A routing optimization method based on traffic matrix completion prediction

Abstract

The application belongs to the field of network management and traffic engineering, and discloses a routing optimization method based on traffic matrix completion prediction, which comprises the following steps: step 1, constructing a routing matrix A , collecting a link load vector at the current moment, step 2, constructing a traffic matrix completion model to output a complete traffic matrix; step 3, constructing a space-time traffic prediction model to output a predicted traffic matrix at the next moment, step 4, connecting the traffic matrix completion model and the space-time traffic prediction model in series, defining a joint loss function to update parameters of the entire end-to-end network; step 5, using the output predicted traffic matrix at the next moment, combining network topology information, pre-evaluating an estimated load state, dynamically adjusting link weights and generating an optimal routing strategy, and finally configuring and issuing to the data layer to realize routing optimization of traffic transmission. The application guarantees the integrity of the traffic matrix, realizes adaptive attention to key links, and makes the completed data highly consistent with future trend prediction.

Description

Technical Field

[0001] This application belongs to the field of network management and traffic engineering, specifically relating to a routing optimization method based on traffic matrix completion prediction. Background Technology

[0002] A traffic matrix is ​​a key indicator representing the amount of traffic between any pair of source and destination nodes in a network. The traffic matrix (TM) provides fundamental data support for network management and plays a crucial role in core tasks such as network traffic engineering, abnormal traffic detection, capacity planning, load balancing, and quality of service assurance.

[0003] As modern networks become increasingly complex and scale continuously, real-time dynamic adjustment and prediction of network conditions have become particularly important. This requires calculating optimal routing strategies based on accurate future traffic forecasts, thereby effectively preventing link congestion and ensuring service requirements are met.

[0004] However, in real-world networks, directly measuring the complete source-destination (OD) flow is extremely costly. SDN devices or NetFlow tools in large-scale networks often experience high CPU utilization, leading to packet sampling loss and making it difficult to obtain a complete and accurate historical traffic matrix.

[0005] Existing matrix completion methods, such as linear or statistical methods (e.g., SVD, PCA), struggle to capture the nonlinear characteristics of network traffic and heavily rely on the amount of observation data, experiencing a sharp performance degradation when significant data is missing. Traditional neural network methods (e.g., LSTM, GCN) typically assume complete and accurate input data, making them incapable of directly handling real-world scenarios with randomly missing values. While network tomography (NT) methods infer traffic using link load, their accuracy is limited due to the rank deficiency of the network tomography equations and reliance on prior assumptions such as gravity models. Summary of the Invention

[0006] To address the problem that high measurement costs and equipment limitations lead to random missing data in the traffic matrix, which severely affects the accuracy of network traffic prediction and the effectiveness of route optimization, this application provides a route optimization method based on traffic matrix completion prediction. This method effectively solves the problem that traditional pure data-driven methods are prone to prediction bias when data is missing.

[0007] To achieve the above objectives, this application employs the following technical solution:

[0008] This application presents a routing optimization method based on traffic matrix completion prediction, applicable to large-scale IP backbone networks. The optimization method includes the following steps:

[0009] Step 1: Constructing a network-aware environment and preprocessing data: The network management platform or central server periodically acquires the topology of the target network and constructs a routing matrix. Collect link load vectors ;

[0010] Step 2: Construct a traffic matrix completion model based on physical constraints and graph attention networks (GANN). This model includes a GAT encoder, a route aggregation layer, and a DNN decoder, using the link load vector collected in Step 1. and routing matrix As input to the flow matrix completion model, a graph attention network is used to capture the nonlinear dependencies between links, and physical tomography equations are used as physical constraints to output the completed flow matrix. ;

[0011] Step 3: Construct a spatiotemporal traffic prediction model based on a two-way perspective. This model comprises a deep neural network model (LSTM2D) with two parallel processing branches, vertical and horizontal, utilizing the completed traffic matrix sequence output in Step 2. As input, the spatiotemporal traffic prediction model captures the temporal dependence characteristics of network traffic. Spatial correlation features It also outputs the predicted flow matrix for the next time step. ;

[0012] Step 4: Construct a unified computational graph, concatenating the flow matrix completion model from Step 2 with the spatiotemporal flow prediction model from Step 3, and define a joint loss function. Update parameters across the entire end-to-end network;

[0013] Step 5: Use the predicted flow matrix for the next time step output in Step 3. By combining network topology information, the estimated load on physical links can be assessed in advance. It dynamically adjusts link weights and generates optimal routing strategies, and finally distributes the configuration to the data layer to achieve routing optimization for traffic transmission.

[0014] A further improvement of this application is that step 1 specifically includes the following steps:

[0015] Step 1.1: The network management platform or central server obtains the topology of the target network and abstracts the topology into a directed graph. ,in, This represents the set of router nodes in the target network, with a quantity of . , This represents the set of physical links in the target network, with a quantity of . ;

[0016] Step 1.2: Calculate the path strategy based on the routing protocol running in the target network and construct the routing matrix. ,in, The total number of all source-destination flow pairs in the target network, the th The source-destination stream passes through the first [stage / path] in the transmission path. Links, then matrix elements This indicates that the value is 0 otherwise.

[0017] Step 1.3: Periodically collect load data from the link interface of the device via the SNMP protocol, and define... The link load vector at time t is ;

[0018] Step 1.4, Complete flow matrix at any time Divided into two parts: the set of observed known flows and the set of unobserved missing flows .

[0019] A further improvement in this application is as follows: Step 2, which constructs a flow matrix completion model based on physical constraints and graph attention networks, is implemented as follows:

[0020] Step 21: Transform the topology of the target network into a link graph structure, and use the link load vector as the initial feature of the nodes;

[0021] Step 22, execute the multi-head graph attention mechanism, utilizing... The layered graph attention layer updates the embedding representation of the link nodes, obtaining the link embedding matrix. ;

[0022] Step 23: Utilize the known routing matrix Learned link embedding matrix Projecting onto the flow space generates a high-dimensional feature representation of the flow. :

[0023]

[0024] Step 24: Analyze the high-dimensional features of the flow using a deep neural network containing two nonlinear units. Decode the data and output the final estimated flow matrix. The forward propagation process is as follows:

[0025]

[0026]

[0027]

[0028] in, These represent the state outputs of the hidden layers in the DNN decoder. The first The weight matrix and bias vector of a fully connected layer. , express Activation function This represents the estimated source-destination flow matrix of the final output;

[0029] Step 25: Define the total loss function that includes mapping error and reconstruction error. :

[0030]

[0031]

[0032]

[0033] in, For hyperparameters, For mapping error, For a known observed flow, Given the estimated values ​​of the flow matrix, For reconstruction error, For inferred unknown flow.

[0034] A further improvement of this application is that step 21 specifically includes the following steps:

[0035] Step 211: Configure the routing matrix Convert to a link graph In this context, nodes represent physical links. If two links carry the same flow, then there is an edge between the nodes.

[0036] Step 212: Transfer the link load vector collected in Step 1 Dimensional reshaping is performed, and these features are used as the initial input features for the GAT encoder. Output link embedding matrix ,in is the feature dimension of the hidden layer.

[0037] A further improvement of this application is that, in step 22, for the first... Layer and first One's attention, , The specific calculation process includes the following steps:

[0038] Step 221: Linear Transformation and Feature Extraction of Input Features By applying the linear transformation matrix, we obtain the transformed features:

[0039]

[0040] in, Represents a node In the Layer Transformed feature vectors under attention heads Represents a node In the The input feature vector of the layer, Indicates the first A learnable weight matrix for each attention head, used to assign dimensions to the input features. Mapping to output dimension ;

[0041] Step 222: Calculate adjacent link nodes and nodes Correlation coefficient between and through Function normalization yields the final attention weights. To calculate the dynamic attention coefficient:

[0042]

[0043]

[0044] Where || denotes the vector concatenation operation. Indicates the first Layer A learnable attention weight vector for each attention head. It is a non-linear activation function. For transpose, Represents a node In the Layer Transformed feature vectors under attention heads Represents a node The set of neighboring nodes, Adjacent link nodes and nodes The correlation coefficient between them;

[0045] Step 223, based on the calculated attention weights The features of neighboring nodes are weighted and summed, and a single-head output is obtained through a non-linear activation function. Finally, all nodes are... The outputs of each node are averaged and aggregated to obtain the node features of the next layer. ;

[0046]

[0047]

[0048] in, For the first Each attention head output, node In the The intermediate feature vector of the layer, It is a non-linear activation function. For the total number of attention heads, the feature passes through After layer update, the final link embedding matrix is ​​obtained. .

[0049] A further improvement in this application is that step 3 specifically includes the following steps:

[0050] Step 31: Complete the flow matrix sequence Reconstructed into two different input views, each used to extract time-dependent features. Spatial correlation features ;

[0051] Step 32: Incorporate time-dependent features Spatial correlation features The data is then stitched together and fused, and the final predicted traffic matrix is ​​generated through mapping using a fully connected layer. :

[0052]

[0053] in, This represents a vector concatenation operation that combines time-dependent features. Spatial correlation features Combined into a single comprehensive feature vector, Indicates a fully connected layer. The next-time predicted flow matrix output by the spatiotemporal flow prediction model;

[0054] Step 33: Define the angle loss function for the spatiotemporal flow prediction model. :

[0055]

[0056] in, Represents the estimated value of the flow matrix of The square of the norm, Represents the predicted flow vector With real future traffic vector Cosine similarity between them This indicates directional error.

[0057] A further improvement in this application is that step 31 specifically includes the following steps:

[0058] Step 311, for the network Each flow in the source-destination flow , construct stream In length of Vertical view sequence within the history time window ;

[0059] Step 312: Input the vertical view sequence into a vertical deep neural network model to extract the time-dependent features of the traffic flow. :

[0060]

[0061] in, Indicates the first Source-destination flow of length Traffic sequence data within a historical time window, Represents the processing function of a Long Short-Term Memory (LSTM) network unit;

[0062] Step 313, for a length of Every moment in the historical time window Build a system that includes all of the network Horizontal view sequence of source-destination flow ;

[0063] Step 314: Input the horizontal view sequence into a horizontal deep neural network model to extract spatial correlation features of traffic flow. :

[0064]

[0065] in, Indicates the first A global traffic snapshot vector consisting of all source-destination flows in the network at each time step.

[0066] A further improvement of this application is that, in step 4, a joint loss function is defined. :

[0067]

[0068]

[0069] in, These are the weighting coefficients. To combine numerical error and cosine similarity loss function, For predicting flow vectors With real future traffic vector The angle between them. During backpropagation, the loss function The generated gradients not only update the parameters of the LSTM2D module, but also propagate forward to the GANN module via the chain rule. The weight parameters in the GANN module are not only subject to physical constraints... The updates are also affected by the demands of downstream forecasting tasks ( This directly impacts the prediction of the target flow matrix, thereby achieving optimization of the flow matrix for the prediction objective.

[0070] A further improvement of this application is that step 5 specifically includes the following steps:

[0071] Step 51: Predicted flow matrix based on the next time step and the routing matrix of the current network Calculate the estimated load on each physical link in the network at the next time step:

[0072] ;

[0073] in, This represents the flow vector of the predicted source-destination flow at the next time step, output by the deep neural network model (LSTM2D). The routing matrix represents the network;

[0074] Step 52: Calculate the future link load based on Step 51. With link physical capacity The relationship is used to generate new routing cost weights for each link. ;

[0075] Step 53: For the predicted load Set the capacity threshold to and the weighting coefficient of congestion warning For predicted load satisfy Increase the weight of the physical links corresponding to the source-destination flows in the traffic matrix;

[0076] Step 54: Based on the updated routing cost weights and network topology, run the routing optimization algorithm to generate the optimal routing strategy;

[0077] Step 55: Convert the calculated optimal routing policy into configuration instructions for the network device and execute them.

[0078] The beneficial effects of this application are:

[0079] This application employs a physically constrained deep learning framework to address the problem of significant random missing data in the network traffic matrix. It introduces the network tomography equation as a physical constraint into the loss function, achieving high-precision traffic completion under sparse observations. This mechanism utilizes physical consistency verification constraints to generate data, ensuring the authenticity of the traffic matrix and effectively solving the prediction bias problem inherent in traditional pure data-driven methods when data is missing.

[0080] This application improves upon traditional link feature extraction methods by replacing static graph convolution with a graph attention network. It dynamically calculates the attention weights of neighboring links to achieve adaptive attention to critical links. This method accurately captures the nonlinear dynamic mapping relationship between link load and traffic, thus solving the problem of traditional methods being slow to respond to sudden changes in network traffic due to their reliance on static topology weights.

[0081] This application employs an end-to-end joint optimization algorithm to collaboratively train traffic completion and prediction tasks, constructing a joint objective function that incorporates both errors. During training, the feedback gradient from the prediction task is used to guide the update of the completion module, making the completed data more aligned with future trend predictions. This mechanism achieves global performance optimization, avoiding the accumulation of completion errors that can mislead routing decisions, as seen in traditional staged methods. Attached Figure Description

[0082] Figure 1 This is a flowchart of the optimization method in this application.

[0083] Figure 2 This is a flowchart outlining the specific steps involved in this application.

[0084] Figure 3 This is a comparison chart of traffic matrix completion performance.

[0085] Figure 4 Performance comparison chart of traffic matrix completion prediction route optimization.

[0086] Figure 5 This is a schematic diagram of the flow matrix completion model (GANN) of this application. Detailed Implementation

[0087] The embodiments of the present invention will be disclosed below with reference to the drawings. For clarity, many practical details will be described in the following description. However, it should be understood that these practical details are not intended to limit the present invention. That is, in some embodiments of the present invention, these practical details are not essential. In addition, for the sake of simplicity, some conventional structures and components will be shown in the drawings in a simple schematic manner.

[0088] like Figure 1 As shown, this application is a routing optimization method based on traffic matrix completion prediction, which is applicable to large-scale IP backbone networks. The routing optimization system mainly runs in the network management center or SDN controller and includes a data acquisition and preprocessing module, a traffic completion module based on GAT (GANN), a traffic prediction module based on LSTM2D, and a routing policy generation module.

[0089] like Figure 2 As shown, the route optimization method based on traffic matrix completion prediction includes the following steps:

[0090] Step 1: Constructing a network-aware environment and preprocessing data: The network management platform or central server periodically acquires the topology of the target network and constructs a routing matrix. Collect link load vectors This involves building the mathematical model foundation for subsequent completion and prediction. Specifically, it includes the following steps:

[0091] Step 1.1: The network management platform or central server obtains the topology of the target network and abstracts the topology into a directed graph. ,in, This represents the set of router nodes in the target network, with a quantity of . , This represents the set of physical links in the target network, with a quantity of . ;

[0092] Step 1.2: Calculate the path strategy based on the routing protocol (such as OSPF, IS-IS, or BGP) running in the target network, and construct the routing matrix. ,in, The matrix elements represent the total number of all source-destination flow pairs in the target network. For a binary variable, the first... The source-destination stream passes through the first [stage / path] in the transmission path. Links, then matrix elements The value is 0 if otherwise indicated, and the routing matrix is ​​used. It describes the spatial mapping relationship of network traffic and satisfies the physical constraint equations. .

[0093] Step 1.3: Periodically collect load data from the link interface of the device via the SNMP protocol (at set time intervals), and define... The link load vector at time t is ;

[0094] Step 1.4: Due to measurement cost limitations, Complete flow matrix at any time Divided into two parts: through measuring tools (such as...) ) Observed set of known flows and the set of unobserved missing flows The core of this invention is the use of link load vectors. and the observed set of known flows To infer the set of unobserved missing flows And predict the forecast flow matrix for the next time step. .

[0095] Step 2: Construct a flow matrix completion model based on physical constraints and graph attention networks (GANN), such as... Figure 5 As shown, the traffic matrix completion model based on physical constraints and graph attention networks (GANN) includes a GAT encoder, a route aggregation layer, and a DNN decoder, using the link load vector collected in step 1. and routing matrix As input to the Flow Matrix Completion Model (GANN), a graph attention network is used to capture the nonlinear dependencies between links, and physical tomography equations are used as physical constraints to output the completed flow matrix. This step uses physical tomography equations as hard constraints to transform the flow completion problem into a feature reconstruction problem with physical consistency verification.

[0096] Step 2: Constructing a flow matrix completion model based on physical constraints and graph attention networks. The implementation steps are as follows:

[0097] Step 21: Transform the topology of the target network into a link graph structure, using the link load vector as the initial feature of the nodes. This includes the following steps:

[0098] Step 211: Configure the routing matrix Convert to a link graph In this context, nodes represent physical links. If two links carry the same flow, then there is an edge between the nodes.

[0099] Step 212: Transfer the link load vector collected in Step 1 Dimensional reshaping is performed, and these features are used as the initial input features for the GAT encoder. The encoder's goal is to output a link embedding matrix containing rich topology information. ,in is the feature dimension of the hidden layer.

[0100] Step 22, execute the multi-head graph attention mechanism, utilizing... The layered graph attention layer updates the embedding representation of the link nodes, obtaining the link embedding matrix. Specifically:

[0101] Step 221: Linear Transformation and Feature Extraction of Input Features By applying a specific linear transformation matrix, the transformed features are obtained:

[0102]

[0103] in, Represents a node In the Layer Transformed feature vectors under attention heads Represents a node In the The input feature vector of the layer, Indicates the first A learnable weight matrix for each attention head, used to assign dimensions to the input features. Mapping to output dimension ;

[0104] Step 222: Calculate adjacent link nodes and nodes Correlation coefficient between and through Function normalization yields the final attention weights. To calculate the dynamic attention coefficient:

[0105]

[0106]

[0107] Where || denotes the vector concatenation operation. Indicates the first Layer A learnable attention weight vector for each attention head. It is a non-linear activation function used to handle negative input values. For transpose, Represents a node In the Layer Transformed feature vectors under attention heads Represents a node The set of neighboring nodes, Represents a node For nodes The importance weight, or attention weight, is a mechanism that allows the model to dynamically focus on the most relevant neighbor links based on the current load. Adjacent link nodes and nodes The correlation coefficient between them;

[0108] Step 223, based on the calculated attention weights The features of neighboring nodes are weighted and summed, and a single-head output is obtained through a non-linear activation function. Finally, all nodes are... The outputs of each node are averaged and aggregated to obtain the node features of the next layer. To perform multi-head feature aggregation;

[0109]

[0110]

[0111] in, For the first Each attention head output, node In the The intermediate feature vector of the layer, It is a non-linear activation function. For the total number of attention heads, the feature passes through After layer update, the final link embedding matrix is ​​obtained. .

[0112] Step 23: Utilize the known routing matrix Learned link embedding matrix Projecting onto the flow space generates a high-dimensional feature representation of the flow. The physical link layer feature information is aggregated to the corresponding source-destination flow layer:

[0113]

[0114] Step 24: Process high-dimensional features of traffic using a deep neural network (DNN) containing two nonlinear units. Decode the data and output the final estimated flow matrix. The forward propagation process is as follows:

[0115]

[0116]

[0117]

[0118] in, These represent the state outputs of the hidden layers in the DNN decoder. The first The weight matrix and bias vector of a fully connected layer. , express Activation functions are used to introduce nonlinearity. This represents the estimated source-destination flow matrix of the final output;

[0119] Step 25: To train the model in the presence of a large number of unknown flows, define a total loss function that includes mapping error and reconstruction error. :

[0120]

[0121]

[0122]

[0123] in, For hyperparameters, Mapping error is used to measure the model's accuracy in predicting known observation flows. The degree of fit, For a known observed flow, Given the estimated values ​​of the flow matrix, For reconstruction error, For the inferred unknown flow, the reconstruction error Utilizing physical constraints to evaluate the completed flow matrix (Including inferred unknown flows) After being projected back to the link load space through the routing matrix, it is compared with the actual link load. By comparing the results, the total loss function is obtained. The total loss function is derived through hyperparameters. The weights of both are balanced to ensure that the completed flow matrix fits the known observation data while satisfying the network tomography equations. Constraints.

[0124] Step 3: Construct a spatiotemporal traffic prediction model based on a two-way perspective. This model comprises a deep neural network model (LSTM2D) with two parallel processing branches, vertical and horizontal, utilizing the completed traffic matrix sequence output in Step 2. As input, the spatiotemporal traffic prediction model captures the temporal dependence characteristics of network traffic. Spatial correlation features It also outputs the predicted flow matrix for the next time step. Specifically, it includes the following steps:

[0125] Step 31: To fully utilize the multidimensional characteristics of traffic data, the completed traffic matrix sequence is... Reconstructed into two different input views, each used to extract time-dependent features. Spatial correlation features Specifically, the steps include the following:

[0126] Step 311: Extract the change sequence of each source-destination (OD) flow on the time axis as a vertical view sequence. For the network... Each flow in the source-destination (OD) flow , construct stream In length of Vertical view sequence within the history time window ;

[0127] Step 312: Input the vertical view sequence into a vertical deep neural network model to extract the time-dependent features of the traffic flow. :

[0128]

[0129] in, Indicates the first Source-destination (OD) flow in length of Traffic sequence data within a historical time window, This represents the processing function of a Long Short-Term Memory (LSTM) network unit, used to capture long-term dependencies in sequential data. This indicates the time-dependent characteristics of the output of the vertical LSTM network, representing the evolution of traffic over time.

[0130] Step 313: Extract global traffic snapshots of the network at each time step as a horizontal view sequence. For a length of... Every moment in the historical time window Build a system that includes all of the network Horizontal view sequence of source-destination (OD) flows ;

[0131] Step 314: Input the horizontal view sequence into a horizontal deep neural network model to extract spatial correlation features of traffic flow. :

[0132]

[0133] in, Indicates the first A global traffic snapshot vector consisting of all source-destination (OD) flows in the network at each time step. The spatial correlation characteristics of the horizontal LSTM output represent the spatial distribution patterns and mutual influences between different OD flows at the same time.

[0134] Step 32: Incorporate time-dependent features Spatial correlation features The data is then stitched together and fused, and the final predicted traffic matrix is ​​generated through mapping using a fully connected layer. :

[0135]

[0136] in, This represents a vector concatenation operation that combines time-dependent features. Spatial correlation features Combined into a single comprehensive feature vector, This represents a fully connected layer, used to map a high-dimensional feature space back to the numerical space of OD flow. The next-time predicted flow matrix output by the spatiotemporal flow prediction model;

[0137] Step 33: To improve the accuracy of prediction and the performance of downstream routing optimization, define the angle loss function of the spatiotemporal traffic prediction model. This loss function not only focuses on the accuracy of the predicted values, but also pays special attention to the consistency of the traffic distribution pattern.

[0138]

[0139] in, Represents the estimated value of the flow matrix of The square of the norm is used to constrain the numerical range of the flow rate. Represents the predicted flow vector With real future traffic vector Cosine similarity between them This represents the direction error; the closer the predicted vector's direction (i.e., the distribution pattern of network traffic) is to the true value, the smaller this term becomes. By introducing this loss function, the model can effectively avoid the problem of only fitting numerical values ​​while ignoring the overall distribution structure of network traffic.

[0140] Step 4: Use the Adam optimizer, based on the total loss function. The parameters of the entire end-to-end network are updated. A unified computation graph is constructed, and the parameters of the flow matrix completion model and the spatiotemporal flow prediction model are updated synchronously using the error gradient of the prediction task through backpropagation. This ensures that the flow matrix generated by the flow matrix completion model not only satisfies physical constraints but also possesses spatiotemporal characteristics beneficial for downstream prediction. A joint loss function is defined. Update parameters for the entire end-to-end network:

[0141]

[0142]

[0143] in, is a weighting coefficient used to adjust the proportion of completion and prediction tasks in the overall optimization objective. These are the parameters to be optimized in the neural network model. Matrix completion provides high-quality data for prediction, and the accuracy of the prediction results, in turn, serves as a supervisory signal, guiding the completion module on how to better fill in missing values. To combine numerical error and cosine similarity loss function, To ensure that the predicted results are not only numerically accurate, but also consistent with the actual flow patterns in terms of spatial distribution, For predicting flow vectors With real future traffic vector The angle between them. By constructing a total loss function that includes both completion error and prediction error. A joint learning mechanism is used to simultaneously optimize network parameters to prevent overfitting. This is to achieve mutual promotion between the two tasks.

[0144] Step 5: Generate routing policy and distribute network configuration based on predicted traffic. Utilize the predicted traffic matrix for the next time step output in Step 3. By combining network topology information, the estimated load on physical links can be assessed in advance. The system dynamically adjusts link weights and generates optimal routing strategies, finally distributing the configuration to the data layer to optimize traffic transmission routing. Specifically, it includes the following steps:

[0145] Step 51: Predicted flow matrix based on the next time step and the routing matrix of the current network By utilizing the forward computation logic of the Network Tomography Equation, the load distribution of future network physical links is derived. Although during the completion phase... Solve It's the inverse problem, but in the prediction application phase, the predicted traffic is known. Solving for future link load It is a forward calculation, calculating the estimated load of each physical link in the network at the next time step:

[0146] ;

[0147] in, This represents the flow vector of the next-time predicted source-destination (OD) flow output by the deep neural network model (LSTM2D). This represents the routing matrix of the network. It defines the mapping relationship between flows and links. This calculates the projected load values ​​for each physical link at future times. This step maps the logical-level traffic forecast back to the physical-level link states to identify potential congestion risks.

[0148] Step 52: Calculate the future link load based on Step 51. With link physical capacity The relationship is used to generate new routing cost weights for each link. ;

[0149] Step 53: For the predicted load Set the capacity threshold to and the weighting coefficient of congestion warning For predicted load satisfy The link increases the weight of the physical link corresponding to the source-destination flow in the traffic matrix; this prediction-based dynamic weight adjustment enables the routing algorithm to sense the upcoming traffic peak and thus actively avoid high-risk links when calculating the path.

[0150] Step 54: Based on the updated routing cost weights and network topology, run the routing optimization algorithm to generate the optimal routing strategy; using the principle of selecting nodes with smaller weight values ​​as routing nodes, adopt the Dijkstra routing algorithm to minimize the weighted cost of the entire network, recalculate the optimal transmission path for all OD flows, and ensure that network traffic has been reasonably allocated to each path before the traffic peak arrives.

[0151] Step 55: Transform the calculated optimal routing policy into configuration commands for the network devices and execute them. Update the routing table or flow table of the network devices based on the optimized path results.

[0152] Simulation Experiment

[0153] The simulation results of the matrix completion method in this application compared with traditional completion methods are as follows: Figure 3 As shown in the figure. The simulation results comparing the performance of the proposed method with four other routing optimization methods are as follows. Figure 4 As shown. By Figure 3 Simulation results show that, compared with traditional gravity model methods, sparse singular value methods, and neural network methods based on Moore-Penrose inverses, the normalized mean absolute error of the completed model in this application is significantly lower than other methods, indicating a clear advantage in completion performance. Figure 4Simulation results show that, under different network resource conditions, the routing optimization method of this application achieves zero blocking when the capacity factor is ≥1, and maintains a high throughput of over 0.88, significantly outperforming equal-cost routing and adaptive routing. This demonstrates that the routing method of this application can significantly avoid network congestion. The above description is merely an embodiment of the present invention and is not intended to limit the invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A route optimization method based on traffic matrix completion prediction, characterized in that: The optimization method includes the following steps: Step 1: Constructing a network-aware environment and preprocessing data: The network management platform or central server periodically acquires the topology of the target network and constructs a routing matrix. Collect link load vectors ; Step 2: Construct a traffic matrix completion model based on physical constraints and graph attention networks. This model includes a GAT encoder, a route aggregation layer, and a DNN decoder, using the link load vector collected in Step 1. and routing matrix As input to the flow matrix completion model, a graph attention network is used to capture the nonlinear dependencies between links, and physical tomography equations are used as physical constraints to output the completed flow matrix. ; Step 3: Construct a spatiotemporal traffic prediction model based on a two-way perspective. This model comprises a deep neural network with two parallel processing branches, vertical and horizontal, utilizing the completed traffic matrix sequence output in Step 2. As input, the spatiotemporal traffic prediction model captures the temporal dependence characteristics of network traffic. Spatial correlation features It also outputs the predicted flow matrix for the next time step. ; Step 4: Construct a unified computational graph, concatenating the flow matrix completion model from Step 2 with the spatiotemporal flow prediction model from Step 3, and define a joint loss function. Update parameters across the entire end-to-end network; Step 5: Use the predicted flow matrix for the next time step output in Step 3. Based on network topology information and the traffic matrix at the next time step, the estimated load on physical links is pre-assessed. The system dynamically adjusts link weights and generates optimal routing strategies, then distributes the configuration to the data layer to optimize traffic transmission routing. This process includes the following steps: Step 51: Predicted flow matrix based on the next time step and the routing matrix of the current network Calculate the estimated load on each physical link in the network at the next time step: ； in, This represents the flow vector of the predicted source-destination flow at the next time step, as output by the deep neural network model. The routing matrix represents the network; Step 52: Calculate the future link load based on Step 51. With link physical capacity The relationship is used to generate new routing cost weights for each link. ; Step 53: For the predicted load Set the capacity threshold to and the weighting coefficient of congestion warning For predicted load satisfy Increase the weight of the physical links corresponding to the source-destination flows in the traffic matrix; Step 54: Based on the updated routing cost weights and network topology, run the routing optimization algorithm to generate the optimal routing strategy; Step 55: Convert the calculated optimal routing policy into configuration instructions for the network device and execute them.

2. The routing optimization method based on traffic matrix completion prediction according to claim 1, characterized in that: Step 1 specifically includes the following steps: Step 1.1: The network management platform or central server obtains the topology of the target network and abstracts the topology into a directed graph. ,in, This represents the set of router nodes in the target network, with a quantity of . , This represents the set of physical links in the target network, with a quantity of . ; Step 1.2: Calculate the path strategy based on the routing protocol running in the target network and construct the routing matrix. ,in, The total number of all source-destination flow pairs in the target network, the th The source-destination stream passes through the first [stage / path] in the transmission path. Links, then matrix elements Otherwise, it is 0; Step 1.3: Periodically collect load data from the link interface of the device via the SNMP protocol, and define... The link load vector at time t is ; Step 1.4, Complete flow matrix at any time Divided into two parts: the set of observed known flows and the set of unobserved missing flows .

3. The routing optimization method based on traffic matrix completion prediction according to claim 1, characterized in that: Step 2: Constructing a flow matrix completion model based on physical constraints and graph attention networks. The implementation steps are as follows: Step 21: Transform the topology of the target network into a link graph structure, and use the link load vector as the initial feature of the nodes; Step 22, execute the multi-head graph attention mechanism, utilizing... The layered graph attention layer updates the embedding representation of the link nodes, obtaining the link embedding matrix. ; Step 23: Utilize the known routing matrix Learned link embedding matrix Projecting onto the flow space generates a high-dimensional feature representation of the flow. : Step 24: Analyze the high-dimensional features of the flow using a deep neural network containing two nonlinear units. Decode the data and output the final estimated flow matrix. The forward propagation process is as follows: in, These represent the state outputs of the hidden layers in the DNN decoder. The first The weight matrix and bias vector of a fully connected layer. , express Activation function This represents the estimated source-destination flow matrix of the final output; Step 25: Define the total loss function that includes mapping error and reconstruction error. : in, For hyperparameters, For mapping error, For a known observed flow, Given the estimated values ​​of the flow matrix, For reconstruction error, For inferred unknown flow.

4. The routing optimization method based on traffic matrix completion prediction according to claim 3, characterized in that: Step 21 specifically includes the following steps: Step 211: Configure the routing matrix Convert to a link graph In this context, nodes represent physical links. If two links carry the same flow, then there is an edge between the nodes. Step 212: Transfer the link load vector collected in Step 1 Dimensional reshaping is performed, and these features are used as the initial input features for the GAT encoder. Output link embedding matrix ,in is the feature dimension of the hidden layer.

5. The routing optimization method based on traffic matrix completion prediction according to claim 4, characterized in that: In step 22, for the first Layer and first One's attention, , The specific calculation process includes the following steps: Step 221: Linear Transformation and Feature Extraction of Input Features By applying the linear transformation matrix, we obtain the transformed features: in, Represents a node In the Layer Transformed feature vectors under attention heads Represents a node In the The input feature vector of the layer, Indicates the first A learnable weight matrix for each attention head, used to assign dimensions to the input features. Mapping to output dimension ; Step 222: Calculate adjacent link nodes and nodes Correlation coefficient between and through Function normalization yields the final attention weights. To calculate the dynamic attention coefficient: Where || denotes the vector concatenation operation. Indicates the first Layer A learnable attention weight vector for each attention head. It is a non-linear activation function. For transpose, Represents a node In the Layer Transformed feature vectors under attention heads Represents a node The set of neighboring nodes, Adjacent link nodes and nodes The correlation coefficient between them; Step 223: Based on the calculated attention weights The features of neighboring nodes are weighted and summed, and a single-head output is obtained through a non-linear activation function. Finally, all nodes are... The outputs of each node are averaged and aggregated to obtain the node features of the next layer. ; in, For the first Each attention head output, node In the The intermediate feature vector of the layer, It is a non-linear activation function. For the total number of attention heads, the feature passes through After layer update, the final link embedding matrix is ​​obtained. .

6. The routing optimization method based on traffic matrix completion prediction according to claim 1, characterized in that: Step 3 specifically includes the following steps: Step 31: Complete the flow matrix sequence Reconstructed into two different input views, each used to extract time-dependent features. Spatial correlation features ; Step 32: Incorporate time-dependent features Spatial correlation features The data is then stitched together and fused, and the final predicted traffic matrix is ​​generated through mapping using a fully connected layer. : in, This represents a vector concatenation operation that combines time-dependent features. Spatial correlation features Combined into a single comprehensive feature vector, Indicates a fully connected layer. The next-time predicted flow matrix output by the spatiotemporal flow prediction model; Step 33: Define the angle loss function for the spatiotemporal flow prediction model. : in, Represents the estimated value of the flow matrix of The square of the norm, Represents the predicted flow vector With real future traffic vector Cosine similarity between them This indicates directional error.

7. The routing optimization method based on traffic matrix completion prediction according to claim 6, characterized in that: Step 31 specifically includes the following steps: Step 311, for the network Each flow in the source-destination flow , construct stream In length of Vertical view sequence within the history time window ; Step 312: Input the vertical view sequence into a vertical deep neural network model to extract the time-dependent features of the traffic flow. : in, Indicates the first Source-destination flow of length Traffic sequence data within a historical time window, Represents the processing function of a Long Short-Term Memory (LSTM) network unit; Step 313, for a length of Every moment in the historical time window Build a system that includes all of the network Horizontal view sequence of source-destination flow ; Step 314: Input the horizontal view sequence into a horizontal deep neural network model to extract spatial correlation features of traffic flow. : in, Indicates the first A global traffic snapshot vector consisting of all source-destination flows in the network at each time step.

8. The routing optimization method based on traffic matrix completion prediction according to claim 1, characterized in that: In step 4, the joint loss function is defined. : in, These are the weighting coefficients. To combine numerical error and cosine similarity loss function, For predicting flow vectors With real future traffic vector The angle between them.

Images