Electric power information network performance prediction method and system based on graph neural network

Abstract

The invention discloses an electric power information network performance prediction method and system based on a graph neural network, and relates to the technical field of network performance estimation. The problems of difficulty in adapting to network configuration change, lagging of large-scale network performance estimation and insufficient prediction precision caused by lack of flow-level statistics exist in the prior art. The method comprises the following steps: constructing a graph neural network model, and extracting network topology features through the graph neural network model; calculating the mutual influence between the network flows based on a time sequence technology; extracting node features related to the target flow through the topological information of the target flow; calculating correlation degrees between all streams and the target stream; and fusing the features to obtain a network performance prediction result. According to the technical scheme, the accuracy of performance estimation is improved, universality and adaptability suitable for different types of networks are achieved, a real-time performance estimation result can be provided for network operation and maintenance personnel, and network troubleshooting and optimization are facilitated.

Description

Technical Field

[0001] The present invention relates to the technical field of network performance estimation, and specifically to a method and system for predicting the performance of a power information network based on a graph neural network. Background Art

[0002] With the deepening of the construction of smart grids, the power information network carries the real-time data transmission of key services such as relay protection, security automation devices, and dispatching automation. Its network performance is directly related to the safe and stable operation of the power grid. In order to achieve effective optimization of network performance, accurate network performance estimation plays a crucial role. Such estimation is particularly useful for studying new protocols and mechanisms, as it allows administrators to evaluate their performance before actually deploying them to the production network.

[0003] Most traditional network performance estimation methods are achieved through network deduction. However, these methods have some problems. First, they cannot provide accurate estimation results when facing changes in network configurations (such as topology, traffic patterns, traffic management mechanisms, etc.), so they lack generality. Second, most existing methods focus on packet-level performance estimation and lack attention to flow-level statistics, which leads to the inability to estimate the performance of large-scale networks in a timely manner. In addition, there is still a large room for improvement in the accuracy of the current methods for predicting the performance of power information networks. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by the present invention is to solve the problems existing in traditional methods in terms of network configuration changes, flow-level statistics, and prediction accuracy.

[0006] To solve the above technical problems, the present invention provides the following technical solution: A method for predicting the performance of a power information network based on a graph neural network, which includes the following steps: Construct a graph neural network model, and extract the topological features of the power information network through the graph neural network model; calculate the mutual influence between network flows based on time series technology; extract node features related to the target flow through the topological information of the target flow; calculate the correlation degree between all flows and the target flow; fuse the features to obtain the network performance prediction result.

[0007] This technical solution deeply integrates graph neural networks with time series technology. Through dynamic-static collaborative modeling and target-oriented feature purification, it solves the problem of insufficient prediction accuracy caused by the coupling of topological complexity, traffic time-variability, and business relevance in power information networks. Specifically: directly model the complex topological structure of the power information network (such as node connection relationships and path dependencies) through graph neural networks to capture spatial correlations that are difficult to describe by traditional methods (such as cascading effects between relay protection nodes in the power grid). Introduce time series technologies (such as Transformer, LSTM, etc.) to model the time-varying correlations of network flows (such as traffic bursts and periodic fluctuations) to solve the problem of insufficient adaptability of traditional static traffic analysis to real-time scenarios. Screen key node features through target flow path information (such as the transmission path of power service flows), suppress noise interference (such as redundant data of irrelevant nodes), and avoid the problem of feature dilution caused by "equal treatment of all nodes" in traditional methods. Quantify the correlation between all flows and the target flow (such as bandwidth competition and routing overlap) to achieve fine-grained traffic impact modeling (such as the squeezing effect of high-priority service flows on the target flow), which is superior to traditional fixed-rule-based traffic priority allocation. Non-linearly fuse topological features (spatial dimension) and traffic interaction features (time dimension) (such as MLP, attention mechanism) to uniformly model the spatio-temporal coupling effect (such as traffic redistribution caused by topological changes), and solve the problem of error accumulation in traditional step-by-step modeling. Support real-time changes in network configuration (such as node failures and routing adjustments) through the topological self-adaptation ability of graph neural networks and the traffic dynamic modeling of time series technologies, which is superior to traditional deduction methods that rely on fixed topological assumptions. Reduce redundant computational volume through feature screening and correlation calculation to meet the real-time requirements of power information networks (such as microsecond-level latency-sensitive services).

[0008] As a preferred solution of a power information network performance prediction method based on graph neural networks according to the present invention, wherein: the topological features of the power information network include the connection relationship between nodes and topological features.

[0009] The extraction of topological features of the power information network includes that the topological feature extractor uses a graph neural network to extract topological features, encodes the network topological information using the connection pattern and structural features between nodes, the graph neural network uses a graph convolutional network for extraction, models the network topology as a graph structure, extracts edge features in the topological information and node features in the state information respectively, and represents them as , where V represents the set of all nodes in the graph, E represents the set of all edges, each node is represented as a feature vector , the edge between node i and node j is represented as , message passing in the graph convolutional network is performed at each layer, and is represented as,

[0010] Among them, represents the th feature vector of the node, represents the activation function, represents the set of neighbor nodes of node is the normalization factor for normalizing the edge , represents the weight matrix of the

[0011] This technical solution can realize the end-to-end dynamic modeling of the power information network topology, solve the limitations of traditional methods in capturing high-order relationships, dynamic adaptability, and feature expression efficiency, and provide a basis for high-precision and low-latency topology feature extraction for power network performance prediction, especially suitable for complex topology and dynamic traffic coupling scenarios in smart grids.

[0012] As a preferred solution of the power information network performance prediction method based on graph neural network described in the present invention, among them: the calculation of the mutual influence between network flows includes that the encoder receives the feature vectors of the flows as inputs, uses the multi-head attention mechanism to extract the mutual influence relationships between the flows, and models the flows as a sequence , where represents the number of flows in the network, and each flow is represented by the feature vector .

[0013] The encoder uses the multi-head attention mechanism to linearly transform the input feature vectors into query, key, and value vectors, calculates the dot product of the query and the key to obtain the attention weights, which are used to measure the correlation between each flow and other flows, multiplies the attention weights by the value vectors to obtain the weighted flow feature representations, and the multi-head attention mechanism is expressed as

[0014] where respectively represent the query, key, and value vectors generated by linear transformation, corresponding to the flow features in different subspaces, represents the dimension of the input features.

[0015] The flow features are mapped to multiple subspaces (corresponding to multiple groups of query-key-value vectors) through the multi-head attention mechanism. Each head independently captures the correlations of different dimensions among the flows (such as traffic volume, time interval, protocol type, etc.), and finally fuses the multi-dimensional feature representations. This avoids the limitations of a single perspective and enhances the richness of feature expression. For example, it can simultaneously identify different influence patterns of "high-traffic long connections" and "low-traffic short connections". The original flow features are decoupled into query, key, and value vectors through linear transformation, corresponding to "query target", "matching key-value", and "information transmission carrier" respectively, clarifying the logical roles of the interactions among the flows; structurally separating the feature uses enables the model to more accurately learn the dependencies among the flows (for example, when a certain flow acts as a "query", it matches the relevance of other flows through the "key" and then aggregates effective information through the "value"). Through dot product operation and normalization to generate attention weights, which measure the correlation between any two flows in real time. The larger the weight value, the more significant the influence between the flows, and the weights between the flows are adaptively adjusted. For example, in a power grid, it can dynamically identify the associations between key flows (such as relay protection signal flows) and other flows and suppress the interference of irrelevant traffic. The flows are modeled as time series, and combined with the sensitivity of the attention mechanism to the sequence order, the time evolution pattern of the dependencies among the flows is captured (such as the attenuation of the influence of a certain flow on the target flow at adjacent time steps), which is applicable to scenarios where the traffic in the power grid fluctuates over time (such as load changes during morning and evening rush hours), improving the accuracy of dynamic performance prediction. By multiplying the attention weights with the value vectors, only the information highly relevant to the current flow is retained, filtering out the interference of low-correlation flows. In a large-scale network, redundant calculations are reduced and the computational efficiency is improved. For example, background traffic with little influence on the target flow will be assigned low weights to avoid "noisy" features dominating the prediction. A scaling factor is introduced to normalize the attention scores, avoiding the vanishing of the softmax gradient due to overly large dot product values in the high-dimensional feature space and stabilizing the training process, especially applicable to scenarios where the flow feature dimensions are relatively high (such as including multiple attributes such as traffic rate, protocol type, source and destination addresses, etc.). The multi-head attention mechanism can learn the non-linear and asymmetric dependencies among the flows, breaking through the limitations of traditional linear models and adapting to complex traffic interaction scenarios in the power grid (such as the competitive access of multi-source concurrent traffic to the same device).

[0016] As a preferred solution of the power information network performance prediction method based on graph neural network described in the present invention, wherein: the extraction of node features related to the target flow includes purifying topological features based on a topological feature purifier, purifying node features related to the target flow, and the topological feature purifier identifies the relationship between each node and the target flow through a node scoring mechanism, and extracts the node features related to the target flow from the output of the topological feature extractor through a weight multiplication operation.

[0017] This technical solution systematically solves the "information overload" problem of traditional graph neural networks in power network analysis through target-oriented node feature purification, achieving dual optimization of feature quality and computational efficiency.

[0018] As a preferred embodiment of the power information network performance prediction method based on graph neural network according to the present invention, wherein: the purification of topological features by the topological feature purifier includes outputting, by the topological feature extractor, a feature matrix representing the entire topological structure, where each row represents the feature vector of a node, and the feature vector contains information about the node's direct and indirect nodes. The node weights related to the target flow need to be set to 1 through the node scoring mechanism, and the weights of irrelevant nodes are set to 0.

[0019] The node features related to the target flow are extracted by performing an operation of multiplying the scores with the topological features. By multiplying the node scores with the topological feature matrix, the feature values of the nodes related to the target flow are retained, and the feature values of the irrelevant nodes are suppressed to zero.

[0020] The purification of the node features related to the target flow is expressed as

[0021] where x is the output of the topological feature extractor, representing a node feature, is the path of the target flow.

[0022] The technical solution purifies the topological features through a binary node scoring mechanism and matrix multiplication operations, completely eliminating the interference of irrelevant nodes (such as edge devices on non-target flow paths in the power network), enabling subsequent predictions to rely only on topological elements directly affecting the performance of the target flow and improving feature purity. The scoring mechanism encodes the prior knowledge of the "target flow path" as a feature selection rule, enabling the purified feature matrix to explicitly contain path structure information (such as node connection order, hop count, etc.). Compared with unsupervised feature selection methods (such as the attention mechanism automatically learning weights), this rule has a clear physical meaning and avoids the model learning false associations that do not conform to the business logic of the power network. The weight multiplication operation transforms the original feature matrix into a sparse matrix, significantly reducing the computational complexity of subsequent graph neural network layers. The non-zero nodes after purification directly correspond to the physical path of the target flow, which can assist in quickly locating network fault nodes. When the topology of the power network changes (such as a line switch causing the target flow path to change to ), only the scoring matrix needs to be regenerated , and there is no need to retrain the topological feature extractor, meeting the stringent real-time requirements of the power system.

[0023] As a preferred solution of a power information network performance prediction method based on a graph neural network according to the present invention, wherein: calculating the correlation degree between all flows and the target flow includes calculating the correlation score between each flow and the target flow through a scoring mechanism, measuring the similarity or correlation degree between two flows through the correlation score, and performing an operation of multiplying the correlation score by the flow features weighted to extract features highly correlated with the target flow.

[0024] The scoring mechanism is expressed as

[0025] wherein respectively represent the features of all flows and the target flow in the network.

[0026] Through the cosine similarity scoring mechanism and weighted feature extraction in this technical solution, the efficient quantification of the correlation degree between flows and the directional extraction of highly correlated features are realized. While ensuring the accuracy of calculating the correlation degree between flows, the optimal balance of calculation cost, generalization ability and interpretability is achieved.

[0027] As a preferred solution of a power information network performance prediction method based on a graph neural network according to the present invention, wherein: obtaining the network performance prediction result by fusing features includes fusing the features to obtain a feature representation, and predicting the target flow based on the feature representation.

[0028] The prediction is expressed as

[0029] wherein represents the output of the topological feature purifier, represents the output of the inter-flow feature purifier, respectively represent the weight and the bias.

[0030] The structural information output by the topological feature purifier ( , such as node connection relationships, path features) and the dynamic information output by the inter-flow feature purifier ( , such as traffic correlation, temporal dependence) are integrated through a concatenation operation to form a joint representation; after fusion, the collaborative modeling of "structure-dynamics" is realized; for example, when the load of a certain switch node is too high (topological feature), and multiple flows related to it simultaneously have burst traffic (inter-flow feature), the model can more accurately predict the performance degradation. Through sharing the weight matrix Map features from different sources to a unified latent space, solve the problems of feature dimension and semantic differences, avoid training instability caused by scale differences of different features, and ensure balanced weights of the two types of features during the fusion process. Adopt two-layer ReLU non-linear transformation to enable the model to learn high-order non-linear relationships between features; for example: the first layer of ReLU captures basic interactions (such as the direct association between "node congestion" and "traffic surge"), and the second layer of ReLU models compound effects (such as the non-linear amplification of "node congestion + multi-flow competition" on latency). By and two weight transformations, combine low-order features into high-order semantic units, for example may represent "critical path status", and then combine with to generate a more abstract representation of "network performance bottleneck". The ReLU activation function alleviates the vanishing gradient problem and makes the training of deep networks more stable.

[0031] Another object of the present invention is to provide a power information network performance prediction system based on a graph neural network, which can improve the accuracy and comprehensiveness of network performance estimation by comprehensively considering the topological characteristics of the power information network and the time-variability of traffic, and solves the problems existing in the traditional methods in terms of network configuration changes, flow-level statistics, and prediction accuracy.

[0032] To solve the above technical problems, the present invention provides the following technical solution: A power information network performance prediction system based on a graph neural network, comprising: a topological feature extraction module, an inter-flow feature extraction module, a target flow topological feature extraction module, a target flow inter-flow feature extraction module, and a fusion module.

[0033] The topological feature extraction module is used to calculate the topological features of the power information network and can learn the representation of nodes.

[0034] The inter-flow feature extraction module is used to calculate the mutual influence between network flows and consider the time-variability of traffic.

[0035] The target flow topological feature extraction module is used to extract node features related to the target flow.

[0036] The target flow inter-flow feature extraction module is used to calculate the correlation degree between all flows and the target flow and correct the inter-flow features.

[0037] The fusion module is used to fuse features and obtain the network performance estimation result.

[0038] The topological feature extraction module focuses on network topology structure modeling. By learning node representations (such as node connection relationships and topological roles) through graph neural networks, it realizes the deep encoding of static network structures. The inter-flow feature extraction module analyzes traffic dynamics and uses time series techniques to capture the time-variability between flows (such as the temporal dependence of traffic fluctuations), forming a "structure-dynamics" two-dimensional modeling with the topology module. The topology and inter-flow feature extraction modules can run in parallel, reducing the overall processing latency and supporting the real-time synchronous analysis of multi-source data in power networks. The target flow topological feature extraction module filters out node features irrelevant to a specific target flow (such as non-path nodes), only retaining key nodes (such as source / target nodes and routing nodes), reducing the processing of invalid nodes, and improving feature relevance while reducing the computational load. The target flow inter-flow feature extraction module suppresses background traffic with low correlation to the target flow through correlation calculation (such as cosine similarity), enhancing the feature weights of highly correlated flows (such as collaborative flows in the same business cluster); it compresses the flow feature dimension from the full set of flows to a subset of highly correlated flows, reducing the input noise for the subsequent fusion module. The fusion module integrates topological structure features (such as path connectivity) and inter-flow dynamic features (such as traffic interaction patterns), reducing the prediction error rate. Nonlinear transformation can improve the prediction accuracy in complex scenarios. When the network topology changes (such as path switching caused by a link failure), the target flow topological feature extraction module can update the relevant node set in real time without retraining the entire system, meeting the real-time requirements of the power system fault rapid recovery scenario.

[0039] A computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the above-mentioned power information network performance prediction method based on graph neural networks are implemented.

[0040] A computer-readable storage medium stores a computer program thereon. When the computer program is executed by a processor, the steps of the above-mentioned power information network performance prediction method based on graph neural networks are implemented.

[0041] Advantages of the present invention: The advantage of this method lies in improving the accuracy of performance estimation, and having the generality and adaptability applicable to different types of networks, which can provide real-time performance estimation results for network operation and maintenance personnel, contributing to network fault troubleshooting and optimization. Description of the Drawings

[0042] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0043] Figure 1 The overall flowchart of a power information network performance prediction method based on graph neural network provided in the first embodiment of the present invention.

[0044] Figure 2 The schematic diagram of the topology feature extractor of a power information network performance prediction method based on graph neural network provided in the first embodiment of the present invention.

[0045] Figure 3 The schematic diagram of the flow feature extractor in a power information network performance prediction method based on graph neural network provided in the first embodiment of the present invention.

[0046] Figure 4 The schematic diagram of the topology feature purification extractor of a power information network performance prediction method based on graph neural network provided in the first embodiment of the present invention.

[0047] Figure 5 The schematic diagram of the flow feature purifier of a power information network performance prediction method based on graph neural network provided in the first embodiment of the present invention.

[0048] Figure 6 The system framework diagram of a power information network performance prediction system based on graph neural network provided in the second embodiment of the present invention.

[0049] Figure 7 The bar chart of performance evaluation under different network scales of different methods in a power information network performance prediction method based on graph neural network provided in the third embodiment of the present invention.

[0050] Figure 8 The bar chart of performance evaluation under different network topologies of different methods in a power information network performance prediction method based on graph neural network provided in the third embodiment of the present invention.

[0051] Figure 9 The curve graph of the influence of the feature purification module on the overall method accuracy in a power information network performance prediction method based on graph neural network provided in the third embodiment of the present invention. Detailed implementation manners

[0052] To make the above objects, features and advantages of the present invention more obvious and understandable, the following will describe the detailed implementation manners of the present invention with reference to the accompanying drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of them. All other embodiments obtained by those of ordinary skill in the art without creative efforts based on the embodiments of the present invention shall fall within the protection scope of the present invention.

[0053] Embodiment 1 Refer toFigures 1 - 5 , which is an embodiment of the present invention, provides a method for predicting the performance of a power information network based on a graph neural network, including the following steps: S1: Construct a graph neural network model, and extract network topology features through the graph neural network model.

[0054] The principle of the topology feature extractor is as Figure 2 shown. The topology feature extractor uses GNN to extract topology features. It can encode network topology information using the connection patterns and structural features between nodes. GNN can effectively learn the representations of nodes, including neighbor information and the positions of nodes in the topology structure.

[0055] This case uses a classic implementation in GNN - the Graph Convolutional Network (GCN). Of course, this part can also be assembled by other GNNs. We model the network topology as a graph structure, denoted as , where V represents the set of all nodes in the graph, E represents the set of all edges, and each node is represented as a feature vector , and the edge between node i and node j is represented as . Message passing in the graph convolutional network is performed at each layer, denoted as,

[0056] where, represents the feature vector of the th node in the th layer, represents the activation function, represents the set of neighbor nodes of node , is the normalization factor for normalizing the edge , represents the th layer's weight matrix.

[0057] S2: Calculate the mutual influence between network flows based on time series technology.

[0058] The schematic diagram of the inter - flow feature extractor is as Figure 3 shown. The inter - flow feature extractor uses a classic encoder - decoder architecture for flow feature extraction.

[0059] Furthermore, the time series technology adopts Transformer, considering the time - variability between flows to accurately describe the mutual influence between network flows.

[0060] Furthermore, the encoder part of the present invention adopts a Transformer encoder structure without positional encoding. Calculating the mutual influence between network flows includes: the encoder receives the feature vectors of the flows as inputs, uses the multi-head attention mechanism to extract the mutual influence relationships between the flows, and models the flows as a sequence , where represents the number of flows in the network, and each flow is represented by a feature vector .

[0061] The encoder uses the multi-head attention mechanism to linearly transform the input feature vectors and convert them into query ( ), key ( ), and value ( ) vectors. By calculating the dot product of the query and the key, the attention weights are obtained, which are used to measure the correlation between each flow and other flows. Then, the attention weights are multiplied by the value vectors to obtain the weighted flow feature representation. In this way, the encoder can capture the important correlation information between each flow and other flows.

[0062] The multi-head attention mechanism is expressed as

[0063] , where represents the flow features in different subspaces in the high-dimensional space, and represents the dimension of the input features.

[0064] The decoder part is a simple multi-layer perceptron that extracts the inter-flow features from the high-dimensional space feature representation output by the encoder.

[0065] S3: Extract the node features related to the target flow through the topological information of the target flow.

[0066] Extracting the node features related to the target flow includes: purifying the topological features based on the topological feature purifier. The schematic diagram of the topological feature purifier is as shown in Figure 4 . The topological feature purifier identifies the relationship between each node and the target flow through the node scoring mechanism, and uses the weight multiplication operation to extract the node features related to the target flow from the output of the topological feature extractor.

[0067] In the present invention, the output of the topological feature extractor is a feature matrix representing the entire topological structure, where each row represents the feature vector of a node, and this feature vector contains the information of the direct and indirect nodes of this node. However, for the target flow, we only focus on the nodes related to it. Therefore, we only need to set the weights of the nodes related to the target flow to 1 and the weights of the unrelated nodes to 0 through the node scoring mechanism.

[0068] Then, through the operation of multiplying the scores with the topological features, the node features related to the target flow are extracted. By performing a multiplication operation on the node scores and the topological feature matrix, the eigenvalue of the node related to the target flow can be retained, while the eigenvalue of the unrelated node is suppressed to zero.

[0069] The purified node features related to the target flow are represented as

[0070] where x is the output of the topological feature extractor, representing a node feature. is the path of the target flow.

[0071] Furthermore, the topological information of the target flow is used to extract the node features related to the target flow to enhance the accuracy of network performance estimation.

[0072] S4: Calculate the correlation degree between all flows and the target flow.

[0073] The schematic diagram of the flow feature purifier is as Figure 5 shown. Calculate the correlation scores between all flows and the target flow. We use these scores to correct the encoder output and obtain more fine-grained features for prediction.

[0074] Furthermore, the correlation degree is used to represent the correlation degree between all flows and the target flow, and based on this, the features between flows are corrected to improve the accuracy of network performance estimation.

[0075] In the present invention, the output of the inter-flow feature purifier is a feature matrix representing all flows in the network. Each row represents the feature vector of a network flow, and these feature vectors contain the information of the target flow and other flows. However, for the target flow, we are more concerned about the features of the flows highly correlated with it. Therefore, we need to calculate the correlation degree between all flows and the target flow through a scoring mechanism.

[0076] Through the scoring mechanism, we can calculate the correlation scores between each flow and the target flow. These scores can be used to measure the similarity or correlation degree between two flows. Then, we perform an operation of weighted multiplication on these scores and the flow features to extract the features highly correlated with the target flow. In this way, we can obtain a more accurate and targeted feature representation, which pays more attention to the information related to the target flow.

[0077] The scoring mechanism is represented as

[0078] where respectively represent the features of all flows and the target flow in the network.

[0079] S5: Obtain the prediction result of the fusion feature acquisition network.

[0080] By fusing the features of S3 and S4, a more comprehensive and rich feature representation can be obtained for the prediction of the target flow.

[0081] By fusing the features of S3 and S4, we can obtain a more comprehensive and rich feature representation for the prediction of the target flow.

[0082] The feature fusion method adopted in this case is to splice these features, and then, the fused features are sent to a Multilayer Perceptron (MLP) for processing. In this case, we use a three-layer MLP structure, where each neuron has weights and biases, and the ReLU function is used as the activation function to perform a non-linear transformation on the input.

[0083] The prediction is expressed as,

[0084] where, represents the output of the topological feature purifier, represents the output of the inter-flow feature purifier, represent weights and biases respectively.

[0085] Embodiment 2 Refer to Figure 6 , an embodiment of the present invention, provides a system for predicting the performance of a power information network based on a graph neural network. A system for predicting the performance of a power information network based on a graph neural network includes a topological feature extraction module, an inter-flow feature extraction module, a target flow topological feature extraction module, a target flow inter-flow feature extraction module, and a fusion module.

[0086] The topological feature extraction module is used to calculate the topological features of the power information network and can learn the representation of nodes.

[0087] The inter-flow feature extraction module is used to calculate the mutual influence between network flows and consider the time-varying nature of traffic.

[0088] The target flow topological feature extraction module is used to extract node features related to the target flow.

[0089] The target flow inter-flow feature extraction module is used to calculate the correlation degree between all flows and the target flow and correct the inter-flow features.

[0090] The fusion module is used to fuse features to obtain the network performance estimation result.

[0091] If the above-mentioned functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs.

[0092] The logic and / or steps represented in the flowchart or described in other ways herein, for example, can be considered as a definite sequence list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other systems that can fetch instructions from the instruction execution system, apparatus, or device and execute the instructions), or used in combination with these instruction execution systems, apparatus, or devices. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device.

[0093] More specific examples (non-exhaustive list) of computer-readable media include the following: electrical connection parts with one or more wirings (electronic devices), portable computer disk cartridges (magnetic devices), random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memories), fiber optic devices, and portable compact disc read-only memories (CDROM). Additionally, the computer-readable medium can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other media, then editing, interpreting, or otherwise processing it as appropriate, and then storing it in a computer memory.

[0094] It should be understood that each part of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one or a combination of the following techniques well known in the art can be used: discrete logic circuits having logic gate circuits for implementing logical functions on data signals, application specific integrated circuits having appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), etc.

[0095] Embodiment 3 Referring to Figures 7 - 9 , in this embodiment, in order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments. In order to demonstrate the superiority of the method we proposed, two types of experimental tests were conducted: generalization experiments and ablation experiments.

[0096] First, generalization experimental tests were carried out, and the results are shown in Figure 7 and Figure 8 . Figure 7 shows the comparison results of the mean absolute percentage error of two models under different network topologies (NsfNet, Geant2) and different numbers of nodes (25, 30, 35). The errors of the flow vision network model proposed by this method are all less than those of the routing network model, and the minimum error is only 49.20% of the error of the routing network model. This means that the streaming network model can more accurately estimate the network performance under various scale network topologies, indicating that the streaming network model shows better stability than the routing network model in different network topologies. Therefore, it can be concluded that the streaming network model has stronger general ability and stability under various network topologies. Further, Figure 8Shows the performance of this method under five typical traffic scenarios (on-off traffic, constant bit rate traffic, modulated traffic, autocorrelated traffic, and mixed traffic). The bar chart shows the comparison of the predicted mean absolute percentage error of different models under different topology and traffic combinations. First, in the on-off traffic scenario, the FlowSeer model shows significant performance advantages compared to all baseline models (including the routing network model and the routing network - Erlang model). Specifically, in the NsfNet topology, the proposed streaming network model we developed achieved a 21.26% performance improvement, while in the Geant2 topology, the performance improvement was even more significant, reaching 28.9%. These results indicate that the FlowSeer network model can more accurately predict network performance in the on-off traffic scenario. In the constant bit rate traffic scenario, the FlowSeer network model outperformed the routing network - Erlang model, which means that the FlowSeer network model can achieve better performance than the baseline model in the constant bit rate traffic scenario. In the modulated traffic scenario, the FlowSeer network model demonstrated significant performance improvement, with a performance increase of at least 28.32%. This further proves the ability of the FlowSeer model to more accurately predict network performance in the modulated traffic scenario. In the autocorrelated traffic scenario, the FlowSeer network model had better performance than other benchmark models, with a performance increase of up to 21.26%. In the mixed traffic scenario, the FlowSeer network model also outperformed other benchmark models. In all scenarios, the FlowSeer network model achieved significant performance improvement, demonstrating its generalization ability and robustness.

[0097] Secondly, ablation experiments were conducted, and the results are as Figure 9 shown. The curve graph showing the impact of the feature purification module on the overall method accuracy is presented. Vertical axis: MAPE% (Mean Absolute Percentage Error), which measures the relative error between the predicted value and the true value, and a smaller value indicates higher accuracy. The solid line in the figure is the model curve with the feature purification module; the dashed line is the model curve with the feature purification module removed. The MAPE% of the model with the purification module (solid line) is basically lower than that of the model with the module removed (dashed line) at each training epoch, proving that the purification module effectively filters out irrelevant features and focuses on key information. Under different topologies (NSFNET and GEANT2), the purification module can reduce the error, indicating its universality for various network structures. As the number of training epochs increases, the MAPE% of the model with the purification module decreases more steadily, indicating that the purification mechanism helps the model converge to a better solution. Through the curve graph, we can more comprehensively understand the impact degree of the feature purification module on the performance of the overall method, further verifying the effectiveness of our method.

[0098] Through these experimental tests, strong experimental evidence has been obtained, demonstrating the superiority and reliability of the proposed method. These results have important guiding significance for further improving and applying our method.

[0099] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. A method for predicting the performance of a power information network based on a graph neural network, characterized in that Including: Construct a graph neural network model, and extract the topological features of the power information network through the graph neural network model; Calculate the mutual influence between network flows based on time series technology; Extract node features related to the target flow through the topological information of the target flow; Calculate the correlation degree between all flows and the target flow; Fuse features to obtain the prediction result of the power information network performance.

2. The method for predicting the performance of a power information network based on a graph neural network according to claim 1, wherein: The topological features of the power information network include the connection relationship and topological features between nodes; The extracted network topology features include that the topology feature extractor uses a graph neural network for topology feature extraction, encodes the topology information of the power information network by using the connection pattern and structural features between nodes, and the graph neural network uses a graph convolutional network for extraction. The power information network topology is modeled as a graph structure, expressed as , where V represents the set of all nodes in the graph, E represents the set of all edges, and each node is represented as a feature vector , and the edge between node i and node j is represented as . Message passing in the graph convolutional network is performed at each layer, expressed as Among them, represents the feature vector of the th node in the layer, represents the activation function, represents the set of neighbor nodes of node is the normalization factor for normalizing the edge represents the weight matrix of the th layer.

3. The performance prediction method of a power information network based on a graph neural network according to claim 2, characterized in that: The calculation of the mutual influence between network flows includes that the encoder receives the feature vectors of the flows as inputs, uses the multi-head attention mechanism to extract the mutual influence relationships between the flows, and models the flows as sequences , where represents the number of flows in the network, and each flow is represented by a feature vector ; The encoder uses the multi-head attention mechanism to linearly transform the input feature vector into query, key, and value vectors. By calculating the dot product of the query and the key, the attention weight is obtained, which is used to measure the correlation between each flow and other flows. Multiply the attention weight by the value vector to obtain the weighted flow feature representation. The multi-head attention mechanism is expressed as Among them, respectively represent the query, key, and value vectors generated by linear transformation, corresponding to the flow features of different subspaces, represents the dimension of the input feature.

4. The performance prediction method of a power information network based on a graph neural network according to claim 3, characterized in that: The extraction of node features related to the target flow includes refining the topological features based on the topological feature refiner, and refining the node features related to the target flow. The topological feature refiner identifies the relationship between each node and the target flow through the node scoring mechanism, and extracts the node features related to the target flow from the output of the topological feature extractor through the weight multiplication operation.

5. The performance prediction method of a power information network based on a graph neural network according to claim 4, characterized in that: The refining of the topological features based on the topological feature refiner includes outputting a feature matrix representing the entire topological structure based on the topological feature extractor, where each row represents the feature vector of a node, and the feature vector contains the information of the node's direct and indirect nodes. The node weights related to the target flow need to be set to 1 through the node scoring mechanism, and the irrelevant node weights are set to 0; Extract the node features related to the target flow through the operation of multiplying the score by the topological features. By multiplying the node score by the topological feature matrix, the eigenvalue of the node related to the target flow is retained, and the eigenvalue of the irrelevant node is suppressed to zero; The refining of the node features related to the target flow is expressed as where x is the output of the topological feature extractor, representing a node feature, is the path of the target flow.

6. The power information network performance prediction method based on a graph neural network according to claim 5, characterized in that: The calculation of the correlation degree between all flows and the target flow includes calculating the correlation score between each flow and the target flow through the scoring mechanism, measuring the similarity or correlation degree between the two flows through the correlation score, and performing the weighted multiplication operation of the correlation score and the flow features to extract the features highly related to the target flow; The scoring mechanism is expressed as Among them, respectively represent the characteristics of all flows and the target flow in the network.

7. The performance prediction method of a power information network based on a graph neural network according to claim 6, characterized in that: The fusion of features to obtain the prediction result of the power information network performance includes fusing the features to obtain the feature representation, and predicting the target flow based on the feature representation; The prediction is expressed as Among them, represents the output of the topological feature purifier, represents the output of the inter-flow feature purifier, represent the weight and bias respectively.

8. A system using a power information network performance prediction method based on a graph neural network as described in any one of claims 1 to 7, characterized in that: Including a topological feature extraction module, an inter-flow feature extraction module, a target flow topological feature extraction module, a target flow inter-flow feature extraction module, and a fusion module; The topological feature extraction module is used to calculate the topological features of the power information network and can learn the representation of nodes; The inter-flow feature extraction module is used to calculate the mutual influence between network flows and consider the time-varying nature of traffic; The target flow topological feature extraction module is used to extract node features related to the target flow; The target flow inter-flow feature extraction module is used to calculate the correlation degree between all flows and the target flow and correct the inter-flow features; The fusion module is used to fuse features to obtain the power information network performance estimation result.

9. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, the steps of a method for predicting the performance of a power information network based on a graph neural network according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, the steps of a method for predicting the performance of a power information network based on a graph neural network according to any one of claims 1 to 7 are implemented.

Images