Unmanned aerial vehicle wind environment spatio-temporal prediction method
Patent Information
- Application Number
- CN202410974391.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-19
- Publication Date
- 2026-10-27
- Estimated Expiration
- 2044-07-19
AI Technical Summary
但是现有的风环境预测方法多采用单点预测，只能够得到单个站点的预测信息，无法提供给无人机更准确的飞行环境信息
[0070]本发明公开了一种无人机风环境时空预测方法，通过集合气象条件和风场间的空间影响建立风环境时空预测模型，基于图注意力网络的框架，可以实现无人机风环境的智能预测，提高了预测效率，改善了无人机的工作环境，具有广阔的应用潜力。
Smart Images
Figure CN118965130B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of unmanned aerial vehicle (UAV) operation safety, and specifically relates to a method for spatiotemporal prediction of wind environment for UAVs. Background Technology
[0002] Drone logistics refers to the use of radio remote control equipment and onboard program control devices to operate drones to automatically deliver packages to their destinations. The advantages of drone logistics include solving delivery problems in remote areas, improving delivery efficiency, and reducing labor costs.
[0003] The safety of existing drone logistics flights is affected by a variety of factors. Aside from the drone's own hardware, drones are subject to numerous external conditions during flight, such as wind, rain, clouds, temperature, altitude, and magnetic fields. Wind is an unavoidable operating environment for drones, directly impacting their flight trajectory and attitude. Drones used in drone logistics require multiple sensors and are themselves expensive, while the goods they carry also possess value; therefore, ensuring the safety of the drone's flight environment is crucial.
[0004] Currently, drones used for unmanned aerial vehicle (UAV) logistics are generally small to medium-sized. These drones have limited wind resistance; excessive wind speeds can directly cause them to lose control, resulting in damage to both the drone and the cargo. Therefore, to ensure drone flight safety, it is necessary to monitor the wind environment along the drone's flight path. In actual flight, wind is a major factor affecting the drone's heading and speed, and can also increase energy consumption, thus affecting its endurance. Therefore, relying solely on wind environment monitoring for drone route planning has a lag; it is necessary to predict the wind environment during drone flight.
[0005] In recent years, with the development of numerical weather prediction technology, the accuracy and timeliness of wind environment forecasts have been improved. This has provided a foundation for wind environment forecasting for UAVs. However, existing wind environment forecasting methods mostly adopt single-point forecasts, which can only obtain forecast information from a single site and cannot provide UAVs with more accurate flight environment information. Summary of the Invention
[0006] The purpose of this invention is to provide a method for spatiotemporal prediction of wind environment using unmanned aerial vehicles (UAVs) that can provide multi-site wind speed prediction.
[0007] This invention provides a method for spatiotemporal prediction of wind environment using unmanned aerial vehicles (UAVs), comprising the following steps:
[0008] S1. Obtain wind speed data;
[0009] S2. Process the wind speed data obtained in step S1 to obtain the UAV wind environment dataset;
[0010] S3. Based on the gated recurrent unit and Shapley additive interpretation model, meteorological factors are screened from the UAV wind environment dataset obtained in step S2 to obtain the training dataset;
[0011] S4. Based on graph attention network, multi-feature fusion and graph optimization are added to construct an initial model for spatiotemporal prediction of UAV wind environment;
[0012] S5. Train the initial model for UAV wind environment spatiotemporal prediction obtained in step S4 based on the training dataset obtained in step S3 to obtain the UAV wind environment spatiotemporal prediction model.
[0013] S6. Based on the UAV wind environment spatiotemporal prediction model obtained in step S5, perform actual UAV wind environment spatiotemporal prediction.
[0014] Step S1 specifically involves: acquiring initial wind speed data through a wind speed sensor, including temperature, air pressure, and wind speed; performing preliminary analysis and filtering on the initial wind speed data to obtain wind speed data; the preliminary analysis includes statistical regularity analysis, rationality analysis, and comparability analysis.
[0015] Statistical regularity analysis involves using mathematical methods or common physical methods to statistically analyze and compare initial wind speed data in order to find the inherent regularity in the initial wind speed data.
[0016] Reasonableness analysis involves modeling the correlation between initial wind speed data to determine the reasonableness of the data;
[0017] Comparability analysis involves comparing and contrasting different data to identify and compare differences, optimize different data variables, and ensure the comprehensiveness of the data.
[0018] Step S2 includes the following steps:
[0019] a. For the missing data in the wind speed data obtained in step S1, forward interpolation is used to complete the missing wind speed data;
[0020] b. For the wind speed data obtained in step a, the missing data is filled in, and the outliers with different distribution characteristics are detected and corrected by the data processing method based on probabilistic autoencoder, so as to obtain the processed wind speed data.
[0021] The probabilistic autoencoder calculates the possible fluctuation range corresponding to each sequence point by inputting the original sequence, uses maximum likelihood loss instead of mean square error, and transforms the features compressed by the encoder into the data point distribution of the decoder. It judges whether the data point is abnormal by comparing the relationship between the measured value of the data point and the distribution range.
[0022] c. The processed wind speed data obtained in step b is standardized to obtain a standardized wind environment dataset;
[0023] The standardization process uses Z-score standardization, and the processed data conforms to a standard normal distribution (0,1). The transformation function is expressed by the following formula:
[0024]
[0025] Where μ is the mean of the processed wind speed data; σ is the standard deviation of the processed wind speed data; x * This is a dataset of wind environment conditions for drones; x represents the processed wind speed data.
[0026] d. The one-dimensional time series dataset is converted into a feature vector dataset using the sliding window method. The sliding window method uses N historical wind speed observation data from the wind environment standardized dataset obtained in step c to form a feature vector according to the sampling order, where N is the length of the feature vector. The future wind speed observation data at the next M time points are used as sample labels. M is the number of model output variables, corresponding to the prediction results of different time lengths, to obtain the UAV wind environment dataset.
[0027] Step S3 includes the following steps:
[0028] A. Based on the Shapley additive interpretation model, the degree of influence of meteorological factors on the wind field is obtained;
[0029] Based on the Shapley additive interpretation model, the tree model uses the feature values from the UAV wind environment dataset obtained in step S2 to divide each region into two sub-regions, constructs a binomial decision tree, and selects the optimal cutting variable j and cutting point s to minimize the objective function. The objective function is expressed by the following formula:
[0030]
[0031] Where R1(j,s)={x|x (j) ≤s}; c1 is the first sub-region; c2 is the second sub-region; c1 is the average value of the first sub-region; c2 is the average value of the second sub-region; y i This represents the wind speed value; x i For timestamps;
[0032] The selected regions are segmented, and the corresponding output values are defined, expressed by the following formula:
[0033]
[0034] Where R1(j,s)={x|x (j) ≤s}, R2(j,s)={x|x (j)>s};N m The length of the meteorological factor sequence; This is the average value;
[0035] The system continuously divides sub-regions until the number of samples in a node is less than a preset threshold. Based on the input features, it obtains the space corresponding to the input samples and uses the average value of the samples in that space as the output value. Finally, it determines the degree of influence of each meteorological factor on wind direction and wind speed prediction.
[0036] B. Sort the meteorological factors obtained in step A by their degree of influence on the wind field, select the meteorological factors with the highest degree of influence, and fuse them with the feature vectors in the UAV wind environment dataset obtained in step S2 to form a multi-source dataset and obtain an intermediate dataset.
[0037] C. Establish a single-site prediction model based on the gated cyclic unit, input the intermediate dataset obtained in step B into the single-site prediction model, obtain the prediction results, compare the impact of various meteorological factors on the model, and select the training dataset.
[0038] The single-site prediction model based on the doorway loop unit includes updating the door and resetting the door, and is expressed by the following formula:
[0039] Z(t)=β(W z ·[H(t-1),x t ])
[0040] R(t)=β(W R ·[H(t-1),x t ])
[0041]
[0042] H(t)=(1-Z(t))*H(t-1)+Z(t)*H(t)
[0043] Where R(t) is the reset gate weight of the gated loop unit; Z(t) is the update gate weight of the loop unit at the gate; and H(t) is the candidate activation state. β represents the active state; β is the activation function; x t The input is the current state; t is the state parameter; W z To update the gate weights; W R is the reset gate weight; W is the weight of the current state;
[0044] The update gate is determined by the historical information that needs to be forgotten from the current state and the new information received. The reset gate is determined by the information that the candidate state obtains from the historical information. The core function of the update gate is to determine how much state information from the previous moment can be incorporated into the current state. The larger the value obtained by the update gate, the more state information the gating loop unit ultimately incorporates.
[0045] The main function of the reset gate is to determine the amount of state information from the previous moment that will be discarded. The smaller the value of the reset gate, the more state information will be discarded.
[0046] The initial model for spatiotemporal prediction of UAV wind environment described in step S4 includes a graph structure optimization module and a graph attention network module connected in series.
[0047] The feature extraction module uses a sliding window method to convert a one-dimensional time series dataset into a feature vector dataset. The sliding window method uses N historical wind speed observation data to form a feature vector according to the sampling order, where N is the length of the feature vector. The future wind speed observation data at the next M time points are used as sample labels. M is the number of output variables of the model, corresponding to prediction results of different time lengths.
[0048] The graph structure optimization module is based on the kNN graph structure optimization algorithm. By calculating the distance between each node's features and other features, it finds the k nodes with the smallest distance to the target node, establishes the adjacency relationship between the target node and these k nodes, sets the corresponding position of the adjacency matrix to 1, and after traversing all nodes, it obtains the adjacency matrix A', which is used as the current adjacency matrix A.
[0049] Methods for measuring distance between nodes include Euclidean distance and dynamic time-warped distance;
[0050] In the graph attention network module, each node is assigned different weights based on the features of its neighboring nodes, and features are aggregated by calculating the similarity between each node and its neighboring nodes.
[0051] target node v i With neighboring node v j The similarity between them is expressed by the following formula:
[0052] e ij =Wh i ·(Wh j ) T
[0053] Where j = 1, 2, ..., n, i ≠ j, and n is the number of nodes with adjacency relationships; e ij For target node v i With neighboring node v j Similarity between them; W is the weight of the linear layer; h i target node vi Features; h j For adjacent node v j The characteristics of the nodes; the attention weight coefficients between nodes are expressed by the following formula:
[0054]
[0055] Where, α ij For target node v i With neighboring node v j Attention weight coefficients between; N(v) j ) is the neighboring node v j The set of nodes; L is the activation function; the aggregated target node v i Features h i Express it using the following formula:
[0056]
[0057] Where σ is a nonlinear activation function; after introducing a multi-head attention mechanism, the target node v i Features h i Express it using the following formula:
[0058]
[0059] Where ||· represents the result of merging multiple attention mechanisms; the multi-head attention mechanism has K independent attention mechanisms; k represents the k-th head of the multi-head attention mechanism; W k N represents the weighting coefficients of the attention in the k-th head of the multi-head attention mechanism. i Let i be any neighborhood of node i in the graph;
[0060] Methods for measuring distance between nodes include Euclidean distance and dynamic time-warped distance.
[0061] The training described in step S5 is as follows: the UAV wind environment dataset obtained in step S2 is input into the initial model for spatiotemporal prediction of UAV wind environment obtained in step S4. The UAV wind environment dataset first extracts features through the feature extraction module, and then processes them in the graph structure optimization module to obtain the current adjacency matrix A.
[0062] During training, the attention weight coefficients between nodes in the graph attention network module are replaced with new weight coefficients, expressed by the following formula:
[0063]
[0064] Where, β ij The coefficients are β; the graph attention network module is trained based on the current adjacency matrix A to obtain the coefficients β. ij For coefficient βij Make the following judgments:
[0065] When β ij When the value is greater than 1, let the new adjacency matrix A be... * element at the corresponding position
[0066] When β ij When = 1, let the new adjacency matrix A * element at the corresponding position Where A ij This is the value of the element at the corresponding position in the current adjacency matrix A;
[0067] When β ij When <1, generate a random number r and r∈(0,1), then if r>1-β ij Then let the new adjacency matrix A * element at the corresponding position Otherwise, let the new adjacency matrix A * element at the corresponding position
[0068] For the obtained coefficient β ij After the judgment is completed, the new adjacency matrix A is obtained. * Determine if the stopping condition is met. If not, change the new adjacency matrix A. * As the current adjacency matrix A, repeat the above process of training the graph attention network module based on the current adjacency matrix A; otherwise, output the new adjacency matrix A. * To ensure algorithm convergence, a constant λ is used to replace the 0s in the adjacency matrix A, where λ ∈ (0, 10). -5 );
[0069] After training, a spatiotemporal prediction model for wind environment using unmanned aerial vehicles (UAVs) was obtained.
[0070] This invention discloses a spatiotemporal prediction method for wind environment of unmanned aerial vehicles (UAVs). By combining meteorological conditions and the spatial influence between wind fields, a spatiotemporal prediction model for wind environment is established. Based on the framework of graph attention network, intelligent prediction of wind environment of UAVs can be achieved, which improves prediction efficiency, improves the working environment of UAVs, and has broad application potential. Attached Figure Description
[0071] Figure 1 This is a schematic flowchart of the method of the present invention;
[0072] Figure 2 This illustrates the impact of various meteorological factors on wind speed prediction in the embodiments of the present invention.
[0073] Figure 3 This illustrates the impact of various meteorological factors on wind direction prediction in this embodiment of the invention. Detailed Implementation
[0074] This invention provides a method for spatiotemporal prediction of wind environment using unmanned aerial vehicles (UAVs), the flowchart of which is shown below. Figure 1 As shown, it includes the following steps:
[0075] S1. Obtain wind speed data, specifically:
[0076] Initial wind speed data is acquired through a wind speed sensor, which includes several parameters such as temperature, air pressure, and wind speed. The initial wind speed data is then preliminarily analyzed and filtered to obtain the wind speed data. The preliminary analysis includes statistical regularity analysis, rationality analysis, and comparability analysis.
[0077] Statistical regularity analysis involves using mathematical methods or common physical methods to statistically analyze and compare initial wind speed data in order to find the inherent regularity in the initial wind speed data.
[0078] Reasonableness analysis involves modeling the correlation between initial wind speed data to determine the reasonableness of the data;
[0079] Comparability analysis involves comparing and contrasting different data to identify and compare differences, optimize different data variables, and ensure the comprehensiveness of the data.
[0080] S2. Process the wind speed data obtained in step S1 to obtain the UAV wind environment dataset, including the following steps:
[0081] a. For the missing data in the wind speed data obtained in step S1, forward interpolation is used to complete the missing wind speed data;
[0082] b. For the wind speed data obtained in step a, the missing data is filled in, and the outliers with different distribution characteristics are detected and corrected by the data processing method based on probabilistic autoencoder, so as to obtain the processed wind speed data.
[0083] The probabilistic autoencoder calculates the possible fluctuation range corresponding to each sequence point by inputting the original sequence, uses maximum likelihood loss instead of mean square error, and transforms the features compressed by the encoder into the data point distribution of the decoder. It judges whether the data point is abnormal by comparing the relationship between the measured value of the data point and the distribution range.
[0084] c. The processed wind speed data obtained in step b is standardized to obtain the UAV wind environment dataset;
[0085] The standardization process uses Z-score standardization, and the processed data conforms to a standard normal distribution (0,1). The transformation function is expressed by the following formula:
[0086]
[0087] Where μ is the mean of the processed wind speed data; σ is the standard deviation of the processed wind speed data; x * This is a dataset of wind environment conditions for drones; x represents the processed wind speed data.
[0088] d. The one-dimensional time series dataset is converted into a feature vector dataset using the sliding window method. The sliding window method uses N historical wind speed observation data from the wind environment standardized dataset obtained in step c to form a feature vector according to the sampling order, where N is the length of the feature vector. The future wind speed observation data at the next M time points are used as sample labels. M is the number of model output variables, corresponding to the prediction results of different time lengths, to obtain the UAV wind environment dataset.
[0089] S3. Based on the gated recurrent unit and Shapley additive interpretation model, meteorological factors are filtered from the UAV wind environment dataset obtained in step S2 to obtain the training dataset, including the following steps:
[0090] A. Based on the Shapley additive interpretation model, the degree of influence of meteorological factors on the wind field is obtained;
[0091] Based on the Shapley additive interpretation model, the tree model uses the feature values from the UAV wind environment dataset obtained in step S2 to divide each region into two sub-regions, constructs a binomial decision tree, and selects the optimal cutting variable j and cutting point s to minimize the objective function. The objective function is expressed by the following formula:
[0092]
[0093] Where R1(j,s)={x|x (j) ≤s}; c1 is the first sub-region; c2 is the second sub-region; c1 is the average value of the first sub-region; c2 is the average value of the second sub-region; y i This represents the wind speed value; x i For timestamps;
[0094] The selected regions are segmented, and the corresponding output values are defined, expressed by the following formula:
[0095]
[0096] Where R1(j,s)={x|x (j) ≤s}, R2(j,s)={x|x (j) >s};N m The length of the meteorological factor sequence; This is the average value;
[0097] The system continuously divides the data into sub-regions until the number of samples in a node is less than a preset threshold. Based on the input features, it obtains the space corresponding to the input samples and uses the average value of the samples in that space as the output value. Finally, it determines the degree of influence of each meteorological factor on wind direction and wind speed prediction. Figure 2 , Figure 3 As shown;
[0098] B. Sort the meteorological factors obtained in step A by their degree of influence on the wind field, select the meteorological factors with the highest degree of influence, and fuse them with the feature vectors in the UAV wind environment dataset obtained in step S2 to form a multi-source dataset and obtain an intermediate dataset.
[0099] C. Establish a single-site prediction model based on the gated cyclic unit, input the intermediate dataset obtained in step B into the single-site prediction model, obtain the prediction results, compare the impact of various meteorological factors on the model, and select the training dataset.
[0100] The single-site prediction model based on the doorway loop unit includes updating the door and resetting the door, and is expressed by the following formula:
[0101] Z(t)=β(W z ·[H(t-1),x t ])
[0102] R(t)=β(W R ·[H(t-1),x t ])
[0103]
[0104] H(t)=(1-Z(t))*H(t-1)+Z(t)*H(t)
[0105] Where R(t) is the reset gate weight of the gated loop unit; Z(t) is the update gate weight of the loop unit at the gate; and H(t) is the candidate activation state. The active state is represented by β, which is the activation function. t The input is the current state; t is the state parameter; W z To update the gate weights; W R is the reset gate weight; W is the weight of the current state;
[0106] The update gate is determined by the historical information that needs to be forgotten from the current state and the new information received. The reset gate is determined by the information that the candidate state obtains from the historical information. The core function of the update gate is to determine how much state information from the previous moment can be incorporated into the current state. The larger the value obtained by the update gate, the more state information the gating loop unit ultimately incorporates.
[0107] The main function of the reset gate is to determine the amount of state information from the previous moment that will be discarded. The smaller the value of the reset gate, the more state information will be discarded.
[0108] S4. Based on graph attention network, multi-feature fusion and graph optimization are added to construct an initial model for spatiotemporal prediction of UAV wind environment;
[0109] Combining the selected meteorological conditions mentioned above, we use them as features to establish an initial model for spatiotemporal prediction of UAV wind environment based on graph neural network.
[0110] The initial model for spatiotemporal prediction of wind environment by UAV includes a graph structure optimization module and a graph attention network module connected in series.
[0111] The graph structure optimization module is based on the kNN graph structure optimization algorithm. By calculating the distance between each node's features and other features, it finds the k nodes with the smallest distance to the target node, establishes the adjacency relationship between the target node and these k nodes, sets the corresponding position of the adjacency matrix to 1, and after traversing all nodes, it obtains the adjacency matrix A', which is used as the current adjacency matrix A.
[0112] Methods for measuring distance between nodes include Euclidean distance and dynamic time-warped distance;
[0113] In the graph attention network module, each node is assigned different weights based on the features of its neighboring nodes, and features are aggregated by calculating the similarity between each node and its neighboring nodes.
[0114] target node v i With neighboring node v j The similarity between them is expressed by the following formula:
[0115] e ij =Wh i ·(Wh j ) T
[0116] Where j = 1, 2, ..., n, i ≠ j, and n is the number of nodes with adjacency relationships; e ij For target node v i With neighboring node v j Similarity between them; W is the weight of the linear layer; h i target node v i Features; h j For adjacent node v j The characteristics of the nodes; the attention weight coefficients between nodes are expressed by the following formula:
[0117]
[0118] Where, α ijFor target node v i With neighboring node v j Attention weight coefficients between; N(v) j ) is the neighboring node v j The set of nodes; L is the activation function; the aggregated target node v i Features h i Express it using the following formula:
[0119]
[0120] Where σ is a nonlinear activation function; after introducing a multi-head attention mechanism, the target node v i Features h i Express it using the following formula:
[0121]
[0122] Where ||· represents the result of merging multiple attention mechanisms; the multi-head attention mechanism has K independent attention mechanisms; k represents the k-th head of the multi-head attention mechanism; W k N represents the weighting coefficients of the attention in the k-th head of the multi-head attention mechanism. i Let i be any neighborhood of node i in the graph;
[0123] Methods for measuring distance between nodes include Euclidean distance and dynamic time-warped distance.
[0124] S5. Train the initial UAV wind environment spatiotemporal prediction model obtained in step S4 using the training dataset obtained in step S3 to obtain the UAV wind environment spatiotemporal prediction model. The training specifically involves:
[0125] The training dataset obtained in step S3 is input into the initial model for spatiotemporal prediction of UAV wind environment obtained in step S4. The UAV wind environment dataset first extracts features through the feature extraction module, and then processes them in the graph structure optimization module to obtain the current adjacency matrix A.
[0126] During training, the attention weight coefficients between nodes in the graph attention network module are replaced with new weight coefficients, expressed by the following formula:
[0127]
[0128] Where, β ij The coefficients are β; the graph attention network module is trained based on the current adjacency matrix A to obtain the coefficients β. ij For coefficient β ij Make the following judgments:
[0129] When β ij When the value is greater than 1, let the new adjacency matrix A be...* element at the corresponding position
[0130] When β ij When = 1, let the new adjacency matrix A * element at the corresponding position Where A ij This is the value of the element at the corresponding position in the current adjacency matrix A;
[0131] When β ij When <1, generate a random number r and r∈(0,1), then if r>1-β ij Then let the new adjacency matrix A * element at the corresponding position Otherwise, let the new adjacency matrix A * element at the corresponding position
[0132] For the obtained coefficient β ij After the judgment is completed, the new adjacency matrix A is obtained. * Determine if the stopping condition is met. If not, change the new adjacency matrix A. * As the current adjacency matrix A, repeat the above process of training the graph attention network module based on the current adjacency matrix A; otherwise, output the new adjacency matrix A. * To ensure algorithm convergence, a constant λ is used to replace the 0s in the adjacency matrix A, where λ ∈ (0, 10). -5 );
[0133] After training, a spatiotemporal prediction model for wind environment using unmanned aerial vehicles (UAVs) was obtained.
[0134] S6. Based on the UAV wind environment spatiotemporal prediction model obtained in step S5, perform actual UAV wind environment spatiotemporal prediction.
Claims
1. A spatiotemporal prediction method for wind environment using unmanned aerial vehicles (UAVs), characterized in that, Includes the following steps: S1. Obtain wind speed data; S2. Process the wind speed data obtained in step S1 to obtain the UAV wind environment dataset; S3. Based on the gated recurrent unit and Shapley additive interpretation model, meteorological factors are screened from the UAV wind environment dataset obtained in step S2 to obtain the training dataset; S4. Based on graph attention network, multi-feature fusion and graph optimization are added to construct an initial model for spatiotemporal prediction of UAV wind environment; S5. Train the initial model for UAV wind environment spatiotemporal prediction obtained in step S4 based on the training dataset obtained in step S3 to obtain the UAV wind environment spatiotemporal prediction model. S6. Based on the UAV wind environment spatiotemporal prediction model obtained in step S5, perform actual UAV wind environment spatiotemporal prediction; Step S2 includes the following steps: a. For the missing data in the wind speed data obtained in step S1, forward interpolation is used to complete the missing wind speed data; b. For the wind speed data obtained in step a, the missing data is filled in, and the outliers with different distribution characteristics are detected and corrected by the data processing method based on probabilistic autoencoder, so as to obtain the processed wind speed data. The probabilistic autoencoder calculates the possible fluctuation range corresponding to each sequence point by inputting the original sequence, uses maximum likelihood loss instead of mean square error, and transforms the features compressed by the encoder into the data point distribution of the decoder. It judges whether the data point is abnormal by comparing the relationship between the measured value of the data point and the distribution range. c. The processed wind speed data obtained in step b is standardized to obtain the UAV wind environment dataset; The standardization process uses Z-score standardization, and the processed data conforms to... The transformation function of the standard normal distribution is expressed by the following formula: in, This represents the average of the processed wind speed data; The standard deviation of the processed wind speed data; Data set for drone wind environment; The processed wind speed data; d. Use the sliding window method to convert the one-dimensional time series dataset into a feature vector dataset; use the sliding window method to standardize the wind environment dataset obtained in step c. A feature vector is formed from historical wind speed observation data in the order of sampling. The length of the feature vector; subsequent... The future wind speed observation data at each moment is used as the sample label; The number of variables output by the model corresponds to the prediction results for different pace lengths, thus obtaining the UAV wind environment dataset; Step S3 includes the following steps: A. Based on the Shapley additive interpretation model, the degree of influence of meteorological factors on the wind field is obtained; Based on the Shapley additive interpretation model, a tree model is used, employing the feature values from the UAV wind environment dataset obtained in step S2, to divide each region into two sub-regions, construct a binomial decision tree, and select the optimal splitting variable. and cutting point The objective function is minimized; the objective function is expressed by the following formula: in, ; This is the first sub-region; This is the second sub-region; This represents the average value of the first sub-region. This represents the average value of the second sub-region. This represents the wind speed value. For timestamps; The selected regions are segmented, and the corresponding output values are defined, expressed by the following formula: in, , ; The length of the meteorological factor sequence; This is the average value; The system continuously divides sub-regions until the number of samples in a node is less than a preset threshold. Based on the input features, it obtains the space corresponding to the input samples and uses the average value of the samples in that space as the output value. Finally, it determines the degree of influence of each meteorological factor on wind direction and wind speed prediction. B. Sort the meteorological factors obtained in step A by their degree of influence on the wind field, select the meteorological factors with the highest degree of influence, and fuse them with the feature vectors in the UAV wind environment dataset obtained in step S2 to form a multi-source dataset and obtain an intermediate dataset. C. Establish a single-site prediction model based on the gated cyclic unit, input the intermediate dataset obtained in step B into the single-site prediction model, obtain the prediction results, compare the impact of various meteorological factors on the model, and select the training dataset. The initial model for spatiotemporal prediction of UAV wind environment described in step S4 includes a graph structure optimization module and a graph attention network module connected in series. The graph structure optimization module is based on the kNN graph structure optimization algorithm. It calculates the distance between each node's features and other features to find the node with the smallest distance to the target node. Establish a connection between the target node and this node. To determine the adjacency relationships of nodes, set the corresponding position in the adjacency matrix to 1. After traversing all nodes, the adjacency matrix is obtained. As the current adjacency matrix ; Methods for measuring distance between nodes include Euclidean distance and dynamic time-warped distance; In the graph attention network module, each node is assigned different weights based on the features of its neighboring nodes, and features are aggregated by calculating the similarity between each node and its neighboring nodes.
2. The UAV wind environment spatiotemporal prediction method according to claim 1, characterized in that, Step S1 specifically involves: acquiring initial wind speed data through a wind speed sensor, including temperature, air pressure, and wind speed; performing preliminary analysis and filtering on the initial wind speed data to obtain wind speed data; the preliminary analysis includes statistical regularity analysis, rationality analysis, and comparability analysis. Reasonableness analysis involves modeling the correlation between initial wind speed data to determine the reasonableness of the data.
3. The UAV wind environment spatiotemporal prediction method according to claim 2, characterized in that, The single-site prediction model based on the doorway loop unit includes updating the door and resetting the door, and is expressed by the following formula: in, The reset gate weight for the gated loop unit; Update the door weight for the loop unit at the entrance; In candidate activation state; In active state; For activation functions; Input for the current state; These are state parameters; To update the gate weights; Reset gate weight; The weight of the current state; The update gate is determined by the historical information that needs to be forgotten in the current state and the new information received; the reset gate is determined by the information that the candidate state obtains from the historical information; the core function of the update gate is to determine how much state information from the previous moment can be introduced into the current state; the larger the value obtained by the update gate, the more state information the gated loop unit ultimately introduces. The main function of the reset gate is to determine the amount of state information from the previous moment that will be discarded. The smaller the value of the reset gate, the more state information will be discarded.
4. The UAV wind environment spatiotemporal prediction method according to claim 3, characterized in that, In the graph attention network module, the target node With neighboring nodes The similarity between them is expressed by the following formula: in, , This represents the number of nodes that are adjacent to each other. For the target node With neighboring nodes Similarity between them; The weights of the linear layer; target node Features; Adjacent nodes The characteristics of the nodes; the attention weight coefficients between nodes are expressed by the following formula: in, For the target node With neighboring nodes Attention weight coefficients between them; Adjacent nodes A set; The activation function; the target node after aggregation. Features Express it using the following formula: in, It is a non-linear activation function; after introducing a multi-head attention mechanism, the target node Features Express it using the following formula: in, To merge the results of multiple attention mechanisms; multi-head attention mechanisms have A separate attention mechanism; The first part representing the multi-head attention mechanism Size; For the multi-head attention mechanism The weighting coefficient of attention in terms of size; For nodes in the graph any neighborhood; Methods for measuring distance between nodes include Euclidean distance and dynamic time-warped distance.
5. The UAV wind environment spatiotemporal prediction method according to claim 4, characterized in that, The training in step S5 specifically involves replacing the attention weight coefficients between nodes in the graph attention network module with new weight coefficients, expressed by the following formula: = in, The coefficients are based on the current adjacency matrix. Train the graph attention network module to obtain the coefficients. Each iteration is based on The values in the adjacency matrix are adjusted to obtain the spatiotemporal prediction model of the UAV wind environment after training is completed.
6. The UAV wind environment spatiotemporal prediction method according to claim 5, characterized in that, According to The values are adjusted in the adjacency matrix, specifically as follows: For coefficients Make the following judgments: when At that time, let the new adjacency matrix element at the corresponding position ; when At that time, let the new adjacency matrix element at the corresponding position ,in The current adjacency matrix The value of the element at the corresponding position in the middle; when At that time, generate random numbers and If Then let the new adjacency matrix element at the corresponding position Otherwise, let the new adjacency matrix be... element at the corresponding position ; For the obtained coefficients After the judgment is completed, the new adjacency matrix is obtained. Determine if the stopping condition is met; if not, create a new adjacency matrix. As the current adjacency matrix Repeat the above based on the current adjacency matrix. The process of training the graph attention network module; otherwise, a new adjacency matrix is output. To ensure algorithm convergence, a constant is used. Replace adjacency matrix 0 and .