packet size of each queue were recorded. Finally, the recorded indicators were used to generate a data set to evaluate the performance of the model in a real environment.

[0052] Step 2: Data preprocessing: The network samples are converted into a graph structure, where nodes represent network devices and edges represent links between connected devices. At the same time, features are extracted for each node and edge, such as traffic features, link capacity, queue size, and queue priority.

[0053] First, the network graph is constructed, with the devices in the network as nodes in the graph and the connections between the devices as edges in the graph. Then, relevant feature information is extracted from each node and edge, including device type, function, traffic characteristics (such as traffic size, type, model parameters), queue information (such as queue size, priority, scheduling strategy), link capacity (such as bandwidth, delay), queue information (such as queue size, priority, scheduling strategy), etc. The numerical features are then normalized, and the extracted features are combined with the connection relationship between nodes and edges to construct a network graph dataset.

[0054] Step 3: Construct a message passing neural network: The message passing neural network includes three parts: a GNN cascade module, a Transformer cascade module, and a specific indicator readout module. First, different functions are used to initialize the hidden states of the stream, queue, and link, and the stream, queue, and link features are encoded into vectors of fixed size to represent feature embedding. Then, the GNN network is used to perform T message passing iterations, each of which includes three stages: hidden state message passing and updating of the stream, hidden state message passing and updating of the queue, and hidden state message passing and updating of the link, so as to achieve the purpose of extracting graph features, and then these features are input into the Transformer network, and its self-attention mechanism is used to further process the features or learn the dependencies between graphs. Finally, the readout function is used to calculate the performance indicators of each stream, such as delay, jitter, and packet loss rate.

[0055] The GNN cascade module is constructed by multiple RNNs to form a flow queue, and uses an autoregressive algorithm to perform multiple cycles inside and outside the network and then outputs it to the Transformer cascade module.

[0056] A recurrent neural network (RNN) is a type of neural network that processes sequence data. It computes long-range dependencies of features by using a recursive structure to maintain memory of previous information. Specifically, each input feature is first fed into the network one by one. The network has a trainable hidden state inside that is updated when processing each element in the sequence. In RNN, a specific hidden state plays the role of a latent embedding space. Each input feature is fed into the network one by one and passes through a trainable linear layer (usually called a weight matrix) together with the hidden state of the previous moment. This linear layer maps the input feature and the hidden state of the previous moment to a new hidden state that contains information about the current input and the previous input. The hidden state can be viewed as an encoding of the previous sequence information in the latent embedding space, with each state vector representing an encoding of the previous sequence information. Unlike the position embedding in the Transformer, RNNs usually do not explicitly learn position information, but implicitly capture the order information in the sequence through a recursive structure. In RNNs, there is no explicit query vector matrix, key vector matrix, and value vector matrix. Instead, the hidden state itself is updated in each time step to reflect the information of the current position in the sequence. This process can be described as:

[0057] H t =f(W*X t +U*H t-1 +b)

[0058] Among them, H t is the hidden state at the current time step, X tis the current input feature, W is the weight matrix of the input feature, U is the weight matrix of the hidden state, b is the bias term, and f is a nonlinear activation function (such as tanh or ReLU).

[0059] Then, the graph features extracted by the GNN cascade module are input into the Transformer cascade module. The cascade Transformer module is composed of several Transformer blocks with shared parameters. The Transformer block is composed of a multi-head self-attention mechanism network and a feedforward network. The long-distance dependencies of the features are calculated through the self-attention mechanism. Specifically, each input feature is first expanded into a one-dimensional sequence. Each sequence is mapped to a potential D-dimensional embedding space through a trainable linear projection layer. In addition, specific position embeddings are learned and added to the patch embedding to retain the position information. The patch with the position information added passes through the linear layer to generate three matrices: query vector matrix, key vector matrix and value vector matrix, and then the three matrices are sent to the multi-head attention mechanism operation to calculate the attention between each other:

[0060] Att(Q,K,V)=ω(QK T )V

[0061] Among them, Att(Q,K,V) is the obtained attention value, Q, K, and V are the query vector matrix, key vector matrix, and value vector matrix respectively, and each row in these three matrices represents a corresponding vector.

[0062] Finally, the information containing more long-distance dependencies in long sequences after Transformer processing is sent to the specific indicator readout module, and parameters such as link load, end-to-end flow delay, and the sum of all link delays are calculated.

[0063]

[0064]

[0065]

[0066]

[0067] in, is the link load, is the flow f on link l j The load on f is the average flow rate of flow f, is the queue delay, is the transmission delay, is the readout function, x fps is the average packet size, x lc is the link capacity, dlink is the link delay.

[0068] Step 4: Training the network: First, divide the preprocessed training set into a training set and a validation set for training the model and evaluating the model performance. Then input the training set images into the message passing neural network, and generate predictions of the performance indicators of each stream through the GNN cascade module, the Transformer cascade module, and the specific indicator readout module; then use the validation set to calculate the mean square error and coefficients to determine the loss calculation, back propagate, and optimize the weights through the selected optimizer and corresponding parameters. After multiple trainings, the network model is obtained.

[0069] The network samples in the training set are input into the message passing neural network. After passing through the GNN cascade module, the Transformer cascade module and the specific indicator readout module, the performance indicators of the corresponding network samples are generated. Then, the validation set is used to calculate the mean square error and coefficient to determine the loss calculation:

[0070]

[0071] The mean square error (MSE) is a measure of the difference between the model's predictions and the actual values. It calculates the average of the squares of the differences between the predicted and actual values. MSE gives higher weights to larger errors, so it is very sensitive to outliers.

[0072]

[0073] The coefficient of determination, often called R2 or coefficient of determination, measures how well the model explains the total variation.

[0074]

[0075] Among them, SS res is the residual sum of squares (the sum of squares of the differences between the predicted values ​​and the actual values), SS tot It is the total sum of squares (the sum of the squares of the differences between the actual values ​​and the mean of the actual values). R2 values ​​range from 0 to 1, with values ​​closer to 1 indicating a better fit of the model to the data.

[0076] After the loss calculation is completed, back propagation is performed to optimize the connection weights through the selected optimizer and corresponding parameters. The final network model is obtained after 150 rounds of training.

[0077] Step 5: Model evaluation: Use unseen network samples to evaluate the generalization ability of the model to predict network performance, including different topologies, traffic models, and routing configurations, and compare them with the true values ​​to calculate various evaluation indicators.

[0078] The evaluation process criteria are mean square error (MSE), mean absolute error (MAE), mean absolute percentage error (MAPE), coefficient of determination (R2)

[0079] Among them, the mean square error and coefficient determination have been mentioned above and will not be repeated here.

[0080] The mean absolute error calculates the average of the absolute values ​​of the differences between the predicted values ​​and the actual values. Compared with MSE, MAE is less sensitive to outliers because it only considers the size of the error instead of the square of the error. Its mathematical formula is:

[0081]

[0082] The mean absolute percentage error is another indicator for measuring forecast accuracy. It calculates the average of the absolute value of the difference between each predicted value and the actual value and its actual value. MAPE is often used to measure the accuracy of forecasts, especially when the actual value range is large. However, when the actual value is zero or close to zero, MAPE may become unstable. The mathematical formula is:

[0083]

[0084] According to the above indicators, the errors compared with the simulation results are calculated when the input network samples are Poisson, Deterministic, On-Off, A.exponentials, M.exponentials, and Mixed.

[0085] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions and variations of these embodiments are made without departing from the principles and spirit of the present invention, and still fall within the scope of protection of the present invention.

Claims

1. A network modeling method based on GNN and Transformer model, characterized by: The following steps are involved: Step 1: Data collection: Collect network samples generated in simulation or in actual work with different network topologies, traffic models, routing configurations, and queue scheduling policies, and record the average delay, jitter, and packet loss rate of all flows in each network sample, as well as queue-level statistics; Step 2: Data preprocessing: The network samples are converted into a graph structure, where nodes represent network devices and edges represent links between connected devices. At the same time, features are extracted for each node and edge, such as traffic characteristics, link capacity, queue size, and queue priority. Step 3: Construct a message passing neural network: The message passing neural network consists of three parts: a GNN cascade module, a Transformer cascade module, and a specific indicator readout module. Different functions are used to initialize the hidden states of flows, queues, and links. The flow, queue, and link features are encoded into fixed-size vectors to represent feature embedding. The GNN network is used to perform T message passing iterations. Each iteration includes three stages: hidden state message passing and updating of flows, hidden state message passing and updating of queues, and hidden state message passing and updating of links. These features are then input into the Transformer network, and its self-attention mechanism is used to further process the dependencies between features or learning graphs. The readout function is used to calculate the performance indicators of each flow. Step 4: Training the network: Divide the preprocessed training set into a training set and a validation set, train the model and evaluate the model performance, input the training set images into the message passing neural network, and generate predictions for the performance indicators of each flow through the GNN cascade module, the Transformer cascade module, and the specific indicator readout module; Then use the validation set to calculate the mean square error and coefficient to determine the loss calculation, back propagate, optimize the weights through the selected optimizer and corresponding parameters, and obtain the network model after multiple trainings; Step 5: Model evaluation: Use unseen network samples to evaluate the generalization ability of the model to predict network performance, including different topologies, traffic models, and routing configurations, and compare them with the true values ​​to calculate various evaluation indicators.

2. According to claim 1, a network modeling method based on GNN and Transformer model is characterized in that: The step 2 specifically includes: Construct a network graph, treat the devices in the network as nodes in the graph, and the connections between devices as edges in the graph. Extract relevant feature information from each node and edge, including device type, function, traffic characteristics, queue information, link capacity, and queue information. Normalize the numerical features, combine the extracted features with the connection relationship between nodes and edges, and construct a network graph dataset.

3. According to claim 1, a network modeling method based on GNN and Transformer model is characterized in that: The step three specifically includes: The GNN cascade module is constructed by multiple recurrent neural networks to form a flow queue, and uses an autoregressive algorithm to perform multiple cycles inside and outside the network and then outputs it to the Transformer cascade module. Each input feature of the recurrent neural network is first input into the network one by one. There is a trainable hidden state inside the network, which is updated when processing each element in the sequence. In the recurrent neural network, a specific hidden state plays the role of a potential embedding space. Each input feature is sent into the network one by one, and passes through a trainable linear layer together with the hidden state of the previous moment. The linear layer maps the input feature and the hidden state of the previous moment to a new hidden state. The hidden state contains the information of the current input and the previous input. Recurrent neural networks usually do not explicitly learn position information, and implicitly capture the order information in the sequence through recursive structures. In recurrent neural networks, there is no explicit query vector matrix, key vector matrix, and value vector matrix. The hidden state itself is updated in each time step to reflect the information of the current position in the sequence. This process can be described as: H t =f(W*X t +U*H t-1 +b) Among them, H t is the hidden state at the current time step, X t is the current input feature, W is the weight matrix of the input feature, U is the weight matrix of the hidden state, b is the bias term, and f is a nonlinear activation function; The graph features extracted by the GNN cascade module are input into the Transformer cascade module. The cascade Transformer module consists of several Transformer blocks with shared parameters. The Transformer block consists of a multi-head self-attention mechanism network and a feedforward network. The long-distance dependencies of the features are calculated through the self-attention mechanism. Each input feature is first expanded into a one-dimensional sequence. Each sequence is mapped to a potential D-dimensional embedding space through a trainable linear projection layer. The specific position embedding is learned and added to the patch embedding to retain the position information. The patch with the position information added passes through the linear layer to generate three matrices: query vector matrix, key vector matrix and value vector matrix. The three matrices are then sent to the multi-head attention mechanism operation to calculate the attention between each other: That(Q,K,V)=ω(QK T )V Where Att(Q, K, V) is the obtained attention value, Q, K, V are the query vector matrix, key vector matrix and value vector matrix respectively, and each row in these three matrices represents a corresponding vector; The information containing long-distance dependencies in long sequences after Transformer processing is sent to the specific indicator readout module, and parameters such as link load, end-to-end flow delay, and the sum of all link delays are calculated. in, is the link load, is the flow f on link l j The load on the flow, λf is the average flow rate of flow f, is the queue delay, is the transmission delay, is the readout function, x fps is the average packet size, x lc is the link capacity, d link is the link delay.

4. The network modeling method based on GNN and Transformer model according to claim 1, characterized in that: The step 4 specifically includes: The network samples in the training set are input into the message passing neural network, and the performance indicators of the corresponding network samples are generated through the GNN cascade module, the Transformer cascade module and the specific indicator readout module. Then the mean square error and coefficient are calculated using the validation set to determine the loss calculation: The mean square error calculates the average of the squares of the differences between the predicted value and the actual value. MSE gives higher weights to larger errors and is very sensitive to outliers. The coefficient of determination, often called R2 or coefficient of determination, measures how well the model explains the total variation; Among them, SS res is the residual sum of squares, SS tot is the total sum of squares, and the value of R2 ranges from 0 to 1. The closer the value is to 1, the better the model fits the data; After the loss calculation is completed, back propagation is performed, and the connection weights are optimized through the selected optimizer and corresponding parameters. The final network model is obtained after 150 rounds of training.

5. The network modeling method based on GNN and Transformer model according to claim 1, characterized in that: The step five specifically includes: The evaluation process criteria are mean square error, mean absolute error, mean absolute percentage error, and coefficient determination; The mean absolute error is the average of the absolute values ​​of the differences between the predicted values ​​and the actual values. The mathematical formula is: The mean absolute percentage error calculates the average of the ratio of the absolute value of the difference between each predicted value and the actual value to its actual value. MAPE is used to measure the accuracy of the forecast. When the actual value range is large, when the actual value is zero or close to zero, MAPE becomes unstable. The mathematical formula is: According to the above indicators, the errors compared with the simulation results are calculated when the input network samples are Poisson, Deterministic, On-Off, A.exponentials, M.exponentials, and Mixed.

Images