High dynamic tactical network link prediction method based on space-time synchronous attention mechanism

Abstract

The application discloses a high-dynamic tactical network link prediction method based on a space-time synchronous attention mechanism, which comprises the following steps: extracting node features and communication links of a communication network to form a feature vector; extracting a network topology graph to build a historical space-time graph sequence; performing standardization processing on the node features and the communication link features, and performing sinusoidal time position coding and fusion on the historical space-time graph sequence; extracting the network topology graph to map nodes to global space edges; building bidirectional time edges between the same nodes at adjacent time steps; forming a supergraph, inputting the supergraph into a graph neural network encoder, and outputting feature vectors of all nodes; splicing the feature vectors of two nodes, inputting the feature vectors into a multilayer perceptron decoder, and obtaining a communication link connectivity probability between the nodes. The application can realize synchronous modeling of network space structure information and time evolution information, thereby improving the accuracy and stability of link state prediction in a high-dynamic and complex environment tactical communication network.

Description

Technical Field

[0001] This invention relates to the field of information engineering, and specifically to a method for predicting highly dynamic tactical network links based on a spatiotemporal synchronous attention mechanism. Background Technology

[0002] Tactical Communication Networks (TCNs) are a typical type of Mobile Ad hoc Network (MANET). In real-world tactical scenarios, network nodes (such as armored vehicles and drones) are highly mobile, and the network topology changes frequently over time, exhibiting significant dynamic characteristics. Accurately predicting the link state at future moments is a core prerequisite for achieving smooth convergence of dynamic routing and ensuring the reliable execution of tactical missions. However, highly dynamic tactical networks face extremely severe challenges: high-speed node mobility, complex terrain obstruction, and variable electromagnetic environments lead to frequent link breaks and reconnections, resulting in nonlinear and abrupt topology evolution. Traditional topology evolution theories are insufficient to effectively model such complex dynamics.

[0003] Currently, link prediction methods for mobile ad hoc networks are mainly divided into two categories: non-artificial intelligence methods and artificial intelligence methods.

[0004] Non-AI methods primarily rely on mathematical models, such as Markov chains and Autoregressive Moving Average (ARIMA) models. Markov chain-based methods for node mobility prediction and relay node selection, or ARIMA models for predicting the movement speed of randomly walking nodes, assist in route discovery. These methods are effective for simple mobility models (such as Random Waypoint), but heavily depend on prior kinematic assumptions and struggle to capture nonlinear topological changes caused by complex terrain occlusion and tactical intentions globally. Therefore, they cannot meet the prediction requirements of highly dynamic tactical networks.

[0005] Artificial intelligence approaches incorporate machine learning and deep learning techniques. Researchers in this field have compared the performance of Multilayer Perceptrons (MLPs) and Long Short-Term Memory Networks (LSTMs) in ad hoc network mobility prediction, finding that LSTMs can better capture historical trajectories. However, these traditional deep learning models can only process gridded or sequential data, cannot directly perceive the global graph topology of the network, and struggle to extract spatial cascading features between multi-hop nodes, thus failing to achieve joint modeling of spatiotemporal information.

[0006] While the aforementioned methods have addressed the single-dimensional problems of temporal or spatial data to some extent, none have achieved a deep integration of temporal evolution and spatial physical topology. To address this bottleneck, researchers have introduced graph neural networks (GNNs) into network topology modeling. GNNs can directly process graph-structured data, learning node representations by recursively aggregating neighbor information. To further capture dynamic characteristics, researchers combined GNNs with temporal models, proposing the Spatiotemporal Graph Neural Network (STGNN) for predicting the future state of tactical networks or ad hoc networks. However, most existing STGNN methods (such as STGED) employ a serially decoupled architecture of "spatial graph aggregation first, followed by temporal series modeling." This spatiotemporal separation modeling method has three inherent defects: (1) Spatiotemporal receptive field fragmentation: When nodes perform spatial aggregation, they cannot refer to the topological state at historical moments, resulting in a delayed response to events such as sudden disconnection; (2) Long-range memory forgetting: When processing long-period observations, recurrent neural networks are prone to gradient decay and information forgetting, making it difficult to maintain attention to early key maneuver states; (3) Error accumulation and propagation: The one-way propagation from the spatial module to the temporal module leads to step-by-step accumulation of errors, making it difficult for the model to maintain stable accuracy in extreme environments.

[0007] It is evident that existing methods struggle to achieve accurate and real-time link prediction in highly dynamic, terrain-constrained tactical networks. Therefore, there is an urgent need for a link prediction method that can overcome the bottleneck of spatiotemporal decoupling and achieve synchronous perception of spatiotemporal characteristics, in order to improve the accuracy and stability of link prediction in highly dynamic tactical communication networks under complex environments. Summary of the Invention

[0008] The purpose of this invention is to address the problems of spatiotemporal receptive field fragmentation, long-range memory forgetting, and response lag in existing serial decoupled spatiotemporal prediction models in highly dynamic tactical networks. This invention provides a link prediction method for highly dynamic tactical networks based on a spatiotemporal synchronous attention mechanism. This method can achieve synchronous modeling of network spatial structure information and temporal evolution information, thereby improving the accuracy and stability of link state prediction in highly dynamic and complex tactical communication networks.

[0009] To achieve the above functions, this invention designs a highly dynamic tactical network link prediction method based on a spatiotemporal synchronous attention mechanism. For the target communication network, the following steps S1-S6 are executed to construct a communication link connectivity prediction model, and the following step S7 is executed to predict the communication link connectivity status between any nodes in the future time step:

[0010] Step S1: Obtain the operation log of the target communication network, extract the feature vector of each node, including node features and communication link features, and generate a binary label of communication link status as a criterion for communication link connectivity based on a preset path loss threshold.

[0011] Step S2: Set up a historical observation window consisting of several time steps, extract the network topology graph of several consecutive time steps, and construct a historical spatiotemporal graph sequence with temporal dependencies;

[0012] Step S3: Standardize the extracted node features and communication link features, and perform sinusoidal time-position coding on the historical spatiotemporal graph sequence; then fuse the standardized node features with the sinusoidal time-position coding.

[0013] Step S4: Extract the network topology graph of each time step in the historical observation window and map the nodes to global spatial edges; construct bidirectional temporal edges between the same nodes in adjacent time steps; form a spatiotemporally unified supergraph based on global spatial edges and bidirectional temporal edges.

[0014] Step S5: Input the super graph into the graph neural network encoder, calculate the attention weights between nodes, and extract the feature vectors of all nodes in the last time step from the output matrix of the graph neural network encoder.

[0015] Step S6: Concatenate the feature vectors of any two nodes whose connectivity to be predicted in the supergraph, and input them into the multilayer perceptron decoder to obtain the communication link connectivity probability between the nodes.

[0016] Step S7: Train and test the communication link connectivity prediction model to obtain the trained communication link connectivity prediction model. Apply the communication link connectivity prediction model to predict the communication link connectivity status between any nodes in the future time step.

[0017] Beneficial effects: Compared with the prior art, the advantages of the present invention include:

[0018] 1. This invention, through innovative reconstruction of a "bidirectional time edge + spatiotemporal unified supergraph," completely breaks down the spatiotemporal barriers of the traditional serial decoupling architecture. In a single operation, the communication link connectivity prediction model can simultaneously weigh spatial physical constraints (such as path loss caused by large hills) against temporal maneuvering inertia, thereby accurately capturing the evolutionary patterns of nodes in complex terrain.

[0019] 2. To address the inherent temporal confusion and feature forgetting issues in long-term prediction, this invention introduces sinusoidal temporal positional encoding, which assigns a strict temporal anchor point to each node in the supergraph. Ablation experiments show that even in Continuous Random Walk Networks (CNCMs) with irregular node movement and extremely chaotic topology changes, this invention maintains extremely high prediction robustness thanks to the dual protection of bidirectional temporal edges and positional encoding.

[0020] 3. This invention abandons the complex and gradient-exploding recurrent neural network structure of traditional methods, and adopts a Transformer architecture based on dot product attention to process the spatiotemporal graph throughout the process. This design not only avoids the gradient vanishing problem in long sequence training, but also greatly enhances the parallel computing efficiency and training convergence speed of the model when deployed on actual tactical platforms, providing lightweight and energy-efficient technical support for the future integrated intelligent closed loop of "prediction-decision-routing". Attached Figure Description

[0021] Figure 1 This is a flowchart of a high-dynamic tactical network link prediction method based on a spatiotemporal synchronous attention mechanism provided in an embodiment of the present invention;

[0022] Figure 2 This is a thermal distribution comparison diagram of the hop count evolution over time for two tactical network topologies provided in an embodiment of the present invention;

[0023] Figure 3 This is a bar chart comparing the link prediction F1 scores of the present invention and various baseline models under different historical observation step sizes. Detailed Implementation

[0024] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0025] The high-dynamic tactical network link prediction method based on spatiotemporal synchronous attention mechanism provided in this invention embodiment refers to... Figure 1 For the target communication network, the following steps S1-S6 are executed to build a communication link connectivity prediction model, and the following step S7 is executed to complete the prediction of the communication link connectivity status between any nodes in the future time step; This embodiment is completed on a deep learning workstation configured with an NVIDIA RTX 4090 GPU, and the model is built based on the PyTorch and PyTorchGeometric (PyG) frameworks.

[0026] Step S1: Obtain the operation log of the target communication network, extract the feature vector of each node, including node features and communication link features, and generate a binary label of communication link status as a criterion for communication link connectivity based on a preset path loss threshold.

[0027] The specific steps of step S1 are as follows:

[0028] Step S1.1: Extract the node features and communication link features of the target communication network, including the two-dimensional maneuvering speed of each node as node features; and also the physical distance between nodes (Distance), electromagnetic wave propagation delay (PropDelay), and path loss as communication link features.

[0029] This embodiment uses a Tactical Mobile Communication Network (CNTM) dataset generated by the high-fidelity network simulation tool EXata. The dataset simulates the maneuvering of a single armored tank company (24 nodes) along a winding path around a large hilly terrain during the troop deployment phase, exhibiting a strong physical terrain obstruction effect. Since the communication link connectivity prediction model focuses more on the instantaneous motion trend of nodes than their absolute coordinates, this embodiment extracts the two-dimensional maneuvering speed of the armored nodes as the initial node features (feature dimension 2); simultaneously, it extracts the physical distance between nodes, path loss, and electromagnetic wave propagation delay as the initial communication link features (feature dimension 3).

[0030] Step S1.2: Set a preset path loss threshold. If the path loss between two nodes is lower than the preset threshold, then the communication link between the two nodes is determined to be connected in the current time step, and the communication link status is labeled with a binary tag. Otherwise, determine that the communication link between the two nodes is broken at the current time step, and assign a binary label to the communication link status. ;in, A binary label representing the communication link status between node i and node j at time step t.

[0031] The extracted communication link features (regardless of connectivity) are used as edge features for subsequent model inputs, including: physical distance between nodes, electromagnetic wave propagation delay, and path loss; node features are extracted as the two-dimensional maneuvering speed of each node. In this embodiment, all of the above features are directly obtained from the simulation logs without additional calculation.

[0032] Step S2: Set up a historical observation window consisting of several time steps, extract the network topology graph of several consecutive time steps, and construct a historical spatiotemporal graph sequence with temporal dependencies;

[0033] The specific steps of step S2 are as follows:

[0034] Step S2.1: At time step t, represent the target communication network as a network topology graph. ,in, A set of nodes; This represents a node, where N is the number of nodes. In this example, N=24. Let F represent the node feature matrix, where F=2 in this embodiment, corresponding to the two-dimensional maneuvering velocity of the node. Each row of the node feature matrix represents the two-dimensional maneuvering velocity of each node at time step t.

[0035] ;

[0036] ;

[0037] in, This represents the two-dimensional maneuvering velocity of node i at time step t. N is the number of nodes; , These represent the movement velocities of node i along the x and y directions at time step t, respectively.

[0038] This is an adjacency matrix, where each element is defined as follows:

[0039] ;

[0040] in, This represents the adjacency matrix of nodes i and j at time step t. This represents the path loss between nodes i and j at time step t. This represents a preset path loss threshold; in this embodiment, the path loss threshold is set to 128dBm; if the path loss is ≤128dBm, the binary label of the communication link status is 1, otherwise it is 0.

[0041] Step S2.2: Set the historical observation window length to T, and construct a historical spatiotemporal graph sequence from the network topology graphs of T consecutive time steps. :

[0042] ;

[0043] Where T is the length of the historical observation window, Representing time steps , , ..., The network topology diagram.

[0044] Step S3: Standardize the extracted node features and communication link features, and perform sinusoidal time-position coding on the historical spatiotemporal graph sequence; then fuse the standardized node features with the sinusoidal time-position coding.

[0045] The specific steps of step S3 are as follows:

[0046] Step S3.1: Based on the feature vector formed by node features and communication link features, for the k-th dimension feature... Zero-mean normalization (Z-score normalization) is performed to eliminate the influence of different physical dimensions (such as velocity and propagation delay) on the gradient of the neural network; the normalization formula is as follows:

[0047] ;

[0048] in, For the k-th dimension feature The mean over the entire training set, For the k-th dimension feature Standard deviation; The k-th dimension feature is the standardized feature.

[0049] In this embodiment, the dataset is divided into a training set (90%), a validation set (5%), and a test set (5%) in chronological order.

[0050] Step S3.2: For the historical spatiotemporal graph sequence, in this embodiment, the step size of the historical spatiotemporal graph sequence is 5, and the dimension is dynamically generated based on the sine and cosine functions. Sinusoidal Positional Encoding, in this embodiment The value is 16; an independent time series identifier is generated for each time step in the historical spatiotemporal graph sequence; the encoding formula is as follows:

[0051] ;

[0052] ;

[0053] Where k is the index of the encoding dimension, k=0,1,..., ; and These represent the components of the sinusoidal time position encoding in the 2kth and 2k+1th dimensions, respectively;

[0054] Step S3.3: The sinusoidal time position code is broadcast and matrix-concatenated with the two-dimensional maneuvering velocity of the 24 nodes in each time step. After matrix concatenation, the dimension of the node features is expanded to 18 dimensions (2-dimensional original features + 16-dimensional position code), thereby giving each physical node in the historical spatiotemporal graph sequence a strict temporal anchor label.

[0055] Step S4: Extract the network topology graph of each time step in the historical observation window and map the nodes to global spatial edges; construct bidirectional temporal edges between the same nodes in adjacent time steps; form a spatio-temporal super-graph based on global spatial edges and bidirectional temporal edges.

[0056] The specific steps of step S4 are as follows:

[0057] Step S4.1: Extract the network topology graph for 5 time steps in the historical observation window, and map the nodes at time step t to global space edges. The mapping formula is as follows:

[0058] ;

[0059] in, Node i represents time step t. Let N be the global space edge mapped to node i at time step t, and N be the number of nodes; in this example, N=24. ;

[0060] Step S4.2: Between the same node in adjacent time steps (such as t and t+1), a bidirectional time edge is forcibly constructed in both forward and backward directions. At the same time, a 3-dimensional learnable tensor that follows a standard normal distribution is initialized as the edge feature of the bidirectional time edge.

[0061] Step S4.3: Topologically connect and stitch the global spatial edges and bidirectional temporal edges to completely stitch the topology of the 5 discrete time frames into a single computational entity containing 120 nodes (5*24), thus forming a spatiotemporally unified supergraph. ;in, This represents the set of nodes in the supergraph, containing node instances at all time steps. The set of edges representing a supergraph consists of global spatial edges and bidirectional temporal edges.

[0062] Step S5: Input the super graph into the graph neural network encoder, calculate the attention weights between nodes, and extract the feature vectors of all nodes in the last time step from the output matrix of the graph neural network encoder.

[0063] The specific steps of step S5 are as follows:

[0064] Step S5.1: Input the super graph into the graph neural network encoder; in this embodiment, the graph neural network encoder is a multi-head attention spatiotemporal graph encoder based on TransformerConv, and its specific structure is a stack of two TransformerConv layers:

[0065] The first layer has an input node feature dimension of 18 (2D original 2D maneuvering speed + 16D sinusoidal time-position encoding), and an input edge feature dimension of 3 (physical distance, path loss, propagation delay). The first layer uses a 128-head attention mechanism, with each head having an output dimension of 8, resulting in an output dimension of 128 × 8 = 1024.

[0066] The second layer has an input dimension of 1024 and an output dimension of 1024. It uses the same 128-head attention mechanism and sums the outputs of multiple heads.

[0067] In a single forward propagation, the physical constraints (such as terrain occlusion) experienced by concurrent computing nodes in the spatial dimension and the maneuvering inertia exhibited in the temporal dimension complete the synchronous aggregation of spatiotemporal features.

[0068] Input the supergraph into the graph neural network encoder; for node i and its spatiotemporal neighbor node j... , Let i represent the set of spatiotemporal neighbor nodes of node i. The feature vector update formula for a node based on single-head attention is:

[0069] ;

[0070] in, and Let i and j be the input feature vectors of node i and node j in the current layer, respectively. Let i be the updated feature vector after attention aggregation. and The weight matrix is ​​a learnable matrix; Let the attention weights be those between node i and node j;

[0071] Attention weight The calculation is as follows:

[0072] ;

[0073] in, , , , , All are learnable weight matrices. The dimension of the key vector; The link feature vector between node i and node j is extracted in step S1, including the physical distance between node i and node j, path loss, and propagation delay. In this embodiment, the hidden layer dimension is set to 256, and a 32-head multi-head attention mechanism is used for feature extraction. This indicates that the softmax function is used for processing.

[0074] Step S5.2: Extract the feature vectors of the last 24 nodes of the high-dimensional hidden layer output (t=5) from the output matrix of the graph neural network encoder. These feature vectors have fully absorbed the spatiotemporal evolution law in the past 5 time steps.

[0075] Step S6: Concatenate the feature vectors of any two nodes whose connectivity to be predicted in the supergraph, and input them into the multilayer perceptron (MLP) decoder to obtain the connectivity probability of the communication link between the nodes;

[0076] The specific steps of step S6 are as follows:

[0077] Step S6.1: Concatenate the feature vectors of any two nodes u and v whose connectivity is to be predicted, and then input them into the multilayer perceptron decoder:

[0078] ;

[0079] in, This indicates the operation of the multilayer perceptron decoder; and These are the final hidden layer feature vectors of nodes u and v, respectively, obtained from step S5; This indicates vector concatenation; This is the intermediate output vector obtained by processing the concatenated feature vectors through a multilayer perceptron decoder.

[0080] Step S6.2: The multilayer perceptron decoder consists of three linear mapping layers, ReLU activation, and Dropout regularization. In this embodiment, the input dimension of the multilayer perceptron decoder is 2048 (the concatenation of 1024-dimensional feature vectors from two nodes), the hidden layer dimensions of the three linear mapping layers are 512 and 256 respectively, and the output dimension is 1; finally, it is activated by the Sigmoid function. The output is transformed into the link connectivity probability in the interval (0,1). :

[0081] ;

[0082] in, This represents an exponential function with the natural constant e as its base.

[0083] Step S7: Train and test the communication link connectivity prediction model to obtain the trained communication link connectivity prediction model. Apply the communication link connectivity prediction model to predict the communication link connectivity status between any nodes in the future time step.

[0084] The specific steps of step S7 are as follows:

[0085] Step S7.1: During the training phase, based on the binary cross-entropy loss function (BCELoss), construct a loss function between the true label and the predicted link connectivity probability. As shown in the following formula:

[0086] ;

[0087] Where M is the total number of communication link samples in the current batch. This is the binary label representing the actual communication link state of the k-th communication link. This represents the predicted connectivity probability of the k-th communication link.

[0088] The Adam optimizer (initial learning rate 1e-4) is used to update parameters in reverse, and the batch size is set to 32. All learnable weight matrices (including those in the graph neural network encoder) are used. , , , , The linear layer weights of the multilayer perceptron decoder are all initialized using Xavier uniform initialization, with the bias initialized to 0. Combined with a validation set early stopping mechanism (stopping if the validation loss does not decrease after 10 consecutive rounds), a preset number of iterations is performed; in this example, 50 iterations are set.

[0089] Step S7.2: During the testing phase, the predicted link connectivity probability is... The final link state prediction result is obtained by making a decision based on a preset threshold, as follows:

[0090] ;

[0091] in, This represents the link state prediction result for the k-th communication link;

[0092] Predicting all nodes yields the complete network topology prediction matrix at time step t+1. , Each element Represents a node and The link state prediction results between them are given, where N is the number of nodes; thus, the end-to-end link prediction from the historical topology sequence to the future network state is completed.

[0093] In this embodiment, the multi-head attention mechanism, TransformerConv layer, multilayer perceptron decoder, Adam optimizer, Dropout regularization, ReLU activation function, and Sigmoid activation function used in the graph neural network encoder are all existing algorithms or modules known in the art. This invention directly adopts their standard implementation (based on PyTorch and PyTorch Geometric framework).

[0094] Figure 2 The topological dynamic evolution characteristics of two tactical network datasets used in the embodiments of the present invention—CNTM (Tactical Mobile Communication Network) and CNCM (Continuous Random Walk Network)—are demonstrated. Figure 2 The horizontal axis represents the target node, and the vertical axis represents the source node. The color intensity indicates the minimum number of hops from the source node to the target node (white indicates disconnection). Figure 2 It can be seen that the hop count distribution in the CNTM dataset exhibits obvious formation movement characteristics, with relatively stable connectivity between nodes; while the hop count distribution in the CNCM dataset presents a completely random pattern, with frequent node movements leading to extremely drastic topology changes. This comparison provides an intuitive basis for setting the difficulty of subsequent experiments and highlights the necessity of verifying the robustness of the model in extremely chaotic environments.

[0095] Table 1 compares the ablation performance errors of the method of this invention in the CNCM dataset for the core components (time edge, sinusoidal time position encoding, bidirectional connectivity).

[0096] Table 1. Comparison of Ablation Experiment Performance Errors

[0097]

[0098] Ablation experiments demonstrate that even in Continuous Random Walk Networks (CNCMs) with unpredictable node movements and extremely chaotic topology changes, this invention maintains high prediction robustness thanks to the dual protection of bidirectional temporal edges and positional encoding. Specifically, removing bidirectional temporal edges (without temporal edges) causes the communication link connectivity prediction model to encounter a performance bottleneck, with the F1 score significantly dropping to 79.7%, indicating a substantial increase in misclassifications of connected links. In contrast, the complete model (ST-SyncAttn Full) incorporating both bidirectional temporal edges and sinusoidal temporal positional encoding significantly improves the F1 score to 82.23% while maintaining a relatively stable recall rate. Furthermore, removing sinusoidal temporal positional encoding (without Time PE) or changing bidirectional edges to unidirectional (Forward Only) also leads to varying degrees of performance degradation, further confirming the indispensability and high generalization ability of the core components of this invention in combating extreme dynamic topology changes.

[0099] A bar chart comparing the link prediction F1 scores of the method of this invention and various baseline models under different historical observation step sizes is provided. Figure 3 The performance comparison table of the method of this invention under different historical observation step sizes (T=1,2,5) in the CNTM (Tactical Mobile Communication Network) dataset is shown in Table 2:

[0100] Table 2. Performance Comparison Table

[0101]

[0102] In the target-oriented tactical mobile communication network experiment, when the observation step size T=5, the F1 score of the present invention reached 93.0% and the accuracy reached 93.6%, which achieved a leapfrog performance improvement compared with the traditional optimal baseline model, fully verifying the superiority of synchronous spatiotemporal feature joint modeling.

[0103] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A high-dynamic tactical network link prediction method based on a spatiotemporal synchronous attention mechanism, characterized in that, For the target communication network, perform the following steps S1-S6 to construct a communication link connectivity prediction model, and then perform the following step S7 to predict the communication link connectivity status between any nodes in the future time step: Step S1: Obtain the operation log of the target communication network, extract the feature vector of each node, including node features and communication link features, and generate a binary label of communication link status as a criterion for communication link connectivity based on a preset path loss threshold. Step S2: Set up a historical observation window consisting of several time steps, extract the network topology graph of several consecutive time steps, and construct a historical spatiotemporal graph sequence with temporal dependencies; Step S3: Standardize the extracted node features and communication link features, and perform sinusoidal time-position coding on the historical spatiotemporal graph sequence; then fuse the standardized node features with the sinusoidal time-position coding. Step S4: Extract the network topology graph of each time step in the historical observation window and map the nodes to global spatial edges; construct bidirectional time edges between the same nodes in adjacent time steps; A spatiotemporally unified supergraph is formed based on global spatial edges and bidirectional temporal edges; Step S5: Input the super graph into the graph neural network encoder, calculate the attention weights between nodes, and extract the feature vectors of all nodes in the last time step from the output matrix of the graph neural network encoder. Step S6: Concatenate the feature vectors of any two nodes whose connectivity to be predicted in the supergraph, and input them into the multilayer perceptron decoder to obtain the communication link connectivity probability between the nodes. Step S7: Train and test the communication link connectivity prediction model to obtain the trained communication link connectivity prediction model. Apply the communication link connectivity prediction model to predict the communication link connectivity status between any nodes in the future time step.

2. The high-dynamic tactical network link prediction method based on spatiotemporal synchronous attention mechanism according to claim 1, characterized in that, The specific steps of step S1 are as follows: Step S1.1: Extract the node features and communication link features of the target communication network, including the two-dimensional maneuvering speed of each node, as node features; It also includes the physical distance between nodes (Distance), electromagnetic wave propagation delay (PropDelay), and path loss as characteristics of the communication link; Step S1.2: Set a preset path loss threshold. If the path loss between two nodes is lower than the preset threshold, then the communication link between the two nodes is determined to be connected in the current time step, and the communication link status is labeled with a binary tag. Otherwise, determine that the communication link between the two nodes is broken at the current time step, and assign a binary label to the communication link status. ;in, A binary label representing the communication link status between node i and node j at time step t.

3. The high-dynamic tactical network link prediction method based on spatiotemporal synchronous attention mechanism according to claim 2, characterized in that, The specific steps of step S2 are as follows: Step S2.1: At time step t, represent the target communication network as a network topology graph. ,in, A set of nodes; N represents a node, and N is the number of nodes; Let F represent the node feature matrix, where F represents the dimension of the node features, corresponding to the two-dimensional maneuver velocity of the node; each row of the node feature matrix represents the two-dimensional maneuver velocity of each node at time step t. ； ； in, This represents the two-dimensional maneuvering velocity of node i at time step t. N is the number of nodes; , These represent the movement velocities of node i along the x and y directions at time step t, respectively. This is an adjacency matrix, where each element is defined as follows: ； in, This represents the adjacency matrix of nodes i and j at time step t. This represents the path loss between nodes i and j at time step t. This indicates the preset path loss threshold; Step S2.2: Set the historical observation window length to T, and construct a historical spatiotemporal graph sequence from the network topology graphs of T consecutive time steps. : ； Where T is the length of the historical observation window, Representing time steps , , ..., The network topology diagram.

4. The high-dynamic tactical network link prediction method based on spatiotemporal synchronous attention mechanism according to claim 3, characterized in that, The specific steps of step S3 are as follows: Step S3.1: Based on the feature vector formed by node features and communication link features, for the k-th dimension feature... Zero-mean standardization is performed using the following formula: ； in, For the k-th dimension feature The mean over the entire training set, For the k-th dimension feature Standard deviation; The k-th dimension feature is the standardized feature. Step S3.2: For the historical spatiotemporal graph sequence, dynamically generate the dimension as follows. The sinusoidal time position encoding; the encoding formula is as follows: ； ； Where k is the index of the encoding dimension, k=0,1,... ; and These represent the components of the sinusoidal time position encoding in the 2kth and 2k+1th dimensions, respectively; Step S3.3: The sinusoidal time position code is broadcast and matrix-concatenated with the two-dimensional maneuvering velocity of the node in each time step; after matrix concatenation, the dimension of the node feature is expanded to 18 dimensions.

5. The high-dynamic tactical network link prediction method based on spatiotemporal synchronous attention mechanism according to claim 4, characterized in that, The specific steps of step S4 are as follows: Step S4.1: Extract the network topology graph for each time step in the historical observation window, and map the nodes at time step t to global space edges. The mapping formula is: ； in, Node i represents time step t. Let N be the global space edge mapped to node i at time step t, and N be the number of nodes. Step S4.2: Between the same node in adjacent time steps, force the construction of bidirectional time edges in both forward and backward directions. At the same time, initialize a 3-dimensional learnable tensor that follows a standard normal distribution as the edge feature of the bidirectional time edge. Step S4.3: Connect the global spatial edges and the bidirectional temporal edges topologically to form a spatiotemporally unified supergraph. ;in, This represents the set of nodes in the supergraph, containing node instances at all time steps. The set of edges representing a supergraph consists of global spatial edges and bidirectional temporal edges.

6. The high-dynamic tactical network link prediction method based on spatiotemporal synchronous attention mechanism according to claim 5, characterized in that, The specific steps of step S5 are as follows: Step S5.1: Input the supergraph into the graph neural network encoder; for node i and its spatiotemporal neighbor node j, , Let i represent the set of spatiotemporal neighbor nodes of node i. The feature vector update formula for a node based on single-head attention is: ； in, and Let i and j be the input feature vectors of node i and node j in the current layer, respectively. Let i be the updated feature vector after attention aggregation. and The weight matrix is ​​a learnable matrix; Let the attention weights be those between node i and node j; Attention weight The calculation is as follows: ； in, , , , , All are learnable weight matrices. The dimension of the key vector; This represents the link feature vector between node i and node j; This indicates the processing by the softmax function; Step S5.2: Extract the feature vectors of all nodes at the last time step of the high-dimensional hidden layer output from the output matrix of the graph neural network encoder.

7. The high-dynamic tactical network link prediction method based on spatiotemporal synchronous attention mechanism according to claim 6, characterized in that, The specific steps of step S6 are as follows: Step S6.1: Concatenate the feature vectors of any two nodes u and v whose connectivity is to be predicted, and then input them into the multilayer perceptron decoder: ； in, This indicates the operation of the multilayer perceptron decoder; and These are the final hidden layer feature vectors of nodes u and v, respectively. This indicates vector concatenation; This is the intermediate output vector obtained by processing the concatenated feature vectors through a multilayer perceptron decoder. Step S6.2: The multilayer perceptron decoder consists of three linear mapping layers, ReLU activation, and Dropout regularization, ultimately activating via the Sigmoid activation function. The output is transformed into the link connectivity probability in the interval (0,1). : ； in, This represents an exponential function with the natural constant e as its base.

8. The high-dynamic tactical network link prediction method based on spatiotemporal synchronous attention mechanism according to claim 7, characterized in that, The specific steps of step S7 are as follows: Step S7.1: During the training phase, construct the loss function between the true labels and the predicted link connectivity probabilities. As shown in the following formula: ； Where M is the total number of communication link samples in the current batch. This is the binary label representing the actual communication link state of the k-th communication link. This represents the predicted connectivity probability of the k-th communication link. The Adam optimizer is used to update parameters in reverse, combined with a validation set early stopping mechanism, to perform a preset number of iterations; Step S7.2: During the testing phase, the predicted link connectivity probability is... The final link state prediction result is obtained by making a decision based on a preset threshold, as follows: ； in, This represents the link state prediction result for the k-th communication link; Predicting all nodes yields the complete network topology prediction matrix at time step t+1. , Each element Represents a node and The link state prediction results between them, where N is the number of nodes.

Images