Satellite network traffic intelligent prediction method and system based on time shifting Pearson enhancement

Abstract

The invention provides a satellite network traffic intelligent prediction method and system based on time shifting Pearson enhancement, and the method comprises the steps: obtaining historical traffic time sequence data, and carrying out the preprocessing of the data, and generating a standardized traffic matrix; based on a sliding window mechanism, calculating a time-shifting Pearson's correlation coefficient of any node pair under a plurality of lag step lengths, selecting an absolute value of the maximum correlation coefficient as an adjacency weight, constructing a dynamic weighting adjacency matrix, and normalizing the dynamic weighting adjacency matrix; inputting the traffic matrix and the normalized adjacent matrix into a graph convolutional network, and extracting spatial dependency features; inputting the spatial feature sequence of a plurality of continuous time steps into a gating circulation unit, and extracting time sequence evolution features; and inputting the time sequence coding vector into a full connection layer to generate a flow prediction value at the next moment. According to the method, the problems that in the prior art, a graph structure depends on physical connection, explicit modeling for the flow lag propagation effect is lacked, and model training is difficult and generalization performance is insufficient in a high-dynamic scene can be solved, and high-precision and high-robustness satellite network flow prediction is achieved.

Description

Technical Field

[0001] This invention relates to the field of intelligent traffic sensing and prediction technology in integrated space-ground communication networks, and particularly to a method and system for intelligent prediction of satellite network traffic based on time-shifted Pearson enhancement. Background Technology

[0002] With the rapid development of the global digital economy and the limitations of terrestrial communication network coverage, large-scale low-Earth orbit (LEO) satellite constellation systems, represented by Starlink and OneWeb, are being deployed at an accelerated pace, gradually building a global space-based information infrastructure. These systems utilize hundreds or thousands of satellites operating at altitudes of 500 to 1200 kilometers, interconnected via inter-satellite laser links, and deeply integrated with globally distributed ground gateway stations, forming a three-in-one space-ground network architecture: space segment, ground segment, and user segment. Under this architecture, satellite nodes not only perform data relay functions but are also gradually acquiring intelligent capabilities such as routing calculation, traffic scheduling, and resource allocation, driving the network's continuous evolution towards programmability, awareness, and adaptability.

[0003] Against this backdrop, real-time and accurate perception and prediction of the spatiotemporal evolution of network traffic has become a crucial prerequisite for achieving dynamic allocation of satellite network resources, optimized link scheduling, early congestion avoidance, and quality of service assurance. However, the traffic of low-Earth orbit satellite internet exhibits complex characteristics significantly different from traditional terrestrial networks, specifically in the following three aspects: First, in the time dimension, influenced by both the periodic regression of satellite orbits (e.g., multiple daily overhead passes of sun-synchronous orbits) and the rhythms of user behavior (e.g., day-night differences, weekday / holiday patterns), traffic sequences show self-similarity and long-range correlation at multiple scales such as hours, days, and weeks. Simultaneously, when satellites fly at high speed over densely populated areas or encounter sudden events, they often trigger short-term, intense bursts of traffic, resulting in a time series exhibiting both stable trends and high-frequency fluctuations. Second, in the spatial dimension, traffic distribution is highly coupled with the real-time network topology: due to the high-speed movement of satellites, satellite-to-ground links and inter-satellite links are continuously established and interrupted, causing network connectivity to change dynamically on a second-by-second basis. This leads to significant differences in the load levels of adjacent satellites, creating strong spatial heterogeneity. Third, in terms of propagation mechanism, data needs to be relayed through multiple inter-satellite routes (e.g., local access satellite → co-orbit relay satellite → destination satellite), and there is a non-negligible time delay in the traffic response between the source node and the destination node, meaning that inter-satellite traffic has a significant "lagging propagation" characteristic. These factors collectively determine that satellite network traffic is a highly coupled, non-stationary, and dynamically evolving spatiotemporal process, placing extremely high demands on modeling methods.

[0004] The problem of network traffic prediction has been a subject of long-term research in academia and industry. Early methods were mainly based on statistical models, such as autoregressive moving average models, autoregressive integral moving average models, and their variants, approximating traffic sequences through polynomial fitting or linear weighting. These methods are simple in structure and have low computational cost, but they heavily rely on linear assumptions and manual parameter tuning, making it difficult to capture the highly nonlinear, sudden, and multi-scale periodic characteristics of satellite traffic. Their applicability is mostly limited to stable, short-term prediction scenarios. Subsequently, shallow machine learning models such as support vector machines and wavelet neural networks were introduced into traffic prediction tasks. These models improved the fitting ability to nonlinear features through kernel function mapping or wavelet decomposition, but they still struggled to model the dynamic interaction relationships between multiple nodes and had high requirements for the scale and quality of training data, thus limiting their generalization performance.

[0005] In recent years, with breakthroughs in deep learning technology, spatiotemporal convolutional networks, recurrent neural networks, attention mechanisms, and graph neural networks have gradually become the mainstream paradigms for traffic prediction. Among them, spatiotemporal graph convolutional networks (such as TGCN, DCRNN, and STGCN) have attracted widespread attention due to their natural adaptability to node-based network structures. These methods combine graph convolution operations with temporal modeling modules (such as gated recurrent units and temporal convolutional networks) to achieve joint spatiotemporal prediction of multiple nodes given adjacency relationships. To further improve adaptability in dynamic scenarios, some studies have attempted to introduce adaptive graph convolution, attention weights, or learnable parameters to implicitly infer dependencies between nodes or dynamically adjust the graph structure based on real-time traffic. However, existing solutions still generally suffer from the following common defects: First, graph topology construction relies heavily on physical connections or static orbital positions, reflecting only "whether the satellite is visible" rather than "whether the traffic is relevant," making it difficult to characterize the real business dependency structure driven by user behavior and routing strategies. Second, even when using a dynamic graph mechanism, the node relationship inference process lacks explicit time-delay correlation modeling, failing to effectively capture the delayed propagation effect of satellite traffic in the time dimension. Third, redundant or erroneous graph connections cause the feature propagation path of the graph neural network to be mixed with a large amount of irrelevant noise, the gradient update direction is ambiguous, model training is difficult, convergence is slow, and the generalization ability is significantly reduced in highly dynamic, large-scale satellite network scenarios. Summary of the Invention

[0006] In view of this, embodiments of the present invention provide a method and system for intelligent prediction of satellite network traffic based on time-shifted Pearson enhancement, in order to solve the problems of existing technologies, such as graph structure relying on physical connections and being unable to characterize the real business relevance, lack of explicit modeling of traffic lag propagation effects, and difficulty in model training and insufficient generalization performance in highly dynamic scenarios.

[0007] On one hand, this invention provides a method for intelligent prediction of satellite network traffic based on time-shifted Pearson enhancement, the method comprising: Historical satellite network traffic time-series data is acquired, and the historical satellite network traffic time-series data is preprocessed to generate a standardized traffic matrix that includes the sliding window time step dimension and the satellite node number dimension. Based on the sliding window mechanism, the time-shifted Pearson correlation coefficient of any satellite node pair is calculated at multiple lag steps within each sliding window. The absolute value corresponding to the largest correlation coefficient among the multiple lag steps is selected as the adjacency weight of the node pair at the current time. A dynamic weighted adjacency matrix is ​​constructed based on the adjacency weight. The dynamic weighted adjacency matrix is ​​then symmetrically normalized to generate a normalized adjacency matrix. The current flow matrix and the normalized adjacency matrix are input into a graph convolutional network. The spatial dependency features of satellite nodes are extracted through the neighborhood aggregation operation of the graph convolutional network, and the spatial feature representation is output. The spatial feature representation sequence of multiple consecutive time steps is input into the gated recurrent unit, and the temporal evolution features are extracted through the gating mechanism of the gated recurrent unit, and the temporal coding vector is output. The time-series coding vector is input into the fully connected layer, and the satellite network traffic prediction value for the next time step is generated through linear mapping.

[0008] In some embodiments of the present invention, the time-shifted Pearson correlation coefficient is calculated using the following formula: ; in, Indicates at time Lag step size is Time satellite node With satellite nodes The time-shifted Pearson correlation coefficient between them; Indicates the size of the sliding window; Indicates satellite node At any moment The flow rate value; Indicates satellite node The average flow rate within the sliding window; Indicates satellite node At any moment The flow rate value; Indicates satellite node The average flow rate within the sliding window.

[0009] In some embodiments of the present invention, the adjacency weight is determined by the following formula: ; in, Indicates at time satellite nodes With satellite nodes The adjacency weights between them; Indicates the lag step size The time-shifted Pearson correlation coefficient; This indicates the maximum lag step size.

[0010] In some embodiments of the present invention, the symmetry normalization process is implemented through the following steps: By adding self-loops to the aforementioned dynamically weighted adjacency matrix, an enhanced adjacency matrix is ​​obtained: ; in, Represents an enhanced adjacency matrix; This represents the dynamically weighted adjacency matrix; Represents the identity matrix; Calculate the enhancement matrix, the diagonal elements of which satisfy the following formula: ; in, Represents the enhancement matrix The Line 1 The diagonal elements of the column; Indicates at time satellite nodes With satellite nodes Enhanced adjacency weights between them; The normalized adjacency matrix is ​​generated using the following formula: ; in, This represents the normalized adjacency matrix.

[0011] In some embodiments of the present invention, the graph convolutional network has a two-layer structure, and the spatial feature representation is extracted using the following formula: ; in, Indicates time Spatial feature representation; Represents the linear rectification activation function; Represents the normalized adjacency matrix; Represents the flow matrix; and These represent the weight matrices of the first and second layers of the graph convolutional network, respectively.

[0012] In some embodiments of the present invention, the timing coding vector of the gated loop unit is calculated in the following manner: ; in, Indicates time The timing coding vector; The mapping function representing the gated loop unit; Indicates time Spatial feature representation.

[0013] In some embodiments of the present invention, the predicted satellite network traffic at the next moment is generated by the following formula: ; in, This represents the predicted flow rate at the next moment. Indicates time The timing coding vector; This represents the weight matrix of the fully connected layer; This represents the bias term of the fully connected layer.

[0014] In some embodiments of the present invention, the dynamically weighted adjacency matrix is ​​updated using any of the following update strategies: Timed sliding update: The adjacency matrix is ​​recalculated and updated at fixed step intervals. Threshold-triggered update: If the change in the time-shifted Pearson correlation coefficient calculated at the current time exceeds the preset threshold within the sliding window interval, the window will immediately slide and the adjacency matrix will be updated. Hybrid update: Combining timed sliding update and threshold-triggered update, it mainly uses fixed-step periodic updates and triggers immediate updates when the correlation coefficient changes beyond the threshold.

[0015] On the other hand, the present invention provides a satellite network traffic intelligent prediction system based on time-shifted Pearson enhancement, the system comprising: The data preprocessing module is used to acquire historical satellite network traffic time series data, preprocess the historical satellite network traffic time series data, and generate a standardized traffic matrix that includes the sliding window time step dimension and the satellite node number dimension. The dynamic graph construction module, connected to the data preprocessing module, is used to calculate the time-shifted Pearson correlation coefficient of any satellite node pair under multiple lag steps in each sliding window based on the sliding window mechanism. The absolute value corresponding to the largest correlation coefficient among the multiple lag steps is selected as the adjacency weight of the node pair at the current time. A dynamic weighted adjacency matrix is ​​constructed based on the adjacency weight. The dynamic weighted adjacency matrix is ​​then symmetrically normalized to generate a normalized adjacency matrix. The spatial feature extraction module is connected to the data preprocessing module and the dynamic graph construction module, respectively. It is used to input the current flow matrix and the normalized adjacency matrix into the graph convolutional network, extract the spatial dependency features of satellite nodes through the neighborhood aggregation operation of the graph convolutional network, and output the spatial feature representation. The temporal feature extraction module, connected to the spatial feature extraction module, is used to input the spatial feature representation sequence of multiple consecutive time steps into the gated recurrent unit, extract the temporal evolution features through the gated recurrent unit's gating mechanism, and output the temporal coding vector. The traffic prediction output module is connected to the time series feature extraction module and is used to input the time series encoding vector into the fully connected layer and generate the satellite network traffic prediction value for the next time step through linear mapping. The storage module is connected to each of the above modules and is used to store the historical satellite network traffic time series data, the standardized traffic matrix, the dynamically weighted adjacency matrix, the normalized adjacency matrix, the spatial feature representation, the time series coding vector, and the traffic prediction value. The control module is connected to each of the above modules and is used to control the data interaction and collaborative working sequence between the modules.

[0016] On the other hand, the present invention provides a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of any of the methods mentioned above.

[0017] This invention constructs a dynamic weighted adjacency matrix by introducing a time-shifted Pearson correlation coefficient, overcoming the limitations of traditional methods that rely on physical connections or static topology. It can explicitly capture the lag propagation effects of traffic in the spatiotemporal dimensions of satellite networks. Compared to static graph structures that only reflect "whether a satellite is visible," this invention, guided by the correlation of service traffic, accurately characterizes the real dependencies driven by user behavior and routing, enabling the spatial aggregation process of graph convolutional networks to focus on nodes with genuine causal relationships, significantly improving the accuracy of spatial dependency modeling.

[0018] Furthermore, this invention combines a time-shifted Pearson-enhanced dynamic graph structure with a spatiotemporal graph convolutional network to form a complete end-to-end prediction framework. In the spatial dimension, the graph convolutional network performs neighborhood aggregation based on the business-related graph, effectively avoiding noise interference introduced by irrelevant connections. In the temporal dimension, gated recurrent units encode historical spatial feature sequences, fully exploring the long-term and short-term dependencies in traffic evolution. This spatiotemporal joint modeling approach enables the model to simultaneously grasp both the macro-trends and micro-fluctuations of traffic distribution, significantly improving the prediction accuracy of future traffic distribution.

[0019] Furthermore, this invention introduces a flexible adjacency matrix dynamic update strategy, including timed sliding updates, threshold-triggered updates, and a hybrid update strategy combining both. This enables the graph structure to adaptively follow the topology changes and service bursts of the satellite network, maintaining a sensitive response to changes in inter-node correlations even in highly dynamic scenarios. This effectively overcomes the performance degradation of static graph models over long-term operation, significantly enhancing the model's robustness and generalization ability.

[0020] Thanks to its business-relevance-oriented graph structure design, the feature propagation path of graph convolutional networks is more focused, and the gradient update direction is more explicit, effectively alleviating the training oscillation problem caused by redundant connections in traditional methods. Meanwhile, gated recurrent units have a more streamlined structure and fewer parameters compared to long short-term memory networks, reducing computational overhead while maintaining temporal modeling capabilities. The combined effect of these two factors allows the model to achieve stable convergence in fewer iterations, significantly improving training efficiency and making it more suitable for resource-constrained on-board or ground-based processing scenarios.

[0021] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the description, or may be learned by practice of the invention. The objects and other advantages of the invention can be realized and obtained by means of the structures specifically pointed out in the specification and drawings.

[0022] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description

[0023] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, are not intended to limit the scope of the invention. In the drawings: Figure 1 This is a schematic diagram illustrating the steps of a satellite network traffic intelligent prediction method based on time-shifted Pearson enhancement in one embodiment of the present invention.

[0024] Figure 2 This is a schematic diagram of a spatiotemporal graph convolutional network model architecture based on time-shifted Pearson enhancement in one embodiment of the present invention.

[0025] Figure 3 This is a schematic diagram of a convolutional network model structure in one embodiment of the present invention.

[0026] Figure 4 This is a schematic diagram of a gated loop unit model structure in one embodiment of the present invention.

[0027] Figure 5This is a diagram of the spatiotemporal graph convolutional network prediction model architecture based on time-shifted Pearson enhancement in one embodiment of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0029] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0030] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0031] It should also be noted that, unless otherwise specified, the term "connection" in this article can refer not only to a direct connection, but also to an indirect connection involving an intermediary.

[0032] In the following description, embodiments of the invention will be illustrated with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar parts, or the same or similar steps.

[0033] It should be emphasized here that the step markers mentioned below are not a limitation on the order of the steps, but should be understood as meaning that the steps can be executed in the order mentioned in the embodiments, or in a different order than in the embodiments, or several steps can be executed simultaneously.

[0034] To address the shortcomings of existing technologies, such as graph structures relying on physical connections and failing to characterize real-world business relevance, lack of explicit modeling of traffic lag propagation effects, and difficulties in model training and generalization performance in highly dynamic scenarios, this invention provides a satellite network traffic intelligent prediction method based on time-shifted Pearson augmentation. Figure 1 As shown, the method includes the following steps S101~S105: Step S101: Obtain historical satellite network traffic time series data, preprocess the historical satellite network traffic time series data, and generate a standardized traffic matrix that includes the sliding window time step dimension and the satellite node number dimension.

[0035] Step S102: Based on the sliding window mechanism, calculate the time-shifted Pearson correlation coefficient of any satellite node pair under multiple lag steps within each sliding window. Select the absolute value corresponding to the largest correlation coefficient among multiple lag steps as the adjacency weight of the node pair at the current time. Construct a dynamic weighted adjacency matrix based on the adjacency weight. Perform symmetric normalization on the dynamic weighted adjacency matrix to generate a normalized adjacency matrix.

[0036] Step S103: Input the current flow matrix and normalized adjacency matrix into the graph convolutional network, extract the spatial dependency features of satellite nodes through the neighborhood aggregation operation of the graph convolutional network, and output the spatial feature representation.

[0037] Step S104: Input the spatial feature representation sequence of multiple consecutive time steps into the gated recurrent unit, extract the temporal evolution features through the gating mechanism of the gated recurrent unit, and output the temporal coding vector; Step S105: Input the temporal coding vector into the fully connected layer and generate the satellite network traffic prediction value for the next time step through linear mapping.

[0038] In step S101, data acquisition and preprocessing are performed first. Step S101 aims to collect raw satellite network traffic data and convert it into a standard format that can be used for subsequent model training and prediction, which is the data foundation of the entire prediction method.

[0039] First, it is necessary to collect the time-series traffic information of each satellite node from the historical operational data of the satellite network. This information can include key indicators that reflect the intensity of service traffic, such as on-board link load, port throughput, and buffer occupancy. The collected data must be strictly aligned to a unified timestamp to ensure that the traffic status of all nodes in the entire network can be observed under the same time reference.

[0040] After acquiring the raw time-series data, it is preprocessed. The first core operation in preprocessing is normalization. Since the flow values ​​of different satellite nodes and different time periods may vary greatly, directly inputting them into the model will lead to unstable numerical calculations and unbalanced gradient updates. Therefore, by using methods such as linear normalization, the flow value of each node is mapped to the interval [0,1] or [-1,1] to eliminate the influence of dimensions.

[0041] After alignment and normalization, a standardized multidimensional flow matrix, denoted as X, is constructed. The structure of this matrix contains two key dimensions: one is the sliding window time step dimension, which represents the length of the historical observation sequence used for prediction, usually denoted by T; the other is the satellite node number dimension, which represents the total number of nodes in the network participating in flow prediction, usually denoted by N.

[0042] Therefore, the standardized traffic matrix X can be understood as a two-dimensional array with dimensions T×N, where each row represents a traffic snapshot of all nodes in the network at a certain historical moment, and each column represents the traffic evolution sequence of a specific satellite node within a historical time window.

[0043] This step successfully transformed the chaotic raw log data into a well-structured and numerically stable standardized flow matrix, laying a solid foundation for subsequent complex spatiotemporal feature extraction. This step is a prerequisite for achieving high-precision prediction, ensuring that all subsequent calculations are based on a consistent and reliable data source.

[0044] In step S102, a dynamic weighted adjacency matrix is ​​constructed based on time-shifted Pearson correlation.

[0045] After data preprocessing and obtaining the standardized traffic matrix in step S101, step S102 is executed, which involves constructing a dynamic weighted adjacency matrix based on time-shifted Pearson correlation. Step S102 aims to overcome the limitations of traditional methods that rely on physical connections to construct static graph structures. By using a data-driven approach, it explicitly mines the spatial dependencies between satellite nodes caused by service propagation and exhibiting lag effects, thereby constructing a dynamic graph structure that can truly reflect traffic correlation.

[0046] like Figure 2 The diagram shown is a schematic of the spatiotemporal graph convolutional network model architecture based on time-shifted Pearson enhancement.

[0047] To capture the unique "traffic delay + spatial diffusion" effect in satellite networks, this invention introduces the time-shifted Pearson correlation coefficient. For any two satellite nodes i and j, at time t, instead of focusing solely on their traffic correlation at the same time point, this invention examines the correlation between the traffic sequence of one node and the other node after a time lag of τ steps. Specifically, based on a sliding window mechanism, within a window of length w, for a given lag step τ, the formula for calculating the time-shifted Pearson correlation coefficient is shown in formula (1): ; (1) in, Indicates at time Lag step size is Time satellite node With satellite nodes The time-shifted Pearson correlation coefficient between them; Indicates the size of the sliding window; Indicates satellite node At any moment The flow rate value; Indicates satellite node The average flow rate within the sliding window; Indicates satellite node At any moment The flow rate value; Indicates satellite node The average flow rate within the sliding window.

[0048] Formula (1) measures the degree of linear correlation between the flow sequence of node j and the flow sequence of node i within window w after the flow sequence of node j is shifted backward by τ steps on the time axis. By traversing a series of possible lag steps τ∈[0, K], where K is the preset maximum lag step, the correlation sequence between node pairs (i, j) under different time delays can be calculated.

[0049] To ensure that the constructed graph structure reflects the most significant and stable spatial dependencies between nodes, the absolute value corresponding to the maximum value among these correlation coefficients is selected as the adjacency weight of node pair (i, j) at time t. This is because the maximum value represents the strongest correlation between two sequences under the optimal matching lag, which can effectively reveal the causal or strong correlation caused by route propagation. This process can be achieved through formula (2): ; (2) in, Indicates at time satellite nodes With satellite nodes The adjacency weights between them; Indicates the lag step size The time-shifted Pearson correlation coefficient; This indicates the maximum lag step size.

[0050] By applying the operation of formula (2) to all node pairs (i, j), the dynamic weighted adjacency matrix at time t can be constructed. Each element of the matrix These all represent the strength of relevance at the business level, rather than simple physical accessibility.

[0051] Since this dynamically weighted adjacency matrix will be directly used in subsequent graph convolutional network operations, it needs to be symmetrically normalized to maintain numerical stability and aggregate node features. This process is achieved through the following three steps: First, a self-loop is added to each node based on the dynamically weighted adjacency matrix to obtain an enhanced adjacency matrix, ensuring that the graph convolution operation can aggregate the feature information of the nodes themselves. As shown in formula (3): ; (3) in, Represents an enhanced adjacency matrix; Represents a dynamically weighted adjacency matrix; Represents the identity matrix.

[0052] Next, calculate the enhancement matrix. The enhancement matrix is ​​a diagonal matrix, which can be denoted as . The elements on its diagonal The sum of the weights of all edges connected to node i reflects the connection strength of the node. The diagonal elements of the enhancement matrix satisfy the following formula (4): ; (4) in, Represents the enhancement matrix The Line 1 The diagonal elements of the column; Indicates at time satellite nodes With satellite nodes Enhanced adjacency weights between them.

[0053] Finally, the augmented adjacency matrix is ​​symmetrically normalized using the negative 1 / 2 power of the augmentation degree matrix to generate the final normalized adjacency matrix used in the graph convolutional network, as shown in Equation (5): ; (5) in, This represents the normalized adjacency matrix.

[0054] Through this symmetric normalization operation, the normalized adjacency matrix Each element in In effect, this means that the connection weight between node i and node j is scaled by the square root of their respective degrees. This ensures that during the propagation of graph convolution, the feature update of nodes takes into account both the strength of the adjacency relationship and avoids numerical instability or gradient vanishing / exploding problems caused by excessive differences in node degrees.

[0055] Through the above operations, a normalized adjacency matrix was successfully constructed. However, the topology and traffic of low-Earth orbit satellite networks are highly dynamic, therefore, this adjacency matrix also needs to be continuously updated over time to ensure that it accurately reflects the latest service relationships between nodes. To this end, this invention further introduces a flexible dynamic weighted adjacency matrix update strategy to determine when to re-execute the above matrix construction process. Specifically, any of the following three strategies can be used for updating: The first strategy is a timed sliding update strategy. Under this strategy, the system presets a fixed time step Δt. Every Δt time intervals, the sliding window is used to recalculate the time-shifted Pearson correlation coefficients between all node pairs, thereby updating the adjacency matrix. This strategy is simple to implement, has predictable computational overhead, and is suitable for scenarios where business traffic changes are relatively stable and periodic patterns are obvious.

[0056] The second strategy is a threshold-triggered update strategy. Under this strategy, even within the timed sliding window interval, the system continuously monitors changes in the time-shifted Pearson correlation coefficient between nodes. If the correlation coefficient calculated at the current moment differs from that at the previous update moment by more than a preset threshold, it indicates a significant change in the network service pattern (e.g., a sudden traffic event). In this case, the system will immediately trigger an additional sliding window and matrix update to quickly respond to this change. This strategy can promptly capture abrupt changes in the correlation structure caused by sudden traffic, improving the model's adaptability to dynamic environments.

[0057] The third approach is a hybrid update strategy. This strategy combines the advantages of the two methods mentioned above, primarily using timed sliding updates to ensure a regular response to periodic changes in the network; and secondarily using threshold-triggered updates as a supplementary mechanism to handle unexpected events. Specifically, the system periodically updates the adjacency matrix at a fixed step size Δt, and continuously monitors changes in the correlation coefficients during the intervals between each timed update. Once the change exceeds the threshold, an immediate update is triggered. This hybrid strategy balances computational regularity and responsiveness, achieving a good balance between computational overhead and model performance.

[0058] Those skilled in the art can select appropriate update strategies or combinations based on the actual network size and business characteristics to achieve a balance between computational overhead and model response sensitivity.

[0059] By introducing one or more of the aforementioned dynamic update strategies, the adjacency matrix construction module of this invention can adaptively follow the rapid changes in the satellite network, ensuring that the graph structure input into the graph convolutional network is always up-to-date and highly relevant. This greatly enhances the robustness and generalization ability of the entire traffic prediction model in highly dynamic, large-scale low-Earth orbit satellite network scenarios, and is one of the key features that distinguishes this invention from static graph methods.

[0060] Thus, through step S102, a normalized adjacency matrix that can dynamically change with network status and is guided by the correlation of service traffic has been successfully constructed. This matrix no longer depends on static physical topology, but explicitly captures the lag propagation effect of traffic in the spatiotemporal dimension of satellite networks in a data-driven manner, providing a high-quality graph structure foundation for accurate spatial feature extraction in subsequent steps.

[0061] In step S103, spatial dependency features are extracted based on graph convolutional networks.

[0062] After obtaining the normalized adjacency matrix that accurately reflects business relevance through step S102, step S103 is executed, which involves extracting spatial dependency features based on graph convolutional networks. Step S103 aims to leverage the powerful neighborhood aggregation capability of graph convolutional neural networks to combine the standardized traffic matrix generated in step S101 with the dynamic graph structure constructed in step S102, accurately extracting the spatial features of each satellite node at the current moment from both network topology and business relevance perspectives.

[0063] like Figure 3 The diagram shown is a schematic of the graph convolutional network model structure.

[0064] The core idea of ​​graph convolutional networks is to update the feature representation of a target node by aggregating the feature information of the target node and its neighboring nodes. In this invention, the standardized flow matrix at the current moment is used as the initial feature input of the node, and the normalized adjacency matrix generated in step S102 is used as the connection structure and weights between nodes.

[0065] In a preferred embodiment of the present invention, the graph convolutional network employs a two-layer structure to extract deeper spatial dependencies. The spatial feature extraction process can be expressed by the following formula (6): ; (6) in, Indicates time Spatial feature representation; Represents the linear rectification activation function; Represents the normalized adjacency matrix; Represents the flow matrix; and These represent the weight matrices of the first and second layers of the graph convolutional network, respectively.

[0066] Further explanation of formula (6) is provided.

[0067] First, in the first layer of graph convolution, the normalized adjacency matrix is ​​multiplied by the input flow matrix and the weight matrix of the first layer. Wherein, This operation achieves neighborhood aggregation, meaning that for each target node, its own flow characteristics are weighted and summed with the flow characteristics of all its neighboring nodes according to the weights defined in the normalized adjacency matrix, thereby fusing spatial adjacency information into the node features. Subsequently, this is combined with a trainable weight matrix... Multiplication involves a linear transformation of the aggregated features to learn different dimensions of spatial feature representation. The transformed result is then nonlinearly mapped using the Rectified Luminaire (ReLU) activation function, introducing nonlinearity to enable the model to fit complex spatial dependency patterns. The ReLU function sets all negative inputs to 0 while keeping positive values ​​unchanged, which helps enhance the model's sparsity and convergence speed.

[0068] After the first layer of graph convolution, an intermediate feature representation that incorporates neighbor information is obtained. To capture a wider range and higher-order neighborhood influence, this is used as input to the second layer of graph convolution. In the second layer, a normalized adjacency matrix is ​​used again to perform neighborhood aggregation on the intermediate features. This operation allows each node to aggregate information from its "neighbor's neighbor," thereby expanding the spatial receptive field. Subsequently, it is multiplied with the weight matrix of the second layer for another linear transformation. Finally, the ReLU activation function is applied again to obtain the final spatial feature representation.

[0069] By stacking and nonlinearly transforming the two layers of graph convolutional networks, a feature representation rich in spatial structure information was successfully extracted from the original traffic matrix. This feature representation is an N×d matrix, where N is the number of satellite nodes and d is the spatial feature dimension after encoding by the graph convolutional network. It not only contains the traffic information of each node itself, but more importantly, it deeply integrates information from nodes with strong correlations through a dynamically constructed adjacency matrix that reflects real-world business propagation relationships, thereby achieving accurate modeling of the spatial dependencies of the satellite network. This feature representation will be fully preserved and used as input to the subsequent temporal feature extraction module.

[0070] In step S104, temporal evolution features are extracted based on gated recurrent units. After obtaining the spatial feature representation of the current moment through step S103, step S104 is executed, which is to extract temporal evolution features based on gated recurrent units. Step S104 aims to utilize the powerful temporal modeling capabilities of gated recurrent units, taking the spatial feature representation sequence of multiple consecutive historical moments as input, to capture the dynamic evolution pattern of satellite network traffic in the time dimension, thereby extracting a temporal coding vector containing temporal context information.

[0071] like Figure 4 The diagram shown is a schematic of the gated loop unit model structure.

[0072] Satellite network traffic is not only spatially interconnected but also exhibits complex dynamic characteristics over time, such as long-range correlations caused by the periodic operation of satellite orbits and periodic fluctuations caused by user behavior rhythms. These temporal dependencies are crucial for accurately predicting future traffic. Therefore, a model capable of effectively processing sequential data is needed to uncover these temporal patterns. Gated recurrent units (RNNs), as a special type of recurrent neural network, can selectively memorize important information and forget irrelevant information through their sophisticated gating mechanism, thereby effectively learning long-term and short-term temporal dependencies and mitigating the gradient vanishing problem common in traditional RNNs to some extent.

[0073] Specifically, the spatial feature representation sequence of multiple consecutive time steps is used as the input to the gated recurrent unit. This sequence can be represented as... ,in, Indicates the length of the time window. It is the spatial feature representation of the current moment output by step S103. , These are spatial feature representations corresponding to each past moment. This series of feature representations has been fused with the spatial neighborhood information of each moment through step S103, thus forming an evolutionary sequence containing rich spatiotemporal information.

[0074] The sequence is input sequentially into the gated loop unit according to time. The gated loop unit contains structures such as update gates and reset gates, and can adjust the current input based on the current sequence. and the hidden state passed down from the previous moment The hidden state is dynamically calculated and updated at the current moment. This process can be abstractly expressed by the following formula (7): ; (7) in, Indicates time The timing coding vector; The mapping function representing the gated loop unit; Indicates time Spatial feature representation More specifically, for each time step in the sequence, the gated loop unit performs the following operations: First, a reset gate determines how much past information needs to be discarded, and this is combined with the current input to form a new candidate hidden state. Then, an update gate determines how much information needs to be retained from the previous hidden state and how much new information needs to be accepted from the candidate hidden states, thus generating the current hidden state. After the entire sequence has been processed, the hidden state output at the last time step is the required temporal coding vector.

[0075] This temporal coding vector is a condensed representation of the entire sequence information. It no longer represents only the spatial features at a single moment, but rather, through the recursive encoding of gated recurrent units, it integrates the spatial feature evolution patterns over the past w time steps. The temporal coding vector contains both recent traffic change trends and periodic patterns generated by the combined effects of network dynamics and user behavior.

[0076] In this way, the traffic forecasting problem is successfully elevated from simple spatial feature extraction or basic time series analysis to a deep modeling level of spatiotemporal joint evolutionary characteristics. This time-series encoded vector, rich in temporal context information, will be passed to the next step to generate the final traffic forecast value.

[0077] This invention employs gated recurrent units, which, compared to traditional long short-term memory networks, have a more streamlined structure and fewer parameters. While ensuring good temporal modeling capabilities, they also offer faster training speeds and lower computational overhead, which is a significant advantage for resource-constrained on-board or ground-based processing scenarios.

[0078] In step S105, the flow prediction output is based on the fully connected layer.

[0079] After obtaining the temporal encoding vector incorporating spatiotemporal evolution information in step S104, step S105 is executed, which involves outputting traffic prediction based on a fully connected layer. Step S105 aims to map the high-dimensional spatiotemporal feature representation to the final prediction target space, generating specific traffic prediction values ​​for each satellite node at the next time step, thus completing the entire prediction process.

[0080] A temporal encoded vector is a condensed representation obtained by a gated recurrent unit encoding a sequence of spatial features over the past w time steps. It contains key spatiotemporal information needed to predict future traffic flow. However, the temporal encoded vector itself is an abstract, high-dimensional feature vector, and its value and dimension do not directly correspond to the final, physically meaningful traffic flow prediction value. Therefore, a mapping function is needed to transform the temporal encoded vector from the feature space to the scalar space of traffic flow values.

[0081] In this invention, a fully connected layer is used to implement this mapping function. A fully connected layer, also known as a linear layer, is essentially a simple linear transformation model. It multiplies the input vector by a trainable weight matrix and adds a trainable bias term to obtain the output vector. This process can be clearly expressed by the following formula (8): ; (8) in, This represents the predicted flow rate at the next moment. Indicates time The timing coding vector; This represents the weight matrix of the fully connected layer; This represents the bias term of the fully connected layer, used to provide a learnable baseline offset for each satellite node.

[0082] As can be seen from formula (8), the calculation process of the fully connected layer is very intuitive. The temporal coding vector is first multiplied by the weight matrix to achieve a linear projection from the "feature dimension" to the "node dimension". This step can be understood as allocating the abstract spatiotemporal features to each specific satellite node according to the learned pattern. Then, a bias term is added element by element on the projection result to adjust the prediction benchmark of each node. Finally, the traffic prediction distribution of all satellite nodes in the network at the next moment is obtained.

[0083] Through this linear mapping operation, the abstract features obtained through complex spatiotemporal modeling are successfully transformed into traffic prediction values ​​with clear physical meaning that can be directly used by the network scheduling system. These prediction values ​​can not only be used for immediate decision-making in the next moment, but also, through iterative or sequence generation, output traffic prediction sequences for multiple future moments, providing more comprehensive decision support for upper-layer applications such as resource pre-allocation, congestion warning, and route optimization in satellite networks.

[0084] Thus, through steps S101 to S105, the present invention completes a complete end-to-end prediction process of "data preprocessing → dynamic graph construction → spatial feature extraction → temporal feature extraction → traffic prediction output", realizing high-precision and intelligent prediction of low-orbit satellite network traffic.

[0085] Corresponding to the above method, the present invention also provides a satellite network traffic intelligent prediction system based on time-shifted Pearson enhancement, the system comprising: The data preprocessing module is used to acquire historical satellite network traffic time-series data, preprocess the historical satellite network traffic time-series data, and generate a standardized traffic matrix that includes the sliding window time step dimension and the satellite node number dimension.

[0086] The dynamic graph construction module, connected to the data preprocessing module, is used to calculate the time-shifted Pearson correlation coefficient of any satellite node pair at multiple lag steps within each sliding window based on the sliding window mechanism. The absolute value corresponding to the largest correlation coefficient among multiple lag steps is selected as the adjacency weight of the node pair at the current time. A dynamic weighted adjacency matrix is ​​constructed based on the adjacency weight, and the dynamic weighted adjacency matrix is ​​symmetrically normalized to generate a normalized adjacency matrix.

[0087] The spatial feature extraction module is connected to the data preprocessing module and the dynamic graph construction module, respectively. It is used to input the current flow matrix and the normalized adjacency matrix into the graph convolutional network, extract the spatial dependency features of satellite nodes through the neighborhood aggregation operation of the graph convolutional network, and output the spatial feature representation.

[0088] The temporal feature extraction module, connected to the spatial feature extraction module, is used to input the spatial feature representation sequence of multiple consecutive time steps into the gated recurrent unit, extract the temporal evolution features through the gating mechanism of the gated recurrent unit, and output the temporal coding vector.

[0089] The traffic prediction output module, connected to the temporal feature extraction module, is used to input the temporal encoding vector into the fully connected layer and generate the satellite network traffic prediction value for the next time step through linear mapping.

[0090] The storage module is connected to each of the above modules and is used to store historical satellite network traffic time-series data, standardized traffic matrix, dynamically weighted adjacency matrix, normalized adjacency matrix, spatial feature representation, time-series coding vector, and traffic prediction values.

[0091] The control module is connected to each of the above modules and is used to control the data interaction and collaborative working sequence between the modules.

[0092] like Figure 5 The diagram shown is an architecture diagram of a spatiotemporal graph convolutional network prediction model based on time-shifted Pearson enhancement.

[0093] Corresponding to the above method, the present invention also provides an electronic device including a computer device, the computer device including a processor and a memory, the memory storing computer instructions, the processor executing the computer instructions stored in the memory, and when the computer instructions are executed by the processor, the electronic device performs the steps of the method as described above.

[0094] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned method. The computer-readable storage medium may be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, register, floppy disk, hard disk, removable storage disk, CD-ROM, or any other form of storage medium known in the art.

[0095] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0096] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0097] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0098] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent prediction of satellite network traffic based on time-shifted Pearson enhancement, characterized in that, The method includes: Historical satellite network traffic time-series data is acquired, and the historical satellite network traffic time-series data is preprocessed to generate a standardized traffic matrix that includes the sliding window time step dimension and the satellite node number dimension. Based on the sliding window mechanism, the time-shifted Pearson correlation coefficient of any satellite node pair is calculated at multiple lag steps within each sliding window. The absolute value corresponding to the largest correlation coefficient among the multiple lag steps is selected as the adjacency weight of the node pair at the current time. A dynamic weighted adjacency matrix is ​​constructed based on the adjacency weight. The dynamic weighted adjacency matrix is ​​then symmetrically normalized to generate a normalized adjacency matrix. The current flow matrix and the normalized adjacency matrix are input into a graph convolutional network. The spatial dependency features of satellite nodes are extracted through the neighborhood aggregation operation of the graph convolutional network, and the spatial feature representation is output. The spatial feature representation sequence of multiple consecutive time steps is input into the gated recurrent unit, and the temporal evolution features are extracted through the gating mechanism of the gated recurrent unit, and the temporal coding vector is output. The time-series coding vector is input into the fully connected layer, and the satellite network traffic prediction value for the next time step is generated through linear mapping.

2. The intelligent prediction method for satellite network traffic based on time-shifted Pearson enhancement according to claim 1, characterized in that, The time-shifted Pearson correlation coefficient is calculated using the following formula: ； in, Indicates at time Lag step size is Time satellite node With satellite nodes The time-shifted Pearson correlation coefficient between them; Indicates the size of the sliding window; Indicates satellite node At any moment The flow rate value; Indicates satellite node The average flow rate within the sliding window; Indicates satellite node At any moment The flow rate value; Indicates satellite node The average flow rate within the sliding window.

3. The intelligent prediction method for satellite network traffic based on time-shifted Pearson enhancement according to claim 1, characterized in that, The adjacency weight is determined by the following formula: ； in, Indicates at time satellite nodes With satellite nodes The adjacency weights between them; Indicates the lag step size The time-shifted Pearson correlation coefficient; This indicates the maximum lag step size.

4. The intelligent prediction method for satellite network traffic based on time-shifted Pearson enhancement according to claim 1, characterized in that, The symmetric normalization process is achieved through the following steps: By adding self-loops to the aforementioned dynamically weighted adjacency matrix, an enhanced adjacency matrix is ​​obtained: ； in, Represents an enhanced adjacency matrix; This represents the dynamically weighted adjacency matrix; Represents the identity matrix; Calculate the enhancement matrix, the diagonal elements of which satisfy the following formula: ； in, Represents the enhancement matrix The Line number The diagonal elements of the column; Indicates at time satellite nodes With satellite nodes Enhanced adjacency weights between them; The normalized adjacency matrix is ​​generated using the following formula: ； in, This represents the normalized adjacency matrix.

5. The intelligent prediction method for satellite network traffic based on time-shifted Pearson enhancement according to claim 1, characterized in that, The graph convolutional network has a two-layer structure, and the spatial feature representation is extracted using the following formula: ； in, Indicates time Spatial feature representation; Represents the linear rectification activation function; Represents the normalized adjacency matrix; Represents the flow matrix; and These represent the weight matrices of the first and second layers of the graph convolutional network, respectively.

6. The intelligent prediction method for satellite network traffic based on time-shifted Pearson enhancement according to claim 1, characterized in that, The timing coding vector of the gated recurrent unit is calculated in the following way: ； in, Indicates time The timing coding vector; The mapping function representing the gated loop unit; Indicates time Spatial feature representation.

7. The intelligent prediction method for satellite network traffic based on time-shifted Pearson enhancement according to claim 1, characterized in that, The predicted satellite network traffic for the next moment is generated using the following formula: ； in, This represents the predicted flow rate at the next moment. Indicates time The timing coding vector; This represents the weight matrix of the fully connected layer; This represents the bias term of the fully connected layer.

8. The intelligent prediction method for satellite network traffic based on time-shifted Pearson enhancement according to claim 1, characterized in that, The dynamically weighted adjacency matrix is ​​updated using any of the following update strategies: Timed sliding update: The adjacency matrix is ​​recalculated and updated at fixed step intervals by sliding the window. Threshold-triggered update: If the change in the time-shifted Pearson correlation coefficient calculated at the current time exceeds the preset threshold within the sliding window interval, the window will immediately slide and the adjacency matrix will be updated. Hybrid update: Combining timed sliding update and threshold-triggered update, it mainly uses fixed-step periodic updates and triggers immediate updates when the correlation coefficient changes beyond the threshold.

9. A satellite network traffic intelligent prediction system based on time-shifted Pearson enhancement, characterized in that, The system includes: The data preprocessing module is used to acquire historical satellite network traffic time series data, preprocess the historical satellite network traffic time series data, and generate a standardized traffic matrix that includes the sliding window time step dimension and the satellite node number dimension. The dynamic graph construction module, connected to the data preprocessing module, is used to calculate the time-shifted Pearson correlation coefficient of any satellite node pair under multiple lag steps in each sliding window based on the sliding window mechanism. The absolute value corresponding to the largest correlation coefficient among the multiple lag steps is selected as the adjacency weight of the node pair at the current time. A dynamic weighted adjacency matrix is ​​constructed based on the adjacency weight. The dynamic weighted adjacency matrix is ​​then symmetrically normalized to generate a normalized adjacency matrix. The spatial feature extraction module is connected to the data preprocessing module and the dynamic graph construction module, respectively. It is used to input the current flow matrix and the normalized adjacency matrix into the graph convolutional network, extract the spatial dependency features of satellite nodes through the neighborhood aggregation operation of the graph convolutional network, and output the spatial feature representation. The temporal feature extraction module, connected to the spatial feature extraction module, is used to input the spatial feature representation sequence of multiple consecutive time steps into the gated recurrent unit, extract the temporal evolution features through the gated recurrent unit's gating mechanism, and output the temporal coding vector. The traffic prediction output module is connected to the time series feature extraction module and is used to input the time series encoding vector into the fully connected layer and generate the satellite network traffic prediction value for the next time step through linear mapping. The storage module is connected to each of the above modules and is used to store the historical satellite network traffic time series data, the standardized traffic matrix, the dynamically weighted adjacency matrix, the normalized adjacency matrix, the spatial feature representation, the time series coding vector, and the traffic prediction value. The control module is connected to each of the above modules and is used to control the data interaction and collaborative working sequence between the modules.

10. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method as described in any one of claims 1 to 8.

Images