A method for optimizing routing of a UAV network based on space-time trajectory prediction

Abstract

This invention pertains to wireless communication network routing technology and relates to a method for optimizing UAV network routing based on spatiotemporal trajectory prediction. The method includes: collecting historical motion state sequences of UAV nodes and their neighboring nodes, and the current network state; preprocessing the historical motion state sequences of the UAV nodes and their neighboring nodes to obtain feature tensors; inputting the feature tensors into a trajectory prediction model to predict the positions of the UAV nodes and their neighboring nodes at future time steps; evaluating the predicted connection duration and routing hole risk indicators of neighboring nodes based on the predicted positions, network state, and destination node positions; and filtering neighboring nodes based on network state information, predicted connection duration, and routing hole risk indicators. This invention predicts the future positions of UAV nodes and their neighboring nodes using historical motion state information and evaluates the predicted connection duration and routing hole risk indicators based on the predicted positions, thereby avoiding potential routing holes.

Description

Technical Field

[0001] This invention pertains to wireless communication network routing technology and relates to a method for optimizing UAV network routing based on spatiotemporal trajectory prediction. Background Technology

[0002] With the widespread application of drone technology in military reconnaissance, emergency rescue, and smart cities, self-organizing networks composed of multiple drones are gradually becoming an important communication method for achieving collaborative perception and information sharing. However, due to the characteristics of drone nodes, such as high-speed movement and strong three-dimensional spatial maneuverability, their network topology exhibits highly dynamic changes over time, posing significant challenges to stable and reliable data transmission.

[0003] In complex mission environments, the frequent movement of UAV nodes can easily lead to sparse node distribution in local areas, thus creating so-called "routing holes" in the network. When data packets are forwarded to this area, the lack of reachable next-hop nodes will force the communication process to be interrupted or enter an inefficient detour, severely affecting the network's transmission performance.

[0004] To address the routing hole problem, existing drone networks often employ a geolocation-based greedy forwarding strategy, supplemented by detour rules to attempt to avoid hole areas. However, such methods typically rely on the current location information of nodes for decision-making, triggering the corresponding hole handling mechanism only when a data packet reaches the edge of the hole and cannot continue forwarding—a typical reactive approach. This lagging decision-making pattern can easily lead to data packets being repeatedly forwarded at the hole edge, increasing end-to-end latency and even causing packet loss due to excessively long paths.

[0005] In addition, some existing routing methods introduce intelligent decision-making mechanisms to optimize forwarding strategies, but their decision-making process is mostly based on the current network state and fails to fully consider the characteristics of the evolution of UAV network topology over time. They also lack the ability to effectively predict future link changes and hole formation trends, and therefore still find it difficult to obtain stable and reliable routing performance in highly dynamic scenarios. Summary of the Invention

[0006] To address the aforementioned problems in the prior art, this invention employs a UAV network routing optimization method based on spatiotemporal trajectory prediction, comprising:

[0007] S1: Collect the historical motion state sequence of the current UAV node and its one-hop neighbor nodes in the UAV network and the network state information at the current moment. Preprocess the historical motion state sequence of the current UAV node and its one-hop neighbor nodes to obtain the historical motion state sequence feature tensor.

[0008] S2: Obtain the trained trajectory prediction model, input the historical motion state sequence feature tensor into the trained trajectory prediction model, and predict the position information of the current UAV node and its one-hop neighbor nodes in the future at least one time step.

[0009] S3: Obtain the destination node and its location information of the current drone node. Based on the predicted location information, the network status information at the current moment, and the location information of the destination node, evaluate the predicted connection duration and routing hole risk index of each one-hop neighbor node of the current drone node.

[0010] S4: Based on the current network status information, predicted connection duration, and routing hole risk indicators, filter the one-hop neighbor nodes of the current drone node, and use the filtered one-hop neighbor nodes as the next-hop routing nodes of the current drone node.

[0011] Beneficial effects:

[0012] This invention analyzes the historical motion state information of UAV nodes and their neighboring nodes to predict their future location information. Based on the predicted location information, it evaluates the predicted connection duration and routing hole risk indicators, thereby avoiding potential routing holes in advance during the routing decision-making process and achieving more stable and efficient data transmission. Attached Figure Description

[0013] Figure 1 A flowchart of a UAV network routing optimization method based on spatiotemporal trajectory prediction provided in an embodiment of the present invention;

[0014] Figure 2 A flowchart for spatiotemporal trajectory prediction provided in an embodiment of the present invention;

[0015] Figure 3 This is a schematic diagram of the construction of the routing void potential energy field provided in an embodiment of the present invention;

[0016] Figure 4 A flowchart of next-hop routing decision based on prediction information is provided for embodiments of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] like Figure 1As shown, this embodiment of the invention employs a UAV network routing optimization method based on spatiotemporal trajectory prediction, including:

[0019] S1: Collect the historical motion state sequence of the current UAV node and its one-hop neighbor nodes in the UAV network and the network state information at the current moment. Preprocess the historical motion state sequence of the current UAV node and its one-hop neighbor nodes to obtain the historical motion state sequence feature tensor.

[0020] Current drone node Its real-time motion status is collected by the airborne positioning module and IMU. Thus, the historical motion state sequence is obtained. The historical motion state sequence includes the motion states from the past N times to the current time t; where, For three-dimensional position coordinates, [ Let t be a three-dimensional velocity vector, t be the current time, and N be the length of the historical motion state sequence.

[0021] Current drone node By periodically sending and receiving "Hello" beacon messages, the system interacts with neighboring nodes within its one-hop communication range to obtain information about its one-hop neighbors. The historical motion state sequence and the network state information at the current moment; the network state information includes at least: energy remaining rate. and cache usage .

[0022] Preprocessing the historical motion state sequences of the current UAV node and its one-hop neighbors includes: timestamp alignment and Z-score normalization of the collected historical motion state sequences of the current UAV node and its one-hop neighbors, followed by concatenation according to a unified time window to obtain the historical motion state sequence feature tensor. Specifically, to adapt to the dynamic topology of drones, a threshold for the number of neighboring nodes is set. : When the actual number of one-hop neighbor nodes When zero-padding is used and the mask is built synchronously, when Node truncation is performed, and the historical motion state sequences of the current UAV node and its one-hop neighbor nodes are then concatenated in batches to form a fixed-dimensional feature tensor. ,in, This represents the batch size of samples processed in parallel during a single forward computation of the model. In the offline training phase of the model, it represents the number of samples processed simultaneously, and in the current phase (i.e., the online real-time prediction phase of the drone), it takes the value of 1. For the length of a historical time step, This represents the state feature dimension of a single node at a single time step.

[0023] S2: Obtain the trained trajectory prediction model, input the historical motion state sequence feature tensor into the trained trajectory prediction model, and predict the position information of the current UAV node and its one-hop neighbor nodes in the future at least one time step.

[0024] like Figure 2 As shown, the trajectory prediction model includes: a temporal feature extraction network, a spatial feature extraction network, and a prediction network; the training process of the trajectory prediction model includes:

[0025] S21: Obtain the UAV network state dataset, which includes the historical motion state sequences of UAV nodes and their one-hop neighbor nodes; preprocess the UAV network state dataset to obtain the historical motion state sequence feature tensor.

[0026] S22: Input the historical motion state sequence feature tensor into the temporal feature extraction network to obtain time-related features;

[0027] In this embodiment, temporal feature extraction employs a Transformer Encoder network that includes multi-head self-attention and feedforward fully connected structures.

[0028] The self-attention mechanism is as follows:

[0029]

[0030] in, These are query, key, and value matrices, obtained through a linear mapping of historical states. This step captures the inertial and maneuvering patterns of the UAV flight by calculating the weighted contribution of states at different historical moments to future moments, and outputs time-related features. .

[0031] S23: Construct a neighbor node topology graph based on the feature tensor of the historical motion state sequence, and input the neighbor node topology graph into a pre-trained spatial feature extraction network to obtain spatially related features;

[0032] In this embodiment, the spatial feature extraction network adopts a graph attention network with communication distance constraint edge weights, and performs weighted aggregation of local topology with historical state as node features and communication distance constraint as edge weights.

[0033] Neighbor node topology graph Where V is the set of nodes, including the current UAV node and all its one-hop neighbor nodes; the feature of each node i in the node set V is the historical motion state sequence feature vector corresponding to node i in the historical motion state sequence feature tensor. E is the edge set, representing the pairs of nodes in the node set V that satisfy the communication radius constraint; A is the weight set, including the communication distance constraint weights between nodes in the node set V. :

[0034] in, For nodes With nodes The Euclidean distance between the features; The effective communication radius is a parameter that is pre-calibrated by the hardware attributes of the wireless communication module on the UAV, such as transmit power and receive sensitivity. To prevent extremely small constants with a denominator of zero.

[0035] When aggregating neighbor features, the spatial feature extraction network introduces communication distance constraint weights. The features of neighboring nodes are weighted and aggregated, with weights for nodes exceeding the communication distance being forcibly set to zero. The output is a spatial correlation feature reflecting the influence of local topology. .

[0036] S24: Fusing time-related features and space-related features to obtain spatiotemporal fusion features. By inputting the spatiotemporal fusion features into the prediction network, the predicted location information of the current UAV node i and its one-hop neighbor nodes j in the next at least one time step is obtained. ;in, K is the index of the time step, and K is the total number of prediction time steps.

[0037] In this embodiment, the prediction network uses a multilayer perceptron regression network to map spatiotemporal fusion features into a sequence of future location coordinates.

[0038] Integrate time-related features and space-related features: .

[0039] S25. Calculate the loss function value based on the predicted location information of the UAV node and its one-hop neighbor node in the future at least one time step, update the parameters of the trajectory prediction model based on the loss function value, and obtain the trained trajectory prediction model when the loss function value is minimized.

[0040] S3: Obtain the destination node and its location information of the current drone node. Based on the predicted location information, the network status information at the current moment, and the location information of the destination node, evaluate the predicted connection duration and routing hole risk index of each one-hop neighbor node of the current drone node.

[0041] The location information of the target node is obtained by parsing the header information of the data packet received by the current UAV node or by querying the network location service. Considering that the prediction time window is extremely short, the location of the target node is regarded as a fixed constant coordinate within a single prediction period in this embodiment.

[0042] The evaluation of the predicted connection duration of each one-hop neighbor node of the current drone node includes:

[0043] S301. Based on the predicted location information, extract the current UAV node. and its one-hop neighbor nodes At various moments in the future Predicted coordinates , ;in, Here, K is the index of the time step, K is the total number of prediction time steps, and t is the current time.

[0044] Predicted coordinates , .

[0045] S302. Calculate the current UAV node based on the predicted coordinates. and its one-hop neighbor nodes At various moments in the future Predicted distance ;

[0046] Predicted distance .

[0047] S303, For each one-hop neighbor node The search makes Minimum time step for initial establishment The time length corresponding to the found time step is taken as the current drone node. with one-hop neighbor nodes Predicted connection duration If not found If the time step k is set, then the current drone node will be... with one-hop neighbor nodes Predicted connection duration Set to the preset maximum prediction duration; where, The preset effective communication radius, and the maximum prediction duration is the number of time steps predicted by the trajectory prediction model. The product of the physical time interval of a single time step.

[0048] like Figure 3 As shown, the metrics for assessing the routing hole risk of each one-hop neighbor node of the current drone node include:

[0049] S311. Based on the predicted location information, extract the current UAV node. and its one-hop neighbor nodes At various moments in the future Predicted coordinates , ;

[0050] S312 uses the current drone node as the center and a preset effective communication radius. Construct a communication sphere with a radius, and divide the communication sphere into multiple three-dimensional sectors. Where m is the index of the 3D sector;

[0051] Considering the three-dimensional distribution characteristics of UAV networks, this invention abandons the traditional planar sector division and adopts a three-dimensional mesh division method in spherical coordinates. The current UAV location is taken as the center of the sphere, and the effective communication radius is... Construct a communication sphere with a radius, define the azimuth angle (0° to 360° in the horizontal plane) and the elevation angle (-90° to +90° in the vertical plane), and divide the sphere space into... Three-dimensional sector.

[0052] S313, Based on the location information of the destination node and the current drone node and its one-hop neighbor nodes At various moments in the future Predicted coordinate statistics for each 3D sector Expected number of effective neighbor nodes within and the slope of change ;

[0053] Statistics of each three-dimensional sector Expected number of effective neighbor nodes within include:

[0054] Step 1: Based on the location information of the destination node and the current drone node and its one-hop neighbor nodes At various moments in the future Calculate the predicted coordinates of the current drone node and its one-hop neighbor nodes At various moments in the future Distances from the destination node ;

[0055] Step 2, based on the current drone node and its one-hop neighbor nodes At various moments in the future Calculate the predicted coordinates of one-hop neighbor nodes Relative to the current drone node At various moments in the future Angle ;in, Each is a one-hop neighbor node Relative to the current drone node The relative azimuth and relative elevation angles at various future moments;

[0056] Relative azimuth, relative elevation:

[0057]

[0058] in, These are the current drone nodes. and its one-hop neighbor nodes Predicted coordinates For the current drone node with one-hop neighbor nodes The distance.

[0059] A valid forwarding candidate node is defined as: at the prediction time... The neighbor node With the target node The distance is less than the current node The distance to the destination node provides the forward routing progress.

[0060] Step 3, based on distance and angle Calculate each 3D sector Expected number of effective neighbor nodes within :

[0061] in, Determine one-hop neighbor nodes Is it in a three-dimensional sector? If inside, then ,otherwise, ; Determine one-hop neighbor nodes Is it closer to the destination node? If so, then ;otherwise .

[0062] S314, Based on expected value Calculate each 3D sector Basic risk value Based on the basic risk value and the slope of change Calculate each 3D sector Routing hole risk indicator ;

[0063] Establish a composite risk assessment model based on "static density" and "dynamic trend". Function variant, constructing the expected number of effective neighbors Compared with the basic risk value Negative exponential mapping relationship:

[0064]

[0065] in, The sparsity threshold is set to 2 in this embodiment, meaning that fewer than 2 nodes in a sector are considered high risk. This is the adjustment coefficient (set to 0.5 in this embodiment). The formula indicates that when predicting the number of neighbors... Much smaller than the sparsity threshold hour, Approaching 1 (high risk); conversely, approaching 0 (low risk).

[0066] Calculate each 3D sector Routing hole risk indicator include:

[0067] in, This is a penalty coefficient based on the dynamic trend of topological evolution. When... When effective nodes are lost, the risk of "routing holes" increases dramatically due to local topology deterioration, therefore, a setting is made... (This embodiment takes) Amplify the risk value to guide routing to proactively avoid the high-risk sector; conversely, when When nodes are stable or converging, network connectivity improves, hence the setting. The basic risk assessment remains unchanged.

[0068] S315. For each one-hop neighbor node Based on its predicted coordinates, its future time points are determined. The corresponding 3D sector, if it exists at all future moments If the sectors belong to the same category, then the routing hole risk index of the sector to which it belongs is used as its routing hole risk index; otherwise, its risk index is calculated for all future times. The time weight of each 3D sector is obtained by dividing the time steps of each sector into the total predicted time steps K, and then assigning these weights to all future time steps. The routing hole risk indicators of the corresponding three-dimensional sector are weighted and merged according to time weight to obtain its routing hole risk indicator.

[0069] S4: Based on the current network status information, predicted connection duration, and routing hole risk indicators, filter the one-hop neighbor nodes of the current drone node, and use the filtered one-hop neighbor nodes as the next-hop routing nodes of the current drone node.

[0070] like Figure 4 As shown, filtering the one-hop neighbor nodes of the current drone node includes:

[0071] S41: Construct a candidate set C based on all one-hop neighbor nodes of the current drone node, and remove nodes from the candidate set C whose predicted connection duration is lower than a preset stability threshold. Or the routing hole risk indicator is higher than the preset risk threshold. The nodes are used to obtain the preferred candidate set;

[0072] Set stability threshold ;in, The single-hop average latency is calculated by the drone node using the moving average of historical communication interactions collected from the underlying MAC protocol; a risk threshold is set. =0.8.

[0073] S42: Determine if the preferred candidate set is empty. If it is not empty, use the preferred candidate set as the final available candidate set; otherwise, it means that all neighbors are not up to standard, so the backup strategy is triggered, and candidate set C is used as the final available candidate set to ensure the connectivity of data packets.

[0074] S43: Obtain the trained intelligent decision-making model and construct the input state of the intelligent decision-making model; wherein, the input state includes at least: the current network state information (energy remaining rate, cache utilization rate) of each neighbor node in the available candidate set, the predicted connection duration, the route hole risk index, and the distance to the destination node. And the distance between the current drone node and the destination node. ;

[0075] In this embodiment, the intelligent decision-making model is a deep Q-network.

[0076] S44: Input the constructed input state into the trained intelligent decision-making model, output the evaluation value of each neighbor node in the available candidate set, and select the neighbor node with the highest evaluation value as the next hop routing node.

[0077] In this embodiment, during the offline training phase before the actual deployment and operation of the UAV network, the training process of the intelligent decision model includes: modeling the routing and forwarding process of the UAV network as a Markov decision process; and training the intelligent decision model based on the Markov decision process to obtain a trained intelligent decision model.

[0078] The state space S of the Markov decision process includes at least: the network state information of each neighbor node in the available candidate set of the drone node, the predicted connection duration, the routing hole risk index, the distance to the destination node, and the distance between the drone node and the destination node; the action space A of the Markov decision process includes the neighbor nodes in the available candidate set of the drone node; and the reward function of the Markov decision process. The design is as follows, guiding the iterative update of model parameters to maximize long-term communication benefits:

[0079]

[0080]

[0081] in, For drone nodes Select its neighboring nodes from the available candidate set. The reward function; For drone nodes Actions, i.e., drone nodes Neighboring nodes selected from its available candidate set; For the indicator function, if the drone node Select its neighboring nodes from the available candidate set. As the next-hop routing node, ,otherwise ; , , , , As weight, For drone nodes Distance to the destination node For neighboring nodes Distance to the destination node For neighboring nodes Energy surplus rate, For neighboring nodes Cache usage For neighboring nodes The routing hole risk indicator For drone nodes with neighboring nodes The predicted connection duration, where j is the drone node The index of the neighboring node in the available candidate set. For drone nodes The number of neighboring nodes in the available candidate set.

[0082] In this embodiment, as shown in Table 1, there are five weight parameters. The value setting is not a single static combination, but a dynamic allocation based on three dimensions: "spatial convergence orientation, node resource load, and local quality optimization".

[0083] Table 1 Dynamic allocation of weight parameters

[0084]

[0085] In summary, the weight allocation is stratified based on its core contribution to topological robustness, ensuring the scientific nature of the decision-making logic. Furthermore, the intelligent decision-making model supports... Under the normalization constraint, the weights are adjusted online adaptively based on the evolution characteristics of network topology entropy. The above weight values ​​are only one optimal embodiment.

[0086] Training an intelligent decision-making model based on a Markov decision process includes: obtaining the current state of the UAV network (i.e., executing steps S1, S2, S3, S41, and S42 to obtain the current input state of the intelligent decision-making model); inputting the current state into the intelligent decision-making model to obtain the current optimal action; executing the optimal action to obtain the current reward and the next state (i.e., after executing the optimal action, executing steps S1, S2, S3, S41, and S42 to obtain the next input state of the intelligent decision-making model); constructing experience samples based on the current state, the current optimal action, the current reward, and the next state; adding the experience samples to the experience replay pool; and collecting experience samples from the experience replay pool to iteratively train the intelligent decision-making model to obtain a trained intelligent decision-making model.

[0087] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for optimizing UAV network routing based on spatiotemporal trajectory prediction, characterized in that, include: S1: Collect the historical motion state sequence of the current UAV node and its one-hop neighbor nodes and the network state information at the current moment in the UAV network. Preprocess the historical motion state sequence of the current UAV node and its one-hop neighbor nodes to obtain the historical motion state sequence feature tensor. S2: Obtain the trained trajectory prediction model, input the historical motion state sequence feature tensor into the trained trajectory prediction model, and predict the position information of the current UAV node and its one-hop neighbor nodes in the future at least one time step. S3: Obtain the destination node and its location information of the current drone node. Based on the predicted location information, the network status information at the current moment, and the location information of the destination node, evaluate the predicted connection duration and routing hole risk index of each one-hop neighbor node of the current drone node. S4: Based on the current network status information, predicted connection duration, and routing hole risk indicators, filter the one-hop neighbor nodes of the current drone node, and use the filtered one-hop neighbor nodes as the next-hop routing nodes of the current drone node.

2. The UAV network routing optimization method based on spatiotemporal trajectory prediction according to claim 1, characterized in that, The historical motion state sequence includes the motion state from the past N moments to the current moment t; Motion status includes at least position information and velocity information; Network state information includes at least: energy remaining rate and buffer occupancy rate; where N is the length of the historical motion state sequence.

3. The UAV network routing optimization method based on spatiotemporal trajectory prediction according to claim 1, characterized in that, The trajectory prediction model includes: a temporal feature extraction network, a spatial feature extraction network, and a prediction network; the training process of the trajectory prediction model includes: S21: Obtain the UAV network state dataset, which includes the historical motion state sequences of UAV nodes and their one-hop neighbor nodes; preprocess the UAV network state dataset to obtain the historical motion state sequence feature tensor. S22: Input the historical motion state sequence feature tensor into the temporal feature extraction network to obtain time-related features; S23: Construct a neighbor node topology graph based on the feature tensor of the historical motion state sequence, and input the neighbor node topology graph into the spatial feature extraction network to obtain spatially related features; S24: Fuse time-related features and spatial-related features to obtain spatiotemporal fusion features; input the spatiotemporal fusion features into the prediction network to obtain the predicted location information of the UAV node and its one-hop neighbor node in the future at least one time step. S25. Calculate the loss function value based on the predicted location information of the UAV node and its one-hop neighbor nodes at least one time step in the future. Update the parameters of the trajectory prediction model based on the loss function value. When the loss function value is minimized, the trained trajectory prediction model is obtained.

4. The UAV network routing optimization method based on spatiotemporal trajectory prediction according to claim 3, characterized in that, Neighbor node topology graph Where V is the set of nodes, including the current UAV node and all its one-hop neighbor nodes; the feature of each node i in the node set V is the historical motion state sequence feature vector corresponding to node i in the historical motion state sequence feature tensor; A is the set of weights, including the communication distance constraint weights between nodes in the node set V. : in, For nodes With nodes The Euclidean distance between the features, The preset effective communication radius, To prevent extremely small constants with a denominator of zero.

5. The UAV network routing optimization method based on spatiotemporal trajectory prediction according to claim 1, characterized in that, The evaluation of the predicted connection duration of each one-hop neighbor node of the current drone node includes: Based on the predicted location information, extract the current drone node. and its one-hop neighbor nodes At various moments in the future Predicted coordinates , ;in, Here, K is the index of the time step, K is the total number of prediction time steps, and t is the current time. Calculate the current drone node based on predicted coordinates. and its one-hop neighbor nodes At various moments in the future Predicted distance ; For each one-hop neighbor node The search makes Minimum time step for initial establishment The time length corresponding to the found time step is taken as the current drone node. with one-hop neighbor nodes The predicted connection duration; if no connection is found that makes it possible to determine the connection duration. If the time step k is set, then the current drone node will be... with one-hop neighbor nodes The predicted connection duration is set to the preset maximum prediction duration; where, The preset effective communication radius, and the maximum prediction duration is the number of time steps predicted by the trajectory prediction model. The product of the physical time interval of a single time step.

6. The UAV network routing optimization method based on spatiotemporal trajectory prediction according to claim 1, characterized in that, Indicators for assessing the risk of routing holes between the current drone node and its one-hop neighbors include: Based on the predicted location information, extract the current drone node. and its one-hop neighbor nodes At various moments in the future Predicted coordinates , ;in, Here, K is the index of the time step, K is the total number of prediction time steps, and t is the current time. Centered on the current drone node, with a preset effective communication radius Construct a communication sphere with a radius, and divide the communication sphere into multiple three-dimensional sectors. Where m is the index of the 3D sector; Based on the location information of the destination node and the current drone node and its one-hop neighbor nodes At various moments in the future Predicted coordinate statistics for each 3D sector Expected number of effective neighbor nodes within and the slope of change ; Based on expected value Calculate each 3D sector Base risk value Based on the basic risk value and the slope of change Calculate each 3D sector Routing hole risk indicator ; For each one-hop neighbor node Based on its predicted coordinates, its future time points are determined. The corresponding 3D sector, if it exists at all future moments If the sectors belong to the same vertical sector, then the routing hole risk index of the vertical sector to which it belongs is used as its routing hole risk index; otherwise, its risk index is calculated for all future times. The time weight of each 3D sector is obtained by dividing the time steps of each sector into the total predicted time steps K, and then assigning these weights to all future time steps. The routing hole risk indicators of the corresponding three-dimensional sector are weighted and merged according to time weight to obtain its routing hole risk indicator.

7. The UAV network routing optimization method based on spatiotemporal trajectory prediction according to claim 6, characterized in that, Statistics of each three-dimensional sector Expected number of effective neighbor nodes within include: Based on the location information of the destination node and the current drone node and its one-hop neighbor nodes At various moments in the future Calculate the predicted coordinates of the current drone node and its one-hop neighbor nodes At various moments in the future Distances from the destination node ; Based on the current drone node and its one-hop neighbor nodes At various moments in the future Calculate the predicted coordinates of one-hop neighbor nodes Relative to the current drone node At various moments in the future Angle ;in, Each is a one-hop neighbor node Relative to the current drone node The relative azimuth and relative elevation angles at various future moments; Based on distance and angle Calculate each 3D sector Expected number of effective neighbor nodes within : in, Determine one-hop neighbor nodes Is it in a three-dimensional sector? If inside, then ,otherwise, ; Determine one-hop neighbor nodes Is it closer to the destination node? If so, then ;otherwise .

8. The UAV network routing optimization method based on spatiotemporal trajectory prediction according to claim 6, characterized in that, Calculate each 3D sector Routing hole risk indicator include: in, The penalty coefficient is based on the dynamic trend of topological evolution. ,but ;like ,but .

9. The UAV network routing optimization method based on spatiotemporal trajectory prediction according to claim 1, characterized in that, Filtering the one-hop neighbor nodes of the current drone node includes: S41: Construct a candidate set C based on all one-hop neighbor nodes of the current drone node, and remove nodes from the candidate set C whose predicted connection duration is lower than a preset stability threshold. Or the routing hole risk indicator is higher than the preset risk threshold. The nodes are used to obtain the preferred candidate set; S42: Determine whether the preferred candidate set is empty. If it is not empty, use the preferred candidate set as the final available candidate set; otherwise, use the candidate set C as the final available candidate set. S43: Obtain the trained intelligent decision-making model and construct the input state of the intelligent decision-making model; wherein, the input state includes at least: the current network state information of each neighbor node in the available candidate set, the predicted connection duration, the routing hole risk index, the distance to the destination node, and the current distance between the drone node and the destination node; S44: Input the constructed input state into the trained intelligent decision-making model, output the evaluation value of each neighbor node in the available candidate set, and select the neighbor node with the highest evaluation value as the next hop routing node.

10. The UAV network routing optimization method based on spatiotemporal trajectory prediction according to claim 9, characterized in that, The training process of the intelligent decision-making model includes: modeling the routing and forwarding process of the UAV network as a Markov decision process; training the intelligent decision-making model based on the Markov decision process to obtain the trained intelligent decision-making model; the state space of the Markov decision process includes at least: the network state information of each neighbor node in the available candidate set of the UAV node, the predicted connection duration, the routing hole risk index, the distance to the destination node, and the distance between the UAV node and the destination node; the network state information includes: energy remaining rate and cache utilization rate; the action space of the Markov decision process includes the neighbor nodes in the available candidate set of the UAV node; and the reward function of the Markov decision process. for: in, For drone nodes Select its neighboring nodes from the available candidate set. The reward function; For drone nodes Actions, i.e., drone nodes Neighboring nodes selected from its available candidate set; For the indicator function, if the drone node Select its neighboring nodes from the available candidate set. As the next-hop routing node, ,otherwise ; , , , , As weight, For drone nodes Distance to the destination node For neighboring nodes Distance to the destination node For neighboring nodes Energy surplus rate, For neighboring nodes Cache usage For neighboring nodes The routing hole risk indicator For drone nodes with neighboring nodes The predicted connection duration, where j is the drone node The index of the neighboring node in the available candidate set. For drone nodes The number of neighboring nodes in the available candidate set.

Images