Unmanned aerial vehicle power distribution communication network post-disaster emergency communication flow prediction method and device

Abstract

The invention discloses an unmanned aerial vehicle power distribution communication network post-disaster emergency communication traffic prediction method and device, and the method comprises the steps: inputting unmanned aerial vehicle node traffic data into a time sequence modeling network with gating, capturing a time rule through gating to obtain time sequence embedding, generating correction relation embedding based on the time sequence embedding, and carrying out the fusion calculation of the two. And obtaining the predicted flow of the future target time period. According to the post-disaster emergency communication traffic prediction method and device for the unmanned aerial vehicle power distribution communication network provided by the invention, the traffic time dynamic law is captured through time sequence embedding, the spatial conduction effect is quantized through correction relation embedding, and the traffic time dynamic law and the spatial conduction effect are deeply fused through spatial-temporal feature fusion calculation, so that the time change feature of the traffic can be accurately extracted; and flow prediction is realized based on spatial-temporal feature deep fusion, so that the problem of insufficient flow prediction accuracy caused by the fact that a flow time rule and a spatial conduction effect under post-disaster quasi-static topology are not fused in an existing method in a post-disaster emergency scene of the unmanned aerial vehicle power distribution communication network can be solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of communication technology of smart grid, in particular to a UAV power distribution communication network post-disaster emergency communication traffic prediction method and device. BACKGROUND

[0002] In recent years, extreme natural disasters such as typhoons, earthquakes, and floods have occurred frequently, which not only cause serious damage to power infrastructure, but also can directly lead to a large area of power distribution communication network paralysis; in the emergency scenario of such post-disaster communication interruption, unmanned aerial vehicles (UAVs) have become a key support equipment for post-disaster emergency communication due to their core advantages of flexible deployment, rapid response, and strong environmental adaptability. They can quickly build a temporary relay network to replace the functions of damaged fixed communication nodes, restore the data transmission link between power distribution automation terminals (such as FTU / DTU) and control centers, and thus provide stable protection for key emergency operations such as fault location, load transfer, and network reconstruction. However, due to the limitation of communication reliability in dynamic environment, the efficiency of UAV emergency communication is still insufficient, so the node traffic prediction in the post-disaster emergency scenario of UAV power distribution communication network is crucial. In the related technology of UAV node traffic prediction in the post-disaster emergency scenario of UAV power distribution communication network, a single deep learning model can be used to capture the time dependence of traffic. Meanwhile, artificial intelligence has been applied in multiple directions in the field of power emergency, including pre-disaster allocation optimization by prepositioning UAV positions through vulnerability models, post-disaster power distribution network reconstruction using deep reinforcement learning, and related research focusing on UAV trajectory planning or power control.

[0003] However, a single deep learning model can only capture the time dependence of traffic, but it does not model the spatial correlation between nodes and the traffic conduction effect. The vulnerability model focuses on pre-disaster allocation of UAV positions, which is irrelevant to real-time traffic prediction in disaster. Deep reinforcement learning is used for post-disaster power distribution network reconstruction, which relies on online interaction and does not involve traffic spatiotemporal features. Trajectory planning or power control research focuses on UAV parameter regulation and lacks quantitative analysis of traffic spatiotemporal regularity. These limitations collectively result in insufficient accuracy of UAV node traffic prediction. SUMMARY

[0004] The present application provides a UAV power distribution communication network post-disaster emergency communication traffic prediction method and device to solve the problem of insufficient accuracy of UAV node traffic prediction in the post-disaster emergency scenario of UAV power distribution communication network due to the lack of fusion of traffic time regularity and spatial conduction effect under post-disaster quasi-static topology in existing methods.

[0005] To achieve the above-mentioned purpose, the present application provides a UAV power distribution communication network post-disaster emergency communication traffic prediction method, comprising: In the post-disaster emergency communication scenario of the UAV power distribution communication network, traffic data of a plurality of UAV nodes is obtained; input the traffic data into a time series modeling network with gating, capture time rules based on the gating structure, and obtain time series embedding of the plurality of unmanned aerial vehicle nodes; generate a revised relationship embedding of the traffic data based on the time series embedding, specifically: combine the quasi-static topological characteristics formed by the post-disaster unmanned aerial vehicle hovering, encode the topological relationship of the plurality of unmanned aerial vehicle nodes into a multi-hot binary vector and construct an adjacency matrix, input the adjacency matrix and the time series embedding into a graph convolution network, quantify the inter-node traffic conduction effect through inter-layer neighbor feature aggregation, and generate the revised relationship embedding of the traffic data; Through spatiotemporal feature fusion calculation on the time series embedding and the revised relationship embedding, the predicted traffic of the plurality of unmanned aerial vehicle nodes in the future target period is obtained.

[0006] The application uses the gating structure of the time series modeling network with gating to accurately screen the key time rules of unmanned aerial vehicle node traffic, effectively captures the dynamic characteristics of traffic changes over time and converts them into time series embedding, making up for the shortcomings of existing methods in fusing time dynamics, and can improve traffic prediction accuracy. Based on the time series embedding, the revised relationship embedding of the traffic data is generated. On the one hand, by combining the key scene feature of the quasi-static topology formed by the post-disaster unmanned aerial vehicle hovering, the node topological relationship is encoded into an adjacency matrix, and the abstract spatial connection relationship is converted into structured information recognizable by the model, providing a carrier for capturing spatial conduction effects. On the other hand, the adjacency matrix and the time series embedding, which have captured the time dynamics through time series modeling, are input into a graph convolution network, and the spatial conduction effect and the time dynamic characteristics interact and quantitatively fuse in the model through inter-layer neighbor feature aggregation. The finally generated revised relationship embedding not only contains the traffic conduction characteristics brought by the spatial topology, but also associates the dynamic rules in the time dimension, providing spatial feature support for subsequent spatiotemporal feature fusion calculation. Finally, through spatiotemporal feature fusion calculation on the time series embedding carrying the time dynamic rules of unmanned aerial vehicle node traffic and the revised relationship embedding containing the spatial conduction effect based on the post-disaster quasi-static topology, the depth fusion of time features and spatial features is realized, the spatiotemporal correlation characteristics of traffic can be captured, and the traffic prediction accuracy is improved.

[0007] Compared with the prior art, the application first captures the time dynamic rules of traffic through time series embedding, then quantifies the spatial conduction effect with the help of revised relationship embedding, and finally fuses the two through spatiotemporal feature fusion calculation. The time variation characteristics of traffic are accurately extracted, the spatial conduction correlation between nodes is captured based on the post-disaster quasi-static topology, and finally the depth fusion of spatiotemporal features is realized for traffic prediction. Therefore, it can solve the problem of insufficient accuracy of unmanned aerial vehicle node traffic prediction caused by the failure of existing methods to fuse traffic time rules and spatial conduction effects under post-disaster quasi-static topology in the post-disaster emergency scenario of unmanned aerial vehicle power distribution communication network.

[0008] As a preferred solution, the traffic data is input into a time modeling network with a gate, and the time regularity is captured based on the gate structure to obtain the time sequence embedding of the plurality of UAV nodes, specifically: After the traffic data is arranged in a matrix form, it is input into the time modeling network with a gate; The time dimension regularity of the traffic data is filtered based on the reset gate and the update gate built in the time modeling network; For each UAV node, the last hidden state output by the time modeling network after processing the corresponding traffic data is taken as the sequence embedding; wherein the sequence embedding contains the time dimension regularity; The sequence embeddings of all UAV nodes are integrated into a matrix to form the time sequence embedding of the plurality of UAV nodes.

[0009] The preferred solution filters the time dimension regularity by means of the reset gate and the update gate built in the network, can accurately strip the invalid noise in the post-disaster traffic data, and focus on the core time characteristics; taking the last hidden state corresponding to each UAV node as the sequence embedding can maximize the concentration of key time sequence information of node traffic, and then integrating into a time sequence embedding matrix provides structured and high-quality time domain basic data for subsequent airspace feature modeling. This design solves the problem that a single time sequence model cannot accurately extract the time regularity of post-disaster dynamic traffic, improves the pertinence and usability of time sequence features, and lays a reliable time domain data support for spatio-temporal feature fusion calculation.

[0010] As a preferred solution, the reset gate is used to filter the target historical information to be retained, and the update gate is used to control the memory and forgetting proportion of historical data.

[0011] The preferred solution clearly defines the core functions of the reset gate and the update gate in the time modeling network, making the logic of the gate structure more explicit. The reset gate selectively filters the target historical information to be retained, effectively filtering irrelevant historical interference terms in post-disaster traffic data, ensuring that key time sequence features are not lost; the update gate dynamically controls the memory and forgetting proportion of historical data, which can flexibly adapt to complex changes such as post-disaster traffic peak migration and sudden fluctuations, avoiding excessive memory of outdated data or blind forgetting of key regularities. The synergistic effect of the two makes the time modeling network accurately capture the dynamic time regularity of post-disaster traffic, and the generated time sequence embedding is more consistent with the actual traffic variation characteristics, solving the problem of complex post-disaster traffic time sequence regularity and difficulty in accurate extraction, and further improving the quality of basic data for subsequent spatio-temporal feature fusion calculation.

[0012] As a preferred solution, the quasi-static topological characteristics formed by the post-disaster UAV hovering are combined to encode the topological relationship between the plurality of UAV nodes into a multi-hot binary vector and construct an adjacency matrix, the adjacency matrix and the time sequence embedding are input into a graph convolution network, the inter-node flow conduction effect is quantified through inter-layer neighbor feature aggregation, and a corrected relationship embedding of the flow data is generated, specifically: The quasi-static topological characteristics formed by the post-disaster UAV hovering are combined to encode the topological relationship between the plurality of UAV nodes into a multi-hot binary vector, and an adjacency matrix representing whether there is a connection between nodes is constructed based on the multi-hot binary vector. The adjacency matrix and the time sequence embedding are input into the graph convolution network, and the node's own features and neighbor node features are aggregated and calculated between the network layers of the graph convolution network to generate the corrected relationship embedding of the flow data.

[0013] The preferred solution refines the generation logic of the corrected relationship embedding, combines the quasi-static topological characteristics of the post-disaster UAV hovering, encodes the topological relationship into a multi-hot binary vector and constructs an adjacency matrix, accurately represents the connection relationship between the UAV nodes, and ensures that the basic data of the airspace feature modeling is accurate and reliable. The adjacency matrix and the time sequence embedding are input into the graph convolution network, the node's own and neighbor features are aggregated between the layers, the time domain features and the space domain features are preliminarily fused, and the flow conduction effect between nodes is effectively quantified. This design makes up for the defect of traditional models that ignore the spatial correlation between nodes, so that the generated corrected relationship embedding can reflect not only the time sequence features of the node itself but also the mutual influence of the neighbor nodes, providing comprehensive airspace feature support for spatiotemporal feature fusion calculation, improving the consideration of the conduction effect between nodes for flow prediction, and reducing the prediction deviation caused by ignoring the spatial relationship.

[0014] As a preferred solution, the adjacency matrix and the time sequence embedding are input into the graph convolution network, the node's own features and neighbor node features are aggregated and calculated between the network layers of the graph convolution network, and the corrected relationship embedding of the flow data is generated, specifically: The adjacency matrix and the time sequence embedding are input into the graph convolution network, and the neighbor nodes of each node are located according to the adjacency matrix between the layers of the graph convolution network to obtain a neighbor node set. The time sequence embedding of the current node in the graph convolution network and the time sequence embedding of the corresponding node in the neighbor node set are aggregated and calculated to quantize the flow conduction effect between nodes and integrate the flow conduction effect into the features of the current node to obtain the features of each node fused with the conduction effect. All node features fused with the conduction effect are integrated into a matrix to obtain the corrected relationship embedding of the flow data.

[0015] The preferred scheme further refines the feature aggregation process of the graph convolution network, accurately locates the neighbor node set of each node through the adjacency matrix, ensures the accuracy of the object of feature aggregation, and avoids the interference of irrelevant node features. The time sequence embedding of the current node and the neighbor node is aggregated and calculated, which can more accurately quantify the flow conduction effect between nodes and deeply integrate it into the current node feature, so that the feature of each node can fully reflect its interaction with the neighborhood. By integrating the features of all nodes that fuse and conduct effects to form a modified relationship embedding matrix, the structural degree and accuracy of the spatial feature data are further improved. This design solves the problems of ambiguous spatial feature aggregation and inaccurate conduction effect quantization, making the modified relationship embedding more truly reflect the spatial dependence between unmanned aerial vehicle nodes, providing higher quality spatial features for spatio-temporal feature fusion calculation, and helping to improve the accuracy of the prediction results.

[0016] As a preferred scheme, the time sequence embedding and the modified relationship embedding are input into a fully connected layer, and the time sequence embedding and the modified relationship embedding are fused and calculated by the neuron weight parameters of the fully connected layer to realize joint modeling of spatio-temporal features, and the predicted flow of the plurality of unmanned aerial vehicle nodes in the future target period is obtained. The time sequence embedding and the modified relationship embedding are input into a fully connected layer, and the time sequence embedding and the modified relationship embedding are fused and calculated by the neuron weight parameters of the fully connected layer to realize joint modeling of spatio-temporal features, and the predicted flow of the plurality of unmanned aerial vehicle nodes in the future target period is obtained. The fully connected layer is established based on the loss function to calculate the prediction error, and the optimizer is used to adjust the neuron weights by back propagation until the loss value decreases to a preset threshold.

[0017] The preferred scheme specifies the core implementation path of spatio-temporal feature fusion calculation, and the time sequence embedding and the modified relationship embedding are fused and calculated by the neuron weight parameters of the fully connected layer to realize deep interaction and complementarity of time domain and spatial domain features, solving the problem of traditional model spatio-temporal feature fragmentation. At the same time, the establishment logic of the fully connected layer based on the loss function to calculate the prediction error and the optimizer to adjust the weights by back propagation is specified, ensuring that the parameter optimization direction of the fully connected layer is highly consistent with the flow prediction target. This design allows the model to fully utilize the flow time law of the time domain and the node conduction effect of the spatial domain to achieve accurate prediction of the post-disaster unmanned aerial vehicle node flow, effectively avoids the prediction deviation caused by single feature modeling, reduces the secondary interruption risk caused by node overload, and improves the reliability and stability of the post-disaster emergency communication network, adapting to the actual needs of post-disaster emergency communication.

[0018] As a preferred scheme, the fully connected layer is established based on the loss function to calculate the prediction error, and the optimizer is used to adjust the neuron weights by back propagation until the loss value decreases to a preset threshold, specifically: The historical traffic sample set in the post-disaster emergency communication scenario of the unmanned aerial vehicle power distribution communication network is acquired, and the historical time sequence embedding and the historical correction relationship embedding generated by the historical traffic sample set are taken as training inputs and transmitted to a preset initial full connection layer; The initial full connection layer is controlled to perform spatio-temporal feature fusion on the historical time sequence embedding and the historical correction relationship embedding through neuron weights, to obtain a preliminary predicted traffic value of each unmanned aerial vehicle node history corresponding period; Based on the real traffic value in the historical traffic sample set and the preliminary predicted traffic value, a loss value is calculated according to a loss function; According to the optimizer, the first-order moment estimation and the second-order moment estimation of the gradient are calculated by taking the loss value as the input, the learning rate is set, the neuron weights of the initial full connection layer are adjusted, and the error is continuously optimized according to the adjusted weights until the loss value decreases to a preset threshold, to obtain the full connection layer.

[0019] In this preferred scheme, the historical time sequence embedding and the historical correction relationship embedding generated by the historical traffic sample set are taken as training inputs, ensuring the consistency of the training data and the actual application scenario and improving the generalization ability of the model. The initial full connection layer generates a preliminary predicted traffic value, the loss is calculated in combination with the real traffic value, and then the weights are adjusted through the optimizer in reverse propagation until the loss value reaches the preset threshold, forming a closed-loop parameter optimization mechanism. This design can effectively avoid the problem of poor feature fusion effect caused by randomization of the full connection layer parameters, and through step-by-step optimization, it ensures that the neuron weights accurately adapt to the spatio-temporal feature fusion requirements, while preventing overfitting, so that the full connection layer has stable and reliable spatio-temporal feature fusion calculation capability. Finally, the model can output accurate prediction results when facing traffic changes in different stages after the disaster, improving the practicality and landing of the scheme.

[0020] As a preferred scheme, the first-order moment estimation and the second-order moment estimation are used to cooperatively guide the update direction and step length of the neuron weight adjustment of the initial full connection layer.

[0021] This optimized scheme clearly defines the core roles of first-order moment estimation and second-order moment estimation. These two methods work together to guide the weight updates of the initial fully connected layer, resolving the problems of ambiguous weight adjustment direction and unreasonable step size in traditional optimizers. First-order moment estimation effectively controls the direction of weight updates, ensuring that parameter optimization always progresses towards reducing the loss value and avoiding deviation from the optimal solution. Second-order moment estimation dynamically adjusts the update step size, adaptively matching the optimization rhythm based on parameter gradient changes, avoiding oscillations caused by excessively large step sizes or slow convergence caused by excessively small step sizes. The combination of these two methods significantly improves the efficiency and accuracy of weight updates, shortens the model training cycle, allows the fully connected layer to reach a stable optimal state more quickly, and enhances the model's adaptability to complex changes in post-disaster traffic, ensuring the stability and accuracy of prediction results and providing efficient model training support for post-disaster emergency communication traffic prediction.

[0022] As a preferred solution, in the post-disaster emergency communication scenario of the UAV power distribution communication network, traffic data of several UAV nodes is obtained, specifically: In the post-disaster emergency communication scenario of the UAV power distribution communication network, the raw traffic data of the aforementioned UAV nodes is obtained; Remove target outliers from the original traffic data, and group the original traffic data after removing target outliers according to the drone node ID to generate a traffic table for each drone node. The average value of the traffic table for each drone node is taken, and the traffic values ​​at missing times are filled in to obtain the traffic data of several drone nodes.

[0023] This optimized scheme effectively eliminates the impact of traffic surges caused by noise interference on subsequent data processing by removing outliers from the original data. It generates traffic tables based on UAV node IDs, ensuring the independence and specificity of data for each node. The averaging operation standardizes the time granularity, making the data more regular, while filling in missing values ​​ensures data integrity. This design solves the problems of messy, irregular, noisy, and missing values ​​in the original traffic data, producing regular, accurate, and complete output traffic data, providing high-quality input data for gated time-series modeling networks. By improving the quality of the basic data, errors in subsequent time-series embedding and correction relationship embedding processes are reduced from the source, laying a solid data foundation for the accuracy and reliability of the entire prediction scheme and ensuring that the model can accurately capture the spatiotemporal patterns of post-disaster traffic.

[0024] This application also provides a device for predicting post-disaster emergency communication traffic in a UAV power distribution communication network, including a data acquisition module, a gating module, a convolution module, and a prediction module; The acquisition module is used to acquire traffic data of several UAV nodes in a post-disaster emergency communication scenario of the UAV power distribution communication network. The gating module is used to input the traffic data into a gated temporal modeling network, capture the temporal patterns based on the gating structure, and obtain the temporal embedding of the plurality of UAV nodes. The convolutional module is used to generate the corrected relationship embedding of the traffic data based on the temporal embedding. Specifically, it combines the quasi-static topological characteristics formed by the hovering of UAVs after the disaster, encodes the topological relationship of the several UAV nodes into a multi-hot binary vector and constructs an adjacency matrix, inputs the adjacency matrix and the temporal embedding into the graph convolutional network, and quantifies the traffic transmission effect between nodes through inter-layer neighbor feature aggregation to generate the corrected relationship embedding of the traffic data. The prediction module is used to obtain the predicted traffic flow of the plurality of UAV nodes in the future target time period by performing spatiotemporal feature fusion calculation on the temporal embedding and the modified relation embedding.

[0025] This application also provides a storage medium storing a computer program, which is called and executed by a computer to implement the post-disaster emergency communication traffic prediction method for UAV power distribution communication network as described above.

[0026] This application also provides a computer program product, including a computer program or instructions, which, when executed by a communication device, implements the post-disaster emergency communication traffic prediction method for UAV power distribution communication networks as described above. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating a method for predicting emergency communication traffic in a UAV power distribution communication network after a disaster, as provided in an embodiment of this application. Figure 2 This is a flowchart of communication traffic prediction provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a post-disaster emergency communication traffic prediction device for a UAV power distribution communication network provided in an embodiment of this application. Detailed Implementation

[0028] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0029] In the description of this application, unless otherwise stated, "a number" means two or more.

[0030] The present application provides a method for predicting emergency communication traffic in a UAV power distribution communication network after a disaster. This method aims to address the problems that arise when UAVs construct temporary power distribution communication networks after a disaster. These problems include dynamic changes in traffic caused by emergency services, the difficulty of capturing both the temporal patterns of traffic and the spatial topological relationships of nodes using a single model, which leads to low accuracy in overload warnings for key relay nodes and a tendency to cause secondary communication interruptions.

[0031] Example 1: Please see Figure 1 The embodiments of this application provide a method for predicting emergency communication traffic in a UAV power distribution communication network after a disaster, including S1~S4, and the specific implementation steps are as follows: S1. In the post-disaster emergency communication scenario of the UAV power distribution communication network, acquire traffic data of several UAV nodes.

[0032] Step S1 in this embodiment of the application is specifically as follows: In the post-disaster emergency communication scenario of the UAV power distribution communication network, the raw traffic data of several UAV nodes can be obtained, which can directly reflect the real-time communication load status of each node.

[0033] Remove target outliers from the original traffic data to eliminate traffic spikes caused by noise interference; Based on the drone node ID, the raw traffic data after removing outliers is grouped to generate a traffic table for each drone node.

[0034] For the flow tables of each node, a 10-minute time window is used to calculate the average of all flow values ​​within the window as the representative flow for that period. By standardizing the time granularity, the interference of random sampling fluctuations on the flow trend is reduced, making the data change patterns clearer. If there are missing traffic data at certain times in the traffic table, linear interpolation is used to fill them in: taking the two valid times before and after the missing time as references, the traffic value corresponding to the missing time is obtained by linear calculation and then filled in to ensure the continuity of the traffic data sequence, and finally the traffic data of several drone nodes are obtained after processing.

[0035] The formula for calculating the average value is: The calculation formula corresponding to the linear interpolation method is: in, This is the average flow rate over 10 minutes. It is the first Traffic volume within 10 minutes, It is the first The number of traffic values ​​to be taken within a 10-minute period; Is The interpolation result at the given point. Assume there are two points. and They are respectively Flow value at any time ,and Flow value at any time , The result calculated after linear interpolation Flow rate at any given moment.

[0036] In this embodiment, S1 effectively eliminates the impact of traffic surges caused by noise interference on subsequent data processing by removing outliers from the original data. Traffic tables are generated based on UAV node IDs, ensuring the independence and specificity of data for each node. The averaging operation unifies the time granularity, making the data more regular, while filling in missing values ​​ensures data integrity. This design solves the problems of messy, irregular, noisy, and missing values ​​in the original traffic data, resulting in regular, accurate, and complete output traffic data, providing high-quality input data for gated time-series modeling networks. By improving the quality of the basic data, errors in subsequent time-series embedding and correction relationship embedding processes are reduced from the source, laying a solid data foundation for the accuracy and reliability of the entire prediction scheme and ensuring that the model can accurately capture the spatiotemporal patterns of post-disaster traffic.

[0037] S2. Input the traffic data into a gated temporal modeling network, capture the temporal patterns based on the gated structure, and obtain the temporal embedding of several UAV nodes.

[0038] Step S2 in this embodiment of the application is specifically as follows: Based on the temporal embedding layer, the traffic data is organized into a matrix form to represent the features of the sequence input, denoted as... ; where, matrix Represents the features of the sequence input. This represents the feature dimension at each time step. Indicates the length of the sequence. This represents the matrix transpose operation. Assume we have... One drone node, then These are the features collected; matrix The input is a gated time-series modeling network, i.e., a GRU network. The time-dimensional patterns of the traffic data are filtered out based on the reset and update gates of the GRU network. Specifically, the reset gate is used to select target historical information to be retained, and the update gate is used to control the ratio of historical data retention to forgetting. Using a dedicated weight matrix Input x for the current time step t Hidden state from the previous moment The concatenated matrix undergoes a linear transformation, and the result is then input into the sigmoid function to calculate the output value, which is used for subsequent candidate hidden state calculations; update gate Through its own weight matrix It performs the same linear transformation and sigmoid activation process as the reset gate, and dynamically controls the ratio of memory to forgetting of historical data by combining the memory update expression. The closer it is to 1, the more it represents the hidden state of history. The higher the level of memory, the greater the proportion of historical data forgotten; the closer to 0, the greater the proportion of forgotten data. For each drone node, the last hidden state is output after processing the corresponding traffic data using the GRU network. As sequence embedding Among them, sequence embedding contains patterns in the time dimension; The sequence embeddings of all drone nodes are integrated into a matrix to form the temporal embeddings of several drone nodes.

[0039] The specific calculation formulas are as follows: (1) Temporal embedding definition: the drone node time step Historical traffic time series data Input the GRU network and remove the last hidden state. Sequence embedding as drone node traffic (Notice, Then the temporal embedding of all nodes is denoted as: in, This represents the traffic timing embedding of all drone nodes. Indicates the embedding size, i.e. It is the number of hidden units in the GRU; This represents the matrix input to the GRU network.

[0040] (2) A GRU network contains the following units: ① The formula for calculating the door reset is: in, This is the weight matrix for resetting the gate. During the calculation, the hidden state from the previous time step must first be considered. Input at the current moment By splicing them together to form a new matrix, and then through... Perform a linear transformation on the concatenated matrix (specifically, a matrix multiplication operation).

[0041] ②The formula for calculating the candidate hidden state is: ③ The formula for calculating the updated door is: ④ The memory update expression is: in, The closer to 1, the more data is remembered; the closer to 0, the more data is forgotten. W is the weight matrix of the update gate, and W represents the weight matrix designed for the candidate hidden state.

[0042] This embodiment S3 utilizes the network's built-in reset and update gates to filter time-dimensional patterns, accurately removing invalid noise from post-disaster traffic data and focusing on core time features. By using the last hidden state corresponding to each UAV node as a sequence embedding, the key time-series information of node traffic can be condensed to the greatest extent. This information is then integrated into a time-series embedding matrix, providing structured and high-quality time-domain foundational data for subsequent spatial feature modeling. This design solves the problem that a single time-series model is difficult to accurately extract the time patterns of post-disaster dynamic traffic, improves the relevance and usability of time-series features, and lays a reliable time-domain data foundation for spatiotemporal feature fusion calculation. Furthermore, this embodiment clearly defines the core functions of the reset gate and update gate in the time-series modeling network, making the operational logic of the gating structure clearer. The reset gate selectively filters the target historical information to be retained, effectively filtering irrelevant historical interference items in post-disaster traffic data and ensuring that key temporal features are not lost. The update gate dynamically controls the ratio of historical data to memory and forgetting, flexibly adapting to complex changes such as post-disaster traffic peak migration and sudden fluctuations, avoiding excessive memory of outdated data or blind forgetting of key patterns. The synergistic effect of the two gates allows the time-series modeling network to accurately capture the dynamic temporal patterns of post-disaster traffic, and the generated time-series embeddings are more closely aligned with the actual traffic change characteristics. This solves the problem of complex and difficult-to-accurate extraction of post-disaster traffic temporal patterns, further improving the quality of the basic data for subsequent spatiotemporal feature fusion calculations.

[0043] S3. Based on temporal embedding, the modified relationship embedding of traffic data is generated. Specifically, the topological relationship of several UAV nodes is encoded into a multi-hot binary vector and an adjacency matrix is ​​constructed by combining the quasi-static topological characteristics formed by UAV hovering after the disaster. The adjacency matrix and temporal embedding are then used to embed the input graph convolutional network. The traffic transmission effect between nodes is quantified by aggregating the inter-layer neighbor features to generate the modified relationship embedding of traffic data.

[0044] Step S3 in this embodiment of the application is specifically as follows: Based on the relation embedding layer, and combined with the quasi-static topology characteristics formed by UAV hovering after a disaster, the topological relationships between several UAV nodes are encoded into multi-hot binary vectors. Based on multi-heat binary vector Construct an adjacency matrix that represents whether there are connections between nodes. ;in, This represents the number of types of relationships between drone nodes. This embodiment of the invention focuses only on topological connections. The adjacency matrix A is in its first... Line number Column elements Used to characterize the The node and the first Does each node have a direct connection?

[0045] The adjacency matrix A and the temporal embedding are input together into LightGCN (Light Graph Convolutional Network), a lightweight graph convolutional network; during the inter-layer computation of LightGCN, each node is located based on the adjacency matrix A. The neighbor nodes are obtained to get the neighbor node set. ; where, for any given node, the node is... For example, its neighbor node set clearly defines the range of adjacent nodes that will subsequently participate in feature updates, while the nodes Feature updates must follow specific rules, namely, node... The update rule is used to calculate the association between its own features and those of its neighbors; According to nodes The update rule combines the temporal embedding of the current node and the set of neighboring nodes in the graph convolutional network. The temporal embeddings of the corresponding nodes are aggregated and calculated to synchronously quantify the traffic transmission effect between nodes. The traffic transmission effect is then integrated into the features of the current node to obtain the features of the fused transmission effect of each node. Since the calculation of LightGCN is performed iteratively in layers (each layer is denoted as layer l), the "features of the fused transmission effect of each node" obtained in this step are essentially the feature representations of each node at layer l, i.e., the node features of layer l.

[0046] By integrating the features of the transmission effect of all nodes into a matrix, the corrected relationship embedding of the traffic data is obtained.

[0047] The specific calculation method for obtaining the corrected relationship embedding of traffic data based on the adjacency matrix A and temporal embedding is as follows: When performing inter-layer computation in LightGCN, it is assumed that each node in the topology... Each has a feature vector Indicate its characteristics, among which It is the dimension of the features, and has an adjacency matrix. Represents the connectivity relationships in a graph, where This refers to the number of nodes. The elements of the adjacency matrix. Represents a node and nodes Are there any edges connecting them? Given a node. Its set of neighboring nodes is Then the node The update rule can be expressed as: in, It is a node In the Layer feature representation.

[0048] Applying the above update rules to all nodes allows us to build a LightGCN layer. Assume we have... The nth convolutional kernel, then the nth... The node features of a layer can be represented as: in, It is the first The node feature matrix of the layer, It is an adjacency matrix with added self-connections. It is a matrix of angle numbers.

[0049] This embodiment S3 refines the generation logic of the modified relation embedding. Combining the quasi-static topological characteristics of post-disaster UAV hovering, it accurately represents the connection relationships between UAV nodes by encoding topological relationships into multi-hot binary vectors and constructing an adjacency matrix, ensuring the accuracy and reliability of the basic data for spatial feature modeling. The adjacency matrix and temporal embedding are input into a graph convolutional network. By aggregating the features of the node itself and its neighbors between layers, a preliminary fusion of temporal and spatial features is achieved, effectively quantifying the flow transmission effect between nodes. This design overcomes the deficiency of traditional models that ignore the spatial correlation between nodes, allowing the generated modified relation embedding to reflect both the temporal characteristics of the node itself and the mutual influence of neighboring nodes. This provides comprehensive spatial feature support for spatiotemporal feature fusion calculation, enhances the consideration of the flow prediction effect between nodes, and reduces prediction bias caused by ignoring spatial relationships. Furthermore, by accurately locating the neighbor set of each node through the adjacency matrix, the accuracy of feature aggregation is ensured, avoiding interference from irrelevant node features. Targeted aggregation calculations of the temporal embeddings of the current node and its neighbors more accurately quantify the flow transmission effect between nodes and deeply integrate it into the current node's features, ensuring that each node's features fully reflect its interaction with its neighbors. By integrating the features of all nodes to form a modified relationship embedding matrix based on the fusion transmission effect, the structure and accuracy of the spatial feature data are further improved. This design solves the problems of fuzzy spatial feature aggregation and inaccurate quantification of transmission effects, allowing the modified relationship embedding to more realistically reflect the spatial dependencies between UAV nodes, providing higher-quality spatial features for spatiotemporal feature fusion calculations, and helping to improve the accuracy of prediction results.

[0050] S4. By performing spatiotemporal feature fusion calculation on temporal embedding and modified relational embedding, the predicted traffic of several UAV nodes in the future target time period is obtained.

[0051] Step S4 in this embodiment is specifically as follows: Based on the prediction layer, the temporal embedding and the modified relational embedding are input into the fully connected layer. The spatiotemporal feature fusion calculation of the temporal embedding and the modified relational embedding is performed through the neuron weight parameters of the fully connected layer to realize the joint modeling of spatiotemporal features and obtain the predicted traffic of several UAV nodes in the future target time period.

[0052] The fully connected layer is constructed by calculating the prediction error based on the loss function and adjusting the neuron weights through backpropagation using an optimizer until the loss value drops to a preset threshold. Specifically: A historical traffic sample set of the UAV power distribution communication network under the post-disaster emergency communication scenario is obtained. The historical time-series embedding and the historical modified relation embedding are calculated based on the historical traffic sample set according to the above calculation method. The historical time-series embedding and the historical modified relation embedding are used as training inputs and passed into the preset initial fully connected layer. The "initial fully connected layer" refers to the fully connected layer structure that is pre-built in the model before training the fully connected layer in this embodiment. The neurons in each layer are connected to all neurons in the previous layer, and the weights and biases are in a state of random initialization.

[0053] The initial fully connected layer is controlled to perform spatiotemporal feature fusion of historical time-series embedding and historical modified relation embedding through neuron weights to obtain the preliminary predicted traffic value of each UAV node for the corresponding historical time period. Based on the actual flow values ​​in the historical flow sample set and the calculated preliminary predicted flow values, the loss value is obtained according to the loss function. Based on the Adam optimizer, the first and second moment estimates of the gradient are calculated using the loss value as input. After setting the learning rate, backpropagation is performed to adjust the neuron weights of the initial fully connected layer. The error is continuously optimized based on the adjusted weights until the loss value decreases to a preset threshold and tends to stabilize. At this point, training is stopped to avoid overfitting and performance degradation, resulting in the fully connected layer. The first and second moment estimates are used to collaboratively guide the update direction and step size of the neuron weights in the initial fully connected layer.

[0054] The loss function is: in, It is the number of historical samples. These are the actual flow values ​​in the sample. This is the preliminary predicted flow value for the corresponding sample.

[0055] This embodiment S4 clarifies the core implementation path of spatiotemporal feature fusion calculation. By fusing temporal embeddings and corrected relational embeddings through the neuron weight parameters of the fully connected layer, it achieves deep interaction and complementarity between temporal and spatial features, solving the problem of spatiotemporal feature fragmentation in traditional models. Simultaneously, it clarifies the logic of establishing the fully connected layer based on the loss function to calculate prediction errors and adjusting weights through backpropagation using an optimizer, ensuring that the parameter optimization direction of the fully connected layer is highly consistent with the traffic prediction target. This design allows the model to fully utilize the temporal traffic patterns in the temporal domain and the node propagation effect in the spatial domain, achieving accurate prediction of UAV node traffic after disasters. It effectively avoids prediction bias caused by single-feature modeling, reduces the risk of secondary interruptions caused by node overload, improves the reliability and stability of post-disaster emergency communication networks, and adapts to the actual needs of post-disaster emergency communication. Furthermore, using historical time-series embeddings and historical corrected relational embeddings generated from historical traffic sample sets as training inputs ensures the consistency between training data and real-world application scenarios, improving the model's generalization ability. An initial fully connected layer generates preliminary predicted traffic values, which are then combined with real traffic values ​​to calculate the loss. An optimizer is then used for backpropagation to adjust the weights until the loss reaches a preset threshold, forming a closed-loop parameter optimization mechanism. This design effectively avoids the problem of poor feature fusion results caused by the randomization of fully connected layer parameters. Step-by-step optimization ensures that neuron weights accurately adapt to the spatiotemporal feature fusion requirements, while preventing overfitting, giving the fully connected layer stable and reliable spatiotemporal feature fusion computation capabilities. Ultimately, the model can output accurate prediction results when facing traffic changes at different stages after a disaster, improving the practicality and feasibility of the solution. Furthermore, this embodiment clarifies the core roles of first-order moment estimation and second-order moment estimation. These two methods work together to guide the weight updates of the initial fully connected layer, resolving the issues of ambiguous weight adjustment direction and unreasonable step size in traditional optimizers. First-order moment estimation effectively controls the direction of weight updates, ensuring that parameter optimization always progresses towards reducing the loss value and avoiding deviation from the optimal solution. Second-order moment estimation dynamically adjusts the update step size, adaptively matching the optimization rhythm based on parameter gradient changes, avoiding oscillations caused by excessively large step sizes or slow convergence caused by excessively small step sizes. The combined effect significantly improves the efficiency and accuracy of weight updates, shortens the model training cycle, allows the fully connected layer to reach a stable optimal state more quickly, and enhances the model's adaptability to complex changes in post-disaster traffic, ensuring the stability and accuracy of prediction results and providing efficient model training support for post-disaster emergency communication traffic prediction.

[0056] For an explanation of the embodiments of the present invention, please refer to [link / reference]. Figure 2 , Figure 2 This is a flowchart of the communication traffic prediction process provided by the present invention, which illustrates the process of obtaining the UAV node traffic dataset, performing output processing, and finally outputting the predicted future traffic values ​​of the UAV nodes.

[0057] Overall, this embodiment has the following beneficial effects: This invention utilizes a gated temporal modeling network to precisely filter key temporal patterns in UAV node traffic, effectively capturing the dynamic characteristics of traffic changes over time and transforming them into temporal embeddings. This overcomes the shortcomings of existing methods in fusing temporal dynamic patterns and improves the accuracy of traffic prediction. Based on the temporal embeddings, a modified relationship embedding for traffic data is generated. On one hand, by combining the key scene feature of quasi-static topology formed by UAVs hovering after a disaster, the node topological relationships are encoded into an adjacency matrix, transforming abstract spatial connections into structured information recognizable by the model, providing a carrier for capturing spatial transmission effects. On the other hand, this adjacency matrix and the temporal embeddings, whose temporal dynamic patterns have been captured through temporal modeling, are input into a graph convolutional network. Through inter-layer neighbor feature aggregation, spatial transmission effects and temporal dynamic features interact and quantify and fuse in the model. The resulting modified relationship embedding includes both the traffic transmission characteristics brought by spatial topology and the dynamic patterns associated with the time dimension, providing spatial feature support for subsequent spatiotemporal feature fusion calculations. Finally, by performing spatiotemporal feature fusion calculations on the temporal embedding of the dynamic laws of traffic carrying UAV nodes and the modified relationship embedding containing the spatial transmission effect based on post-disaster quasi-static topology, a deep fusion of temporal and spatial features is achieved, which can simultaneously capture the spatiotemporal correlation characteristics of traffic, thereby improving the accuracy of traffic prediction. In summary, this invention addresses the characteristics of post-disaster UAVs deployed as hovering relays with relatively fixed locations. The resulting power distribution communication network topology can be considered a quasi-static structure, where each UAV corresponds to an independent node in the graph model. This invention further expands the capability boundary of the prediction model by incorporating network topology features into the modeling scope, thereby accurately capturing the interaction mechanisms between nodes. In this communication network, each UAV node possesses independent traffic time-series data. The prediction process requires analyzing both the dynamic changes in long-term traffic and the data correlation characteristics between topologically connected nodes. Therefore, this invention uses GRU to mine the temporal evolution relationship of traffic, combined with LightGCN to characterize the spatial topological correlation of nodes, and achieves accurate prediction of future traffic of key relay nodes through spatiotemporal joint modeling, ultimately effectively avoiding the problem of secondary communication interruptions caused by node overload.

[0058] Example 2: Please see Figure 3 The embodiments of this application provide a device for predicting emergency communication traffic after a disaster in a UAV power distribution communication network, including a data acquisition module 10, a gating module 20, a convolution module 30 and a prediction module 40; Among them, the acquisition module 10 is used to acquire traffic data of several UAV nodes in the post-disaster emergency communication scenario of the UAV power distribution communication network; The gating module 20 is used to input traffic data into a gated temporal modeling network, capture time patterns based on the gating structure, and obtain the temporal embedding of several UAV nodes. Convolutional module 30 is used to generate corrected relation embeddings for traffic data based on temporal embedding. Specifically, it combines the quasi-static topological characteristics formed by the hovering of UAVs after the disaster, encodes the topological relationship of several UAV nodes into multi-hot binary vectors and constructs an adjacency matrix, embeds the adjacency matrix and temporal embeddings into the input graph convolutional network, and quantifies the traffic transmission effect between nodes through inter-layer neighbor feature aggregation to generate corrected relation embeddings for traffic data. The prediction module 40 is used to perform spatiotemporal feature fusion calculation on temporal embedding and modified relational embedding to obtain the predicted traffic of several UAV nodes in the future target time period.

[0059] It should be noted that the technical concept of this second embodiment is completely consistent with that of the first embodiment. The two maintain a high degree of synergy at the technical logic level. The specific technical details can be referred to the relevant description of the first embodiment, which will not be repeated here.

[0060] Example 3: This application provides a computer-readable storage medium, which includes a stored computer program, wherein the computer program controls the device where the computer-readable storage medium is located to execute the method for predicting post-disaster emergency communication traffic in a UAV power distribution communication network when it is running. The method for predicting emergency communication traffic in a UAV power distribution communication network after a disaster, if implemented as a software functional unit and used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0061] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer programs or instructions. When the computer program or instructions are loaded and executed on a computer, the processes or functions described in the embodiments of this application are performed entirely or partially. The computer can be a general-purpose computer, a special-purpose computer, a computer network, a network device, a user equipment, or other programmable device. The computer program or instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another. For example, the computer program or instructions can be transferred from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium, such as a floppy disk, hard disk, or magnetic tape; it can also be an optical medium, such as a digital video optical disc; or it can be a semiconductor medium, such as a solid-state drive. The computer-readable storage medium may be a volatile or non-volatile storage medium, or may include both types of storage media.

[0062] The above are preferred embodiments of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for predicting post-disaster emergency communication traffic of a UAV power distribution communication network, characterized in that, The application relates to a method for predicting traffic of unmanned aerial vehicle nodes in a future target period. In a post-disaster emergency communication scenario of a UAV power distribution communication network, traffic data of a plurality of UAV nodes are acquired; The traffic data are input into a time sequence modeling network with a gate, time rules are captured based on the gate structure, and time sequence embeddings of the plurality of UAV nodes are obtained; A corrected relationship embedding of the traffic data is generated based on the time sequence embedding, specifically, the topological relationship of the plurality of UAV nodes is encoded into a multi-hot binary vector and an adjacency matrix is constructed based on the quasi-static topological characteristics formed by the UAV hovering after the disaster, the adjacency matrix and the time sequence embedding are input into a graph convolution network, the traffic conduction effect between nodes is quantified through inter-layer neighbor feature aggregation, and the corrected relationship embedding of the traffic data is generated; The time sequence embedding and the corrected relationship embedding are subjected to spatiotemporal feature fusion calculation, and the predicted traffic of the plurality of UAV nodes in the future target period is obtained. 2.The method of claim 1, wherein, The traffic data are input into a time sequence modeling network with a gate, time rules are captured based on the gate structure, and time sequence embeddings of the plurality of UAV nodes are obtained, specifically: After the traffic data are arranged into a matrix form, the traffic data are input into the time sequence modeling network with a gate; The time dimension rules of the traffic data are filtered based on the reset gate and the update gate built in the time sequence modeling network; For each UAV node, the last hidden state output by the time sequence modeling network after processing the corresponding traffic data is taken as a sequence embedding; wherein the sequence embedding contains the time dimension rules; The sequence embeddings of all UAV nodes are integrated into a matrix to form the time sequence embeddings of the plurality of UAV nodes. 3.The method of claim 2, wherein, The reset gate is used to filter target historical information to be retained, and the update gate is used to control the memory and forgetting proportion of historical data. 4.The method of claim 1, wherein, The topological relationship of the plurality of UAV nodes is encoded into a multi-hot binary vector and an adjacency matrix is constructed based on the quasi-static topological characteristics formed by the UAV hovering after the disaster, the adjacency matrix and the time sequence embedding are input into a graph convolution network, the traffic conduction effect between nodes is quantified through inter-layer neighbor feature aggregation, and the corrected relationship embedding of the traffic data is generated, specifically: The topological relationship between the plurality of UAV nodes is encoded into a multi-hot binary vector based on the quasi-static topological characteristics formed by the UAV hovering after the disaster, and an adjacency matrix representing whether there is a connection between nodes is constructed based on the multi-hot binary vector; The adjacency matrix and the time sequence embedding are input into the graph convolution network together, and the features of the nodes and the neighbor nodes are aggregated and calculated through the network layers of the graph convolution network to generate the corrected relationship embedding of the traffic data.

5. The method of claim 4, wherein the method further comprises: The adjacency matrix and the time sequence embedding are input into the graph convolution network together, and the features of the nodes and the neighbor nodes are aggregated and calculated through the network layers of the graph convolution network to generate the corrected relationship embedding of the traffic data, specifically: The adjacency matrix and the time sequence embedding are input into the graph convolution network together, and the neighbor nodes of each node are located according to the adjacency matrix between the network layers of the graph convolution network to obtain a neighbor node set. The time sequence embedding of the current node in the graph convolution network and the time sequence embedding of the corresponding node in the neighbor node set are aggregated to calculate, synchronize the traffic conduction effect between nodes, and integrate the traffic conduction effect into the feature of the current node to obtain the feature of each node fused with the conduction effect; The features of all nodes fused with the conduction effect are integrated into a matrix to obtain the modified relationship embedding of the traffic data.

6. The unmanned aerial vehicle power distribution communication network post-disaster emergency communication traffic prediction method of claim 1, wherein, The time sequence embedding and the modified relationship embedding are input into a full connection layer, and the time sequence embedding and the modified relationship embedding are calculated by the neuron weight parameters of the full connection layer to realize joint modeling of the space-time features and obtain the predicted traffic of the plurality of UAV nodes in the future target period. The full connection layer is established based on a loss function to calculate a prediction error, and an optimizer is used to adjust the neuron weights by back propagation until the loss value decreases to a preset threshold. The full connection layer is established based on a loss function to calculate a prediction error, and an optimizer is used to adjust the neuron weights by back propagation until the loss value decreases to a preset threshold.

7. The unmanned aerial vehicle power distribution communication network post-disaster emergency communication traffic prediction method of claim 6, wherein, A historical traffic sample set in a UAV power distribution communication network post-disaster emergency communication scenario is obtained, and historical time sequence embedding and historical modified relationship embedding generated by the historical traffic sample set are used as training input and transmitted to a preset initial full connection layer. The initial full connection layer is controlled to perform space-time feature fusion on the historical time sequence embedding and the historical modified relationship embedding by neuron weights to obtain a preliminary predicted traffic value of each UAV node in a corresponding historical period. Based on the real traffic value in the historical traffic sample set and the preliminary predicted traffic value, a loss value is calculated according to a loss function. According to the optimizer, the first and second moment estimations of the gradient are calculated based on the loss value as input, back propagation is performed after setting a learning rate, the neuron weights of the initial full connection layer are adjusted, and the error is continuously optimized based on the adjusted weights until the loss value decreases to a preset threshold to obtain the full connection layer. The first and second moment estimations are used to cooperatively guide the update direction and step length of the initial full connection layer in adjusting the neuron weights. 8.The method of claim 7, wherein, In a UAV power distribution communication network post-disaster emergency communication scenario, traffic data of a plurality of UAV nodes is obtained, specifically: 9.The method of claim 1, wherein, In a UAV power distribution communication network post-disaster emergency communication scenario, the original traffic data of the plurality of UAV nodes is obtained. Target outliers in the original traffic data are removed, the original traffic data after removing the target outliers is grouped according to the UAV node ID to generate a traffic table of each UAV node. The traffic table of each UAV node is subjected to a mean value operation, and the traffic values of missing time points are filled to obtain the traffic data of the plurality of UAV nodes. It includes a collection module, a gating module, a convolution module, and a prediction module.

10. A UAV power distribution communication network post-disaster emergency communication traffic prediction device, characterized in that, ​ The collection module is configured to acquire traffic data of a plurality of unmanned aerial vehicle (UAV) nodes in a UAV power distribution communication network post-disaster emergency communication scenario. The gating module is configured to input the traffic data into a time series modeling network with gating, capture time rules based on a gating structure, and obtain time series embedding of the plurality of UAV nodes. The convolution module is configured to generate a corrected relationship embedding of the traffic data based on the time series embedding, specifically, encode a topological relationship of the plurality of UAV nodes into a multi-hot binary vector and construct an adjacency matrix based on a quasi-static topological characteristic formed by UAV hovering after a disaster, input the adjacency matrix and the time series embedding into a graph convolution network, quantify a traffic conduction effect between nodes by inter-layer neighbor feature aggregation, and generate the corrected relationship embedding of the traffic data. The prediction module is configured to obtain predicted traffic of the plurality of UAV nodes in a future target period by performing spatio-temporal feature fusion calculation on the time series embedding and the corrected relationship embedding.

11. A storage medium, characterized by The storage medium has a computer program stored thereon, the computer program is invoked and executed by a computer, and a UAV power distribution communication network post-disaster emergency communication traffic prediction method according to any one of claims 1 to 9 is implemented.

12. A computer program product comprising computer programs or instructions, characterized in that, The computer program or instructions are executed by a communication device to implement a UAV power distribution communication network post-disaster emergency communication traffic prediction method according to any one of claims 1 to 9.

Images