Unmanned aerial vehicle base station deployment method and system based on GNN and differential evolution

Abstract

This disclosure provides a method and system for deploying unmanned aerial vehicle (UAV) base stations based on GNN and differential evolution. Applied to the fields of wireless communication network optimization and artificial intelligence, the method includes: acquiring a disaster area and constructing a heterogeneous graph model; based on the heterogeneous graph model, using a graph neural network to predict the quality of candidate UAV base station locations and generate a deployment priority distribution of candidate points; initializing a differential evolution algorithm population based on the probability distribution; iteratively updating the initial population through differential evolution optimization to obtain an approximate global optimal solution for the three-dimensional location of the UAV base stations; and after the differential evolution algorithm converges iteratively, extracting the individual with the lowest fitness from the final population to obtain the optimal three-dimensional deployment location of the UAV base station. In this way, the present invention can obtain high-quality three-dimensional location configurations of UAV base stations in complex deployment spaces, effectively improving deployment efficiency, coverage quality, and algorithm stability.

Description

Technical Field

[0001] This disclosure relates to the fields of wireless communication network optimization and artificial intelligence technology, and in particular to a method and system for deploying unmanned aerial vehicle (UAV) base stations based on GNN and differential evolution. Background Technology

[0002] In natural disasters such as earthquakes, floods, and forest fires, traditional ground communication base stations are easily damaged, causing localized or large-scale communication outages and making it difficult to transmit rescue command information. Unmanned Aerial Vehicle Base Stations (UAV-BS), due to their rapid deployment, mobility, and high controllability, are considered an important means of restoring emergency communications at disaster sites. However, the three-dimensional deployment of UAV-BS presents the following problems:

[0003] 1. The location of a drone base station includes two-dimensional horizontal position and continuous height parameters, constituting a high-dimensional non-convex optimization problem, which is difficult to find globally for the optimal solution using traditional methods;

[0004] 2. Uneven user distribution at different heights, building collapses, terrain obstruction, and temporary airspace restrictions significantly affect deployment quality;

[0005] 3. Existing optimization methods converge slowly or are prone to getting trapped in local optima. Although differential evolution has strong global search capabilities, it still needs heuristic guidance to improve search efficiency in highly complex environments.

[0006] 4. Lack of full utilization of spatial structure information. The distribution of users and the extent of damage in the disaster site have significant spatial correlations, but traditional optimization methods cannot model the dependencies between nodes.

[0007] Graph Neural Networks (GNNs) have advantages in processing spatially structured data. They can learn prior knowledge of "deployment quality" from the structural relationships between device distribution and candidate base station locations, thus providing a high-quality search starting point for differential evolution algorithms.

[0008] Therefore, it is necessary to construct a hybrid optimization strategy for UAV base station deployment that combines the spatial learning capabilities of GNNs and differential evolution algorithms with global optimization capabilities. Summary of the Invention

[0009] This disclosure provides a method and system for deploying drone base stations based on GNN and differential evolution. By combining graph neural networks and differential evolution algorithms, it enables rapid deployment of drone base stations and solves the technical problems of low deployment quality and poor efficiency of existing methods.

[0010] According to a first aspect of this disclosure, a method for deploying unmanned aerial vehicle (UAV) base stations based on GNN and differential evolution is provided, comprising the following steps:

[0011] Obtain disaster areas and construct heterogeneous graph models based on these areas;

[0012] Based on the heterogeneous graph model, a graph neural network is used to predict the quality of candidate UAV base station locations and generate a deployment priority distribution of candidate points.

[0013] The differential evolution algorithm population is initialized based on the deployment priority probability distribution of the candidate points;

[0014] By iteratively updating the initial population multiple times through differential evolution, an approximate global optimal solution for the three-dimensional position of the UAV base station is obtained.

[0015] After the differential evolution algorithm converges iteratively, the individual with the lowest fitness is extracted from the final population to obtain the optimal three-dimensional deployment location of the UAV base station.

[0016] Preferably, the process of obtaining the disaster area and constructing a heterogeneous graph model based on the disaster area is as follows:

[0017] Geographic information and user equipment information of the disaster area are collected, and the K-means clustering method is used to generate a set of UAV base station deployment points to obtain a disaster area model;

[0018] Based on the disaster area model, a set of heterogeneous graph nodes and a set of heterogeneous graph edges are constructed.

[0019] Based on the set of nodes and edges of the heterogeneous graph, a node feature matrix, an edge feature matrix, and an adjacency structure are generated, and a heterogeneous graph model is constructed.

[0020] Preferably, the heterogeneous graph node set includes user equipment nodes and candidate UAV deployment nodes;

[0021] The user equipment node is generated based on terrestrial IoT devices and includes location coordinates, data demand, and user priority information;

[0022] The candidate drone deployment nodes include three-dimensional candidate locations and information indicating whether the line of sight is obstructed.

[0023] Preferably, the heterogeneous graph edge set includes user-user edges, user-candidate deployment location edges, and candidate-candidate edges;

[0024] The user-user edges are constructed based on the distance between two users and are used to express the user density and hotspot region structure, wherein the weight of the edge is the distance between two users;

[0025] The user-candidate deployment location edge is used to represent the coverage capability of IoT users when a certain node is used as the drone base station location. Specifically, it includes the estimated path loss, the horizontal distance between the IoT device and the drone base station, and whether there are obstacles on the distance route.

[0026] The candidate-candidate edge is used to represent the minimum safe distance constraint when multiple drones cooperate, specifically including the location information of the drone base station and the distance information of other drone base stations.

[0027] Preferably, the process of using a graph neural network to predict the quality of candidate UAV base station locations based on the heterogeneous graph model and generating a deployment priority distribution of candidate points is as follows:

[0028] The constructed heterogeneous graph model is input into the graph neural network, and neighborhood aggregation and feature fusion are performed on each candidate UAV deployment node through a multi-layer message passing mechanism to update the hidden representation of the node.

[0029] After multiple layers of propagation, the final embedded representation of the candidate UAV deployment node is calculated;

[0030] The final embedded representation is mapped to a scalar deployment quality score;

[0031] The quality scores of all candidate drone deployment nodes are normalized using the softmax function to generate a deployment priority probability distribution.

[0032] Preferably, the process of initializing the differential evolution algorithm population based on the deployment priority probability distribution of the candidate points is as follows:

[0033] A set of candidate UAV deployment three-dimensional locations is obtained, and individual vectors in the set of UAV base station three-dimensional locations are sampled multiple times based on the deployment priority probability distribution to obtain several candidate center point indices.

[0034] Gaussian perturbations are introduced near each center point to generate initial individuals for the differential evolution algorithm, thus obtaining the initialization results;

[0035] Boundary corrections are applied to the initialization results to obtain an initial population that satisfies both physical and operational constraints.

[0036] Preferably, the process of boundary correction of the initialization results to obtain an initial population that satisfies physical and operational constraints includes:

[0037] If individual x i y i If the deployment exceeds the boundary of the deployment area, a mirroring / pruning strategy is used for boundary correction, where x i y i These are the horizontal and vertical coordinates of the candidate drone deployment locations;

[0038] If the individual's h i If the flight altitude exceeds the permitted range, a truncation method will be used to limit the boundary to within the permitted range, where h i Deployment altitude for candidate drones;

[0039] If an individual's location falls into a no-fly zone or a severely obstructed area, the center point index needs to be resampled or a Gaussian perturbation needs to be introduced back into the center point.

[0040] Preferably, the process of iteratively updating the initial population multiple times through differential evolution to obtain an approximate global optimal solution for the three-dimensional location of the UAV base station is as follows:

[0041] The air-to-ground link propagation model was used to calculate the coverage rate and energy consumption index of each individual in the population within the disaster area, and the fitness function was constructed using a multi-objective weighted method.

[0042] For each target individual in the population, three different individuals are randomly selected from the current population to construct a mutation vector, thus obtaining the mutated individual;

[0043] A dimension-by-dimensional crossover strategy is used to combine the target individual with the variant individual to generate experimental individuals;

[0044] For each experimental individual, its fitness is calculated and compared with the fitness of the corresponding target individual. The better individual is selected to enter the next generation. The iteration stops when the number of iterations reaches the preset maximum number of iterations.

[0045] According to a second aspect of this disclosure, a UAV base station deployment system based on GNN and differential evolution is provided, comprising: a model building module, a quality prediction module, a population initialization module, an iterative optimization module, and an optimal deployment output module;

[0046] The model building module is used to acquire disaster areas and construct a heterogeneous graph model based on the disaster areas.

[0047] The quality prediction module is used to predict the quality of candidate UAV base station locations based on the heterogeneous graph model and using a graph neural network to generate a deployment priority distribution of candidate points.

[0048] The population initialization module is used to initialize the differential evolution algorithm population based on the deployment priority probability distribution;

[0049] The iterative optimization module is used to iteratively update the initial population multiple times through differential evolution iterative optimization to obtain an approximate global optimal solution for the three-dimensional position of the UAV base station;

[0050] The optimal deployment output module is used to extract the individual with the lowest fitness from the final population after the differential evolution algorithm has converged, thereby obtaining the optimal three-dimensional deployment location of the UAV base station.

[0051] Compared with the prior art, the present invention has the following technical effects:

[0052] (1) This invention constructs a heterogeneous graph structure of disaster areas and introduces graph neural networks to model user distribution, terrain structure and candidate deployment location relationships. It can learn deep features of deployment quality from spatial topology and achieve accurate prediction of high-quality deployment areas. Furthermore, this invention uses the high-quality deployment probability distribution output by GNN to initialize the population of differential evolution algorithm, so that DE can iterate in a better search subspace, significantly reduce invalid search areas, effectively improve global search ability and convergence speed, and overcome the problem of random initialization and premature convergence of traditional DE in high-dimensional continuous environment.

[0053] (2) Furthermore, this invention employs a multi-objective joint optimization strategy for the deployment of UAV base stations. By comprehensively considering path loss, coverage, airspace restrictions, and the intensity of user demand in disaster areas, it can ensure that UAV base stations meet communication service requirements while improving resource utilization and deployment stability. Compared to existing technologies that only consider coverage or path loss as a single objective, the deployment strategy of this invention is more robust and practical, and is particularly suitable for scenarios with extremely high timeliness requirements, such as disaster relief and emergency communication restoration.

[0054] (3) Through the complementary collaborative optimization of GNN and DE, this invention realizes a hybrid deployment framework of "structure perception + global optimization", which can obtain high-quality three-dimensional position configuration of UAV base stations in complex, dynamic and non-convex deployment space, effectively improving deployment efficiency, coverage quality and algorithm stability, and providing better technical support for the application of UAV base stations in disaster emergency response, low-altitude communication network construction and smart city scenarios.

[0055] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of this disclosure, nor is it intended to restrict the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0056] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the scope of this disclosure. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0057] Figure 1The diagram illustrates the variation of drone base station altitude in a drone base station deployment method based on GNN and differential evolution according to an embodiment of this disclosure.

[0058] Figure 2 A schematic flowchart of a UAV base station deployment method based on GNN and differential evolution according to an embodiment of the present disclosure is shown.

[0059] Figure 3 A schematic diagram illustrating the heterogeneous graph model construction process of a UAV base station deployment method based on GNN and differential evolution according to an embodiment of the present disclosure is shown.

[0060] Figure 4 A schematic diagram of the deployment quality prediction process of a UAV base station deployment method based on GNN and differential evolution according to an embodiment of the present disclosure is shown.

[0061] Figure 5 A schematic diagram of the population initialization process of a UAV base station deployment method based on GNN and differential evolution according to an embodiment of the present disclosure is shown.

[0062] Figure 6 A schematic diagram illustrating the iterative optimization process of a UAV base station deployment method based on GNN and differential evolution according to an embodiment of the present disclosure is shown.

[0063] Figure 7 A schematic diagram of a UAV base station deployment system based on GNN and differential evolution according to an embodiment of the present disclosure is shown. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0065] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0066] This embodiment considers a possessing One drone base station, Emergency disaster scenarios for individual IoT devices. In specific application scenarios... Several IoT devices are randomly distributed in the disaster scenario. A drone base station hovers in the air to provide communication services. Assuming both the drone base station and the IoT device are equipped with omnidirectional antennas capable of transmitting and receiving signals, the drone base station provides communication to the IoT device via an air-to-ground channel. A set of indexes representing time slots.

[0067] Due to the locational relationship between the drone base station and the IoT device, in disaster scenarios, the presence of obstacles such as houses and trees introduces non-line-of-sight components into the line-of-sight link for communication between the drone base station and the IoT device. Therefore, the communication channel between the drone base station and the IoT device includes both line-of-sight and non-line-of-sight link modes. This embodiment uses an air-to-ground channel probability model to calculate the average path loss when the drone base station communicates with the IoT device. Specifically:

[0068] in, Indicates drone base station With IoT devices The angle of elevation between them Indicates drone base station With IoT devices Between at the elevation angle The probability of establishing a line-of-sight link under certain circumstances can be specifically expressed as:

[0069]

[0070] in, and These are environmental parameters; therefore, the probability of a non-line-of-sight link connection can be expressed as: ; and These are the average path losses for line-of-sight links and non-line-of-sight links, respectively, and can be expressed as:

[0071] (3)

[0072] (4)

[0073] in, It is the carrier frequency of the air-to-ground channel. It's the speed of light. Indicates drone base station With IoT devices The distance between them and These are the average additional path loss for line-of-sight links and non-line-of-sight links, respectively.

[0074] According to the air-to-ground channel model, for UAV base stations At a given transmit power In such cases, the coverage range of a drone will vary depending on its altitude. The coverage range of a drone base station varies with altitude in four urban environments, as shown below. Figure 1 As shown.

[0075] For an IoT device If it is connected to a drone base station The horizontal distance between them is less than or equal to the coverage area of ​​the drone. This is defined as the Internet of Things being covered by drone base stations. Specifically, it can be expressed as:

[0076] (5)

[0077] in, Indicates Internet of Things (IoT) devices drone base station Covered, or not covered, r ij This refers to the horizontal distance between the drone and the base station.

[0078] Coverage is a key factor in evaluating the effectiveness of drone base station deployment schemes. In this embodiment, it is defined as the ratio of the number of IoT devices covered by the drone base station to the total number of IoT devices, which can be specifically expressed as:

[0079]

[0080] Furthermore, in the deployment of drone base stations, the energy consumption of drone base stations mainly consists of two parts: transmission energy consumption during auxiliary communication and hovering energy consumption. Drone base station Energy consumption can be expressed as:

[0081]

[0082] in, The energy consumption during UAV base station-assisted communication can be expressed as follows:

[0083] (8)

[0084] in, Indicates drone base station The transmission power, Indicates the time during which the drone performs assisted communication. This represents the hovering energy consumption of the drone base station during auxiliary communication, calculated based on the hovering power of the drone base station.

[0085] On the other hand, in order to ensure service quality, this embodiment defines the first The receiving power of each IoT device must exceed a threshold. Only then can drone base stations provide services to IoT devices. Therefore, drone base stations To IoT devices The minimum transmit power can be expressed as:

[0086] (9)

[0087] Among them, P ij,T This is the minimum transmission power.

[0088] In summary, in this embodiment, the transmission power of the entire UAV base station can be expressed as P. T :

[0089] (10)

[0090] Therefore, the goal of this embodiment is to achieve the maximum coverage with the minimum number of drone base stations deployed while ensuring communication quality. Under the premise of minimizing the drone base station transmission power to reduce energy consumption and meet the minimum communication threshold requirements of IoT devices, the three-dimensional deployment problem of drone base stations can be described as follows:

[0091] (11)

[0092] Wherein, constraint C1 represents the drone base station. With IoT devices Connection decisions between them; constraint C2 states that each IoT device can be served by at most one drone base station; constraint C3 states that the drone base station Maximum service One IoT device The maximum number of IoT devices; constraint C4 represents the number of drone base stations. The transmit power must be greater than the path loss and minimum threshold requirements; constraints C5, C6, and C7 indicate that the location constraints of the UAV base station must not exceed a specific area. i y i h i These are the three-dimensional location coordinates of the drone base station.

[0093] Therefore, in order to provide reliable communication services to IoT devices in disaster scenarios, this embodiment provides a drone base station deployment method based on GNN (Graph Neural Network) and differential evolution. By utilizing the mobility of drones to quickly move to designated deployment areas in disaster zones, reliable communication services are provided to IoT devices, effectively improving the speed of emergency rescue.

[0094] like Figure 2As shown, the method for deploying unmanned aerial vehicle (UAV) base stations based on GNN and differential evolution includes the following steps:

[0095] S101. Obtain the disaster area and construct a heterogeneous graph model based on the disaster area.

[0096] like Figure 3 As shown, this step aims to encode information such as environmental elements, user equipment, and candidate UAV deployment locations within the disaster area into a unified graph structure, enabling the graph neural network to perform high-dimensional feature inference using spatial topological relationships. The entire construction process includes four main stages: data acquisition, node set definition, edge set generation, and node and edge feature extraction.

[0097] (1) Data collection and spatial discretization of disaster areas: Before constructing heterogeneous maps, disaster areas are first modeled, and geographical information and user equipment information of disaster areas are collected. K-means clustering method is used to generate a set of UAV base station deployment points.

[0098] (2) Heterogeneous graph node set Construction

[0099] User equipment node Each terrestrial IoT device, as a node in the diagram, can be specifically represented as follows:

[0100] (12)

[0101] in, For terrestrial IoT devices, each node contains information such as location coordinates, data demand, and user priority.

[0102] Candidate drone deployment nodes :

[0103] (13)

[0104] in, For candidate drone deployment locations, each node contains information such as a 3D candidate location and whether the line of sight is obstructed.

[0105] (3) Heterogeneous graph edge set Construction

[0106] User-user edges: User-user edges are constructed using the distance between two users. The weight of the edge is the distance between the two users, which is used to express the user density and hotspot region structure.

[0107] User-Candidate Deployment Location Edge: This edge represents the coverage capability for IoT users when a node is used as a drone base station location. This edge includes information such as estimated path loss, the horizontal distance between the IoT device and the drone base station, and whether there are obstacles along the route.

[0108] Candidate-Candidate Edge: This edge represents the minimum safe distance constraint when multiple drones cooperate. The edge contains the location information of the drone base station and the distance information of other drone base stations to ensure the minimum safe distance constraint of the drone during the mission.

[0109] (4) Construction of node feature matrix and edge feature matrix

[0110] Input to the graph neural network: node feature matrix Edge feature matrix Adjacency structure , of which F V The number of features contained in each node, F e The number of features for each edge.

[0111] Once the heterogeneous graph is constructed, it can represent the complex spatial relationships in the disaster area, significantly improving the accuracy and generalization ability of UAV base station deployment prediction. The final heterogeneous graph is as follows: .

[0112] S102. Based on the heterogeneous graph model, a graph neural network is used to predict the quality of candidate UAV base station locations and generate a deployment priority distribution of candidate points.

[0113] like Figure 4 As shown, after constructing the heterogeneous map of the disaster area... Subsequently, this embodiment uses a graph neural network (GNN) to predict the deployment quality of candidate UAV base station locations, evaluate the merits of each candidate point as a UAV base station deployment location, and serve as an important basis for the initialization and search guidance of the subsequent differential evolution algorithm (DE).

[0114] Specifically, graph structure input and node embedding computation: the constructed heterogeneous graph Including node feature matrix Edge feature matrix Adjacency structure Input pre-trained model .

[0115] GNN uses a multi-layered message passing mechanism to communicate with each candidate deployment node. Neighborhood aggregation and feature fusion are performed. The information update process can be represented as follows:

[0116] (14)

[0117] in, Indicates the first Layer GNN nodes The implicit representation, Represents a node The neighborhood group, Describe the characteristics of the edges. This indicates attention-weighted aggregation. This represents the parameterized function of the MLP. Let u be the feature vector of node u in the l-th layer of the graph neural network, where u is the number of node features.

[0118] After propagation through layer L, the candidate deployment nodes obtain the final embedded representation. .

[0119] The deployment quality score for the candidate locations is then calculated: the node embeddings output by the GNN are mapped to a scalar deployment quality score. This is used to measure the merits of candidate locations as deployment sites for drone base stations.

[0120] (15)

[0121] in, This represents the MLP mapping layer, used to project high-dimensional node features into scalar space.

[0122] Finally, the deployment probability distribution is obtained through softmax: In order to transform the quality score of GNN into a probabilistic deployment priority, this embodiment uses the softmax function to normalize the quality scores of all candidate points, obtaining the deployment priority probability distribution p. j Specifically:

[0123] (16)

[0124] Where, q j For the quality score of the j-th candidate deployment point, q m Give a quality score to the m-th node among all candidate deployment points.

[0125] Through this step, this embodiment maps the complex spatial relationships of the disaster area into a candidate point deployment quality score, and further normalizes it into a probability distribution, providing high-value initialization and search directions for the subsequent differential evolution algorithm.

[0126] S103. Initialize the differential evolution algorithm population based on the deployment priority probability distribution.

[0127] like Figure 5As shown, the probability distribution of candidate deployment locations calculated based on GNN is obtained. Subsequently, this embodiment uses the probability distribution to guide the construction of the initial population of the Differential Evolution (DE) algorithm, enabling the DE algorithm to start searching from high-quality regions, thereby improving global optimization efficiency and convergence performance.

[0128] Center point sampling based on probability distribution: Let the set of candidate deployment locations be:

[0129] (17)

[0130] Among them, c j Deployment locations for candidate drones, , where are the three-dimensional location nodes of the candidate deployment drone locations, and J is the total number of candidate deployment locations.

[0131] The corresponding deployment probability is:

[0132] (18)

[0133] The population size for the differential evolution algorithm is set to Each individual represents a set of three-dimensional position vectors of a drone base station.

[0134] During the initialization phase, this embodiment selects a "center candidate point" for each individual in the following manner: based on probability distribution. Perform multiple samplings to obtain several candidate centroid indices. During the sampling process, candidate points with higher probability values ​​are more likely to be selected as center points, thus ensuring that the initial population is concentrated in the spatial region where the GNN prediction quality is high.

[0135] Introducing Gaussian perturbations near the center points to generate individuals: To expand discrete candidate points into feasible solutions in continuous space, this embodiment introduces Gaussian perturbations near each center point to generate initial individuals for the differential evolution algorithm. For the ... Each individual, whose corresponding center point is Gaussian noise is applied to the three coordinates in three-dimensional space respectively:

[0136] (19)

[0137] in, This indicates that the mean is 0 and the variance is . Gaussian random variables; when there are multiple drone base stations When there are 1, an individual can be represented as a three-dimensional vector group:

[0138] (20)

[0139] in, Let m be the three-dimensional vector group of the m-th individual.

[0140] Meanwhile, this embodiment adopts a cluster division method: first, the disaster area or user distribution is clustered, and then candidate points are sampled in each cluster as the center position of the corresponding UAV, so as to avoid multiple UAVs being concentrated in the same area to construct multiple UAV individuals.

[0141] To ensure that the generated individuals meet physical and business constraints, this embodiment performs boundary correction on the initialization results:

[0142] (1) If If the deployment exceeds the boundary of the deployment area, a mirroring / pruning strategy will be used to correct the boundary.

[0143] (2) If Exceeding the permitted flight altitude range If the value is ], then it is restricted to that interval by truncation.

[0144] (3) If an individual's location falls into a no-fly zone or a severely obstructed area, the center point can be resampled or a new disturbance can be generated.

[0145] S104. The initial population is iteratively updated multiple times through differential evolution to obtain an approximate global optimal solution for the three-dimensional position of the UAV base station.

[0146] like Figure 6 As shown, after initializing the differential evolution algorithm based on the candidate deployment probability distribution, this embodiment performs multiple generations of iterative updates on the initial population to obtain an approximate global optimal solution for the three-dimensional location of the UAV base station. This step includes fitness calculation, mutation operation, crossover operation, and selection operation.

[0147] First, calculate the fitness value: Let the current population size be... , No. The first generation Individuals are represented as:

[0148] (twenty one)

[0149] in, The number of drone base stations, Let be the three-dimensional location of the m-th individual in the population. For each individual, its coverage rate, energy consumption, and other indicators within the disaster area are calculated based on the air-to-ground link propagation model. This embodiment uses a multi-objective weighted approach to construct the fitness function. , can be represented as:

[0150] (twenty two)

[0151] in, and These are the weighting coefficients. Indicates energy consumption. This indicates coverage. A lower fitness value indicates a better solution.

[0152] Mutation operation: for each target individual Three different individuals are randomly selected from the current population. Construct the mutation vector.

[0153] (twenty three)

[0154] in, This is the mutation scaling factor. The mutation operation achieves a global exploration of the search space by amplifying the difference vector and superimposing it onto the basis vectors.

[0155] Crossover operation: To enhance population diversity and recombine information between original and mutated individuals, this embodiment employs a dimension-by-dimensional crossover strategy, crossovering the target individual... With variant individuals Combined generation of experimental individuals

[0156] (twenty four)

[0157] in, Indicates a dimension index. For crossover probability, The random dimension index is used to ensure that at least one dimension comes from the mutation vector, and rand(·) is a random number matrix. The crossover operation can introduce beneficial genes from the original individual while maintaining the ability to explore mutations, achieving a balance between exploration and development.

[0158] Selection operation: For each test individual Calculate its fitness and the fitness of the corresponding target individual. By comparing and selecting the better individuals for the next generation:

[0159] (25)

[0160] Through iteration, individuals in the population gradually converge towards a deployment scheme with better fitness. Iteration stops when the preset maximum number of iterations is reached.

[0161] S105. After the differential evolution algorithm converges iteratively, the individual with the lowest fitness is extracted from the final population to obtain the optimal three-dimensional deployment location of the UAV base station.

[0162] After the differential evolution algorithm converges iteratively, this invention extracts the individual with the smallest fitness value from the final population. :

[0163] (26)

[0164] in, Indicates the first The optimal three-dimensional deployment location for a drone base station.

[0165] like Figure 7 As shown, this embodiment also provides a UAV base station deployment system based on GNN and differential evolution, including: a model building module 1, a quality prediction module 2, a population initialization module 3, an iterative optimization module 4, and an optimal deployment output module 5;

[0166] Model building module 1 is used to acquire disaster areas and build a heterogeneous graph model based on the disaster areas;

[0167] Quality prediction module 2 is used to predict the quality of candidate UAV base station locations based on a heterogeneous graph model and using a graph neural network to generate a deployment priority distribution of candidate points;

[0168] Population initialization module 3 is used to initialize the differential evolution algorithm population based on the deployment priority probability distribution;

[0169] The iterative optimization module 4 is used to iteratively update the initial population multiple times through differential evolution iterative optimization to obtain an approximate global optimal solution for the three-dimensional position of the UAV base station;

[0170] The optimal deployment output module 5 is used to extract the individual with the lowest fitness from the final population after the differential evolution algorithm has converged, so as to obtain the optimal three-dimensional deployment location of the UAV base station.

[0171] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this disclosure is not limited to the described order of actions, because according to this disclosure, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this disclosure.

[0172] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0173] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for deploying unmanned aerial vehicle (UAV) base stations based on GNN and differential evolution, characterized in that, Includes the following steps: The process involves acquiring disaster areas and constructing a heterogeneous graph model based on these areas. This includes: collecting geographical information and user equipment information of the disaster areas, and using the K-means clustering method to generate a set of UAV base station deployment points to obtain a disaster area model; constructing a set of heterogeneous graph nodes and edges based on the disaster area model; and generating node feature matrices, edge feature matrices, and adjacency structures based on the set of heterogeneous graph nodes and edges to construct the heterogeneous graph model. The heterogeneous graph edge set includes user-user edges, user-candidate deployment location edges, and candidate-candidate edges. The user-user edges are constructed based on the distance between two users and are used to express user density and hotspot structure, where the edge weight is the distance between the two users. The user-candidate deployment location edges represent the coverage capability for IoT users when a node is used as a drone base station location. These user-candidate deployment location edges include estimated path loss, the horizontal distance between the IoT device and the drone base station, and information on whether there are obstacles along the path. The candidate-candidate edges represent the minimum safe distance constraint during multi-drone collaboration. These candidate-candidate edges include the location information of the drone base station and its distance information from other drone base stations. Based on the heterogeneous graph model, a graph neural network is used to predict the quality of candidate UAV base station locations and generate a deployment priority distribution of candidate points. The specific process of initializing the differential evolution algorithm population based on the deployment priority probability distribution of the candidate points is as follows: A set of candidate UAV base station three-dimensional locations is obtained, and individual vectors in the set of candidate UAV base station three-dimensional locations are sampled multiple times based on the deployment priority probability distribution to obtain several candidate center point indices. Gaussian perturbations are introduced near each center point to generate initial individuals for the differential evolution algorithm, thus obtaining the initialization results; The initialization results are corrected by boundary adjustments to obtain an initial population that satisfies both physical and operational constraints; By iteratively updating the initial population multiple times through differential evolution, an approximate global optimal solution for the three-dimensional position of the UAV base station is obtained. After the differential evolution algorithm converges iteratively, the individual with the lowest fitness is extracted from the final population to obtain the optimal three-dimensional deployment location of the UAV base station.

2. The method according to claim 1, characterized in that, The heterogeneous graph node set includes user equipment nodes and candidate drone deployment nodes; The user equipment node is generated based on terrestrial IoT devices and includes location coordinates, data demand, and user priority information; The candidate drone deployment nodes include three-dimensional candidate locations and information indicating whether the line of sight is obstructed.

3. The method according to claim 2, characterized in that, The process of using a graph neural network to predict the quality of candidate UAV base station locations based on the heterogeneous graph model and generating a deployment priority distribution of candidate points is as follows: The constructed heterogeneous graph model is input into the graph neural network, and neighborhood aggregation and feature fusion are performed on each candidate UAV deployment node through a multi-layer message passing mechanism to update the implicit representation of the node. After multiple layers of propagation, the final embedded representation of the candidate UAV deployment node is calculated; The final embedded representation is mapped to a scalar deployment quality score; The quality scores of all candidate drone deployment nodes are normalized using the softmax function to generate a deployment priority probability distribution.

4. The method according to claim 1, characterized in that, The process of boundary correction of the initialization results to obtain an initial population that satisfies physical and operational constraints includes: If individual x i y i If the deployment exceeds the boundary of the deployment area, a mirroring / pruning strategy is used for boundary correction, where x i y i These are the horizontal and vertical coordinates of the candidate drone deployment locations; If the individual's h i If the flight altitude exceeds the permitted range, a truncation method will be used to limit the boundary to within the permitted range, where h i Deployment altitude for candidate drones; If an individual's location falls into a no-fly zone or a severely obstructed area, the center point index needs to be resampled or a Gaussian perturbation needs to be introduced back into the center point.

5. The method according to claim 1, characterized in that, The process of obtaining an approximate global optimal solution for the three-dimensional location of the UAV base station by iteratively updating the initial population multiple times through differential evolution optimization is as follows: The air-to-ground link propagation model was used to calculate the coverage rate and energy consumption index of each individual in the population within the disaster area, and the fitness function was constructed using a multi-objective weighted method. For each target individual in the population, three different individuals are randomly selected from the current population to construct a mutation vector, thus obtaining the mutated individual; A dimension-by-dimensional crossover strategy is used to combine the target individual with the variant individual to generate experimental individuals; For each experimental individual, its fitness is calculated and compared with the fitness of the corresponding target individual. The better individual is selected to enter the next generation. The iteration stops when the number of iterations reaches the preset maximum number of iterations.

6. A UAV base station deployment system based on GNN and differential evolution, used to implement the method as described in any one of claims 1-5, characterized in that, include: The module consists of a model building module (1), a quality prediction module (2), a population initialization module (3), an iterative optimization module (4), and an optimal deployment output module (5). The model building module (1) is used to obtain disaster areas and build a heterogeneous graph model based on the disaster areas; The quality prediction module (2) is used to perform quality prediction on the candidate UAV base station locations based on the heterogeneous graph model and using a graph neural network to generate a deployment priority distribution of candidate points; The population initialization module (3) is used to initialize the differential evolution algorithm population based on the deployment priority probability distribution; The iterative optimization module (4) is used to perform multiple iterations of the initial population through differential evolution iterative optimization to obtain an approximate global optimal solution for the three-dimensional position of the UAV base station; The optimal deployment output module (5) is used to extract the individual with the smallest fitness from the final population after the differential evolution algorithm has converged, so as to obtain the optimal three-dimensional deployment position of the UAV base station.

Images