A wind field prediction method and system based on cooperation of a UAV cluster

Abstract

The application discloses a kind of based on unmanned aerial vehicle cluster cooperation's wind field prediction method and system, it is related to aircraft technical field, including the following steps: the target area is divided into multiple three-dimensional grid nodes and corresponding perception domain, the multiple key features corresponding to each unmanned aerial vehicle are mapped to corresponding three-dimensional grid node according to preset rule;Network node, node feature and edge weight are input into graph neural network, and the global wind field feature map of target area real-time is constructed;Global wind field feature map in the current setting time period is input into pre-trained wind field prediction model, and the three-dimensional wind field prediction value of future setting time period of target area is obtained.The global wind field feature map of the application has the spatial dimension corresponding to the target area and the channel dimension corresponding to the wind field feature, forms the cooperative perception network covering low-altitude wide-area region, and improves the wind field prediction result accuracy of target area.

Description

Technical Field

[0001] This invention relates to the field of aircraft technology, and in particular to a wind field prediction method and system based on UAV swarm collaboration. Background Technology

[0002] As a new economic form with a wide reach and long industrial chain, the low-altitude economy has become an important direction for my country's high-quality economic development. The low-altitude economy encompasses multiple fields such as drone logistics, emergency rescue, agricultural plant protection, and urban transportation. Its core relies on safe and efficient flight activities. These application scenarios place extremely high demands on the flight safety and efficiency of drones, and accurate perception and prediction of the flight environment are the foundation for achieving all of this. Simultaneously, accurate wind field prediction can support dynamic optimization of flight paths. Drones can proactively choose tailwind routes and avoid strong headwind areas based on forecast information, thereby significantly reducing energy consumption, extending flight range, and shortening mission execution time, propelling the low-altitude economy from initial application to a mature and intelligent stage.

[0003] However, the low-altitude airspace presents complex meteorological conditions and dynamic obstacles (such as trees, buildings, and other aircraft), resulting in low-altitude wind fields exhibiting characteristics such as suddenness, complex spatiotemporal variations, multi-factor influence, non-stationarity and nonlinearity, local and regional differences, and uncertainties under complex meteorological conditions. These characteristics pose significant challenges to the perception and prediction of low-altitude wind fields, requiring high-precision, high spatiotemporal resolution sensing technologies to address them.

[0004] Existing wind field prediction methods collect wind field parameters of the target area using fixed sensors (such as weather stations and wind profiler radars) or single UAVs. This can only achieve local and small-scale wind field perception and cannot cover a comprehensive and accurate perception of the dynamic wind field over a wide area, resulting in low accuracy of wind field prediction results for the target area. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the present invention provides a wind field prediction method and system based on UAV swarm collaboration, which solves the existing problems.

[0006] The present invention adopts the following technical solution:

[0007] In a first aspect, the present invention provides a wind field prediction method based on UAV swarm collaboration, comprising the following steps:

[0008] The target area is divided into multiple three-dimensional grid nodes and corresponding perception domains. A drone swarm is deployed in the target area, and the drones fly in multiple perception domains according to a preset trajectory.

[0009] The flight parameters of each UAV are collected in real time, and the three-dimensional wind speed vector of the current UAV area is inverted in real time based on the UAV flight parameters; feature extraction is performed on each three-dimensional wind speed vector to obtain multiple key features for characterizing wind field uncertainty; the multiple key features corresponding to each UAV are mapped to the corresponding three-dimensional grid nodes according to preset rules.

[0010] Using multiple 3D grid nodes as network nodes and multiple key features of each 3D grid node as node features, the edge weights between network nodes are obtained through a Gaussian kernel function based on the 3D spatial distance between different 3D grid nodes. The network nodes, node features, and edge weights are then input into a graph neural network to construct a real-time global wind field feature map of the target area.

[0011] The global wind field feature map within the current set time period is input into the pre-trained wind field prediction model to obtain the three-dimensional wind field prediction value of the target area for the next set time period.

[0012] Preferably, the UAV flight parameters include scalar airspeed, flight attitude angle, and ground speed vector; the real-time inversion of the three-dimensional wind speed vector based on the UAV flight parameters specifically includes the following steps:

[0013] Construct the rotation matrix from the body coordinate system to the inertial coordinate system based on the flight attitude angle;

[0014] Input the scalar airspeed at each position into the rotation matrix to obtain the airspeed vector in the inertial coordinate system;

[0015] The difference between the ground speed vector and the air speed vector is used to obtain the real-time three-dimensional wind speed vector at each location.

[0016] Preferably, the airspeed vector in the inertial coordinate system is as follows:

[0017] ;

[0018] In the formula, The airspeed vector in the inertial coordinate system. For rotation matrix, The airspeed vector in the body coordinate system. b Refers to the body coordinate system. i Refers to the inertial coordinate system;

[0019] The specific three-dimensional wind speed vector is shown below:

[0020] ;

[0021] In the formula, , and These represent the eastward, northward, and vertical components of airspeed in the inertial coordinate system. , and These represent the eastward, northward, and vertical components of the three-dimensional wind speed vector in the inertial coordinate system. , and These are the eastward, northward, and vertical components of the ground velocity.

[0022] Preferably, the UAV flight parameters also include latitude, longitude, and altitude; feature extraction is performed on the real-time three-dimensional wind speed vector at each location to obtain multiple key features for characterizing wind field uncertainty, specifically including the following steps:

[0023] The latitude and longitude are projected onto the east and north coordinates in the inertial coordinate system, and the altitude is used as the vertical coordinate in the inertial coordinate system.

[0024] The horizontal wind speed gradient is obtained by using the eastward and northward components of the three-dimensional wind speed vector in the inertial coordinate system, as well as the eastward and northward coordinates.

[0025] Vertical wind shear is obtained by using the vertical component and vertical coordinate of the three-dimensional wind speed vector in the inertial coordinate system.

[0026] Turbulence intensity is obtained by the eastward component of the three-dimensional wind speed vector in the inertial coordinate system;

[0027] The rate of change of wind speed over time is obtained by using the modulus of the three-dimensional wind speed vector.

[0028] Preferably, the wind field prediction model is a CNN-LSTM model, including CNN layers and LSTM layers; the step of inputting the global wind field feature map within the current set time period into the pre-trained wind field prediction model to obtain the three-dimensional wind field prediction value for each location in the future set time period includes the following steps:

[0029] Acquire theoretical models of atmospheric turbulence and historical wind field data for the target area;

[0030] The CNN layer extracts features from multiple global wind field feature maps within the current time period to obtain multiple corresponding three-dimensional spatial features. During the feature extraction process, the physical constraint parameters of the atmospheric turbulence theoretical model are used as regularization terms and embedded into the weight update process of the convolution kernel.

[0031] The LSTM layer captures multiple three-dimensional spatial features in a temporal sequence to obtain the predicted three-dimensional wind field value for a future time period. During the temporal capture process, the forget gate weights of the LSTM unit are adjusted using historical wind field data.

[0032] Preferably, the preset rules specifically include:

[0033] The UAV collects its own position coordinates in real time and obtains its actual distance to the coordinates of all three-dimensional grid nodes. Based on the actual distance, it matches the nearest three-dimensional grid node that is within its perception domain, and the UAV's multiple key features are preferentially assigned to that three-dimensional grid node.

[0034] If the UAV is located in the overlapping area of ​​the perception domain of multiple 3D grid nodes, obtain the 3D spatial distance from the UAV to each overlapping 3D grid node, allocate data weights according to the inverse distance weighting method, and distribute multiple key features to each overlapping 3D grid node according to the weights.

[0035] Preferably, the wind field prediction model parameters are updated using adaptive Kalman filtering.

[0036] Preferably, the flight parameters of the UAV at each location are synchronized in time using Kalman filtering.

[0037] Secondly, the present invention provides a wind field prediction system based on UAV swarm collaboration, comprising:

[0038] The deployment module is used to divide the target area into multiple three-dimensional grid nodes and corresponding perception domains, and deploy a drone swarm in the target area. The drones fly in multiple perception domains according to preset trajectories.

[0039] The extraction module is used to collect the flight parameters of each UAV in real time, and to invert the three-dimensional wind speed vector of the area where the UAV is located in real time based on the UAV flight parameters; to extract features from each three-dimensional wind speed vector to obtain multiple key features for characterizing the uncertainty of the wind field; and to map the multiple key features corresponding to each UAV to the corresponding three-dimensional grid nodes according to preset rules.

[0040] The module is used to construct a real-time global wind field feature map of the target area by using multiple 3D mesh nodes as network nodes and multiple key features of each 3D mesh node as node features. Based on the 3D spatial distance between different 3D mesh nodes, the edge weights between network nodes are obtained through a Gaussian kernel function. The network nodes, node features and edge weights are then input into the graph neural network.

[0041] The prediction module is used to input the global wind field feature map within the current set time period into the pre-trained wind field prediction model to obtain the three-dimensional wind field prediction value of the target area in the future set time period.

[0042] Compared with the prior art, the above-mentioned at least one technical solution adopted by the present invention can achieve the following beneficial effects:

[0043] This invention uses multiple 3D grid nodes in the target area as network nodes, mapping multiple key features corresponding to each UAV to the corresponding 3D grid nodes according to preset rules, so that each 3D grid node has sensing data at different times. The key features of each 3D grid node are used as node features, and the edge weights between nodes corresponding to different UAVs are obtained. The network nodes, node features, and edge weights are input into a graph neural network to obtain a real-time global wind field feature map of the target area. Based on the current global wind field feature map, the future 3D wind field prediction value of the target area is obtained. The global wind field feature map has a spatial dimension corresponding to the target area and a channel dimension corresponding to the wind field features, forming a collaborative sensing network covering a wide low-altitude area, while effectively capturing the real-time changes of various wind field features, thus improving the accuracy of wind field prediction results for the target area. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a flowchart of a wind field prediction method based on UAV swarm collaboration according to the present invention.

[0046] Figure 2 This is a flowchart of a wind field prediction method based on UAV swarm collaboration according to the present invention. Detailed Implementation

[0047] 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.

[0048] This invention proposes a wind field perception and prediction method based on UAV swarm collaboration. It aims to significantly improve the efficiency and accuracy of wide-area wind field perception and prediction by leveraging the collaborative perception capabilities of UAV swarms and intelligent algorithms that fuse multi-source data, thereby meeting the flight environment requirements of UAVs in various low-altitude economic application scenarios and mission demands. (Refer to...) Figure 1 and Figure 2 Specifically, it includes the following steps:

[0049] S1: Cluster drone wind speed and conventional airborne data acquisition.

[0050] The target area is divided into multiple three-dimensional grid nodes and corresponding perception domains. A drone swarm is deployed in the target area, and the drones fly in multiple perception domains according to preset trajectories.

[0051] Distributed Deployment: A collaborative sensing cluster of 5-50 medium-sized UAVs is deployed in the target area. The target monitoring area is divided into a 100m×100m×20m three-dimensional grid. The horizontal distance between adjacent UAV nodes is controlled within 300m, and the vertical height layer covers 0-500m (divided into 25 height layers at 20m intervals), ensuring a spatial data overlap rate of ≥20% between adjacent nodes, achieving comprehensive wind field sensing of the target area. Each UAV flies along a preset trajectory, forming a distributed communication chain through a self-organizing network (using 802.11p protocol + TDMA scheduling, communication radius ≥600m). One UAV is designated as the "master node," and the rest are "slave nodes." The master node is responsible for synchronizing the clock reference, issuing trajectory commands, and data aggregation. Slave nodes perform local data acquisition and preprocessing, and upload raw data to the master node in real time through the self-organizing network, forming a "master-slave collaboration + distributed backup" communication architecture.

[0052] Parameter acquisition: Real-time acquisition of scalar airspeed by each UAV ( ), flight attitude angle (roll angle) Pitch angle Yaw angle ), ground speed ( V g including the eastward North ,vertical (Components) and location (latitude / longitude and altitude) h ).

[0053] Scalar airspeed (the speed of the UAV relative to the air) is collected at a frequency of 100Hz using an ultrasonic airspeed meter and serves as the core reference parameter for wind field inversion.

[0054] Ground speed is collected via a GPS (Global Positioning System) module at a frequency of 10Hz or via a nine-axis IMU at a frequency of 100Hz, and is used to calculate the UAV's speed in the inertial coordinate system.

[0055] Flight attitude angles are acquired at a frequency of 100Hz using a nine-axis IMU (Inertial Measurement Unit) and used as the attitude reference for coordinate system transformation.

[0056] Location data was collected at a frequency of 10Hz using a dual-frequency GPS module, serving as the basis for UTM (Universal Transverse Mercator) coordinate mapping and elevation correction.

[0057] Air pressure is collected at a frequency of 50Hz using a MEMS air pressure sensor as auxiliary data for altitude correction.

[0058] S2: Preprocess the collected parameters.

[0059] Preprocessing includes time synchronization and parameter correction.

[0060] S21: Time Synchronization: Clock synchronization is achieved using a Kalman filter, with a state vector... , Due to time deviation, (for frequency drift), where for k Time-state vector; Process equation: ( A Here is the state transition matrix. For process noise, variance 10. -8 ); Observation equation: ( This is the measured time deviation. To account for observation noise, variance 10. -6 The state vector is sensor data.

[0061] A Kalman filter update is performed every 10ms to dynamically correct the slave node clock, ensuring that the timestamp synchronization error of all sensor data in the entire cluster is ≤1ms, and finally outputting the time-synchronized scalar airspeed. V a ), three-dimensional ground velocity ( ) and flight attitude parameters ( / / ).

[0062] S22: Three-dimensional ground velocity ( Solution: By fusing GPS ground speed and IMU data through extended Kalman filter (EKF) to eliminate noise interference, a high-precision three-dimensional ground speed component is obtained.

[0063] EKF fusion model: State equation: ( for k Speed ​​of time The acceleration calculated by the IMU. The sampling period is (for process noise); observation equation: ( This is the actual ground speed measured by GPS. (for observation noise).

[0064] The raw data is input into the EKF fusion model to obtain the output results. The raw data includes the ground velocity vector output from GPS (in geodetic coordinates, including the original values ​​of the east, north, and vertical components) and the three-axis acceleration and angular velocity data output from the IMU. The output results include the calculated data in the inertial coordinate system. V g including the eastward direction North ,vertical The component velocity measurement error is ≤0.1m / s.

[0065] S23: Flight attitude parameters ( / / Calibration involves performing multi-source calibration on the raw attitude data acquired by the IMU to ensure parameter accuracy.

[0066] Static calibration: Roll angle is calibrated using a water platform before flight. ) and pitch angle ( Zero bias is achieved to eliminate sensor installation errors; the yaw angle is calibrated using a magnetic compass. To address the magnetic interference error, a magnetic declination correction model was established.

[0067] Dynamic calibration: During flight, attitude drift is corrected in real time by fusing GPS heading and IMU attitude data. A complementary filtering algorithm (cutoff frequency 10Hz) is used to filter out high-frequency noise, and the final output is the calibrated roll angle. ), pitch angle ( ), yaw angle ( ), attitude angle error ≤ 0.5°.

[0068] S24: UTM coordinate system mapping: Determine the UTM projection zone based on the longitude of the target area, and convert latitude and longitude to eastward using the UTM projection formula. x ), North ( y Planar coordinates, with altitude as the vertical coordinate. z Output the position coordinates in the UTM coordinate system. x / y / z (Accuracy ≤ 1m).

[0069] Output: Scalar airspeed after time synchronization ( V a ), three-dimensional ground velocity ( ), flight attitude parameters ( / / ), and the position coordinates in the UTM coordinate system ( x Eastward y North z vertical).

[0070] S3: Inversion from scalar air velocity to three-dimensional vector wind field.

[0071] Coordinate system definition: Body coordinate system ( b Tie): x The axis moves forward along the nose of the drone. y Axis to the right, z Axial direction downwards (right-hand rule);

[0072] Inertial coordinate system ( i System, UTM): x East-facing axis y Axis north, z The axis points vertically upwards (physical meaning: perpendicular to the horizontal plane of the earth, upwards is positive, corresponding to the direction of altitude). z = h (unit: m).

[0073] Attitude transformation matrix: the rotation matrix from the body coordinate system to the inertial coordinate system. Derived through three rotations (yaw → pitch → roll):

[0074] ;

[0075] in:

[0076] ;

[0077] ;

[0078] ;

[0079] Airspeed vector conversion: airspeed scalar Along the body x In the axial direction, the airspeed vector in the body coordinate system is Transforming to the inertial coordinate system yields:

[0080] ;

[0081] Three-dimensional wind speed vector calculation: Wind speed vector in inertial coordinate system ( u Eastward v North w Vertically upward) equals the ground velocity vector Subtract airspeed vector :

[0082] ;

[0083] In the formula, , , These represent the eastward, northward, and vertical components of airspeed in the inertial coordinate system.

[0084] S4: Feature extraction and global wind field construction.

[0085] Key features are extracted from the three-dimensional wind speed vector, including horizontal wind speed gradient, vertical wind shear, turbulence intensity, and wind speed time change rate.

[0086] Horizontal wind speed gradient (rate of change of wind field in the east-north plane) :

[0087] ;

[0088] In the formula, x , y The partial derivatives are calculated using the first-order difference method, given the east and north coordinates of the UTM coordinate system (obtained by latitude and longitude conversion).

[0089] Vertical wind shear (rate of change of wind speed in the z-direction with height):

[0090] ;

[0091] In the formula, , The first k Vertical wind speed and altitude at each sampling point.

[0092] Turbulence intensity (degree of wind speed fluctuation) :

[0093] ;

[0094] In the formula, The average easterly wind speed over time. N The number of sampling points. This represents the standard deviation of wind speed.

[0095] Wind speed change over time (time dynamics of the wind field):

[0096] ;

[0097] In the formula, (100Hz sampling period) W The modulus of the three-dimensional wind speed vector. t For time.

[0098] Based on the three-dimensional wind speed inversion above, four core features (horizontal wind speed gradient, vertical wind shear, turbulence intensity, and wind speed time change rate) are calculated. A fifth-order Butterworth low-pass filter (cutoff frequency 3Hz) is used to filter out noise, and the original data is compressed into a feature vector of <200B / time, which is then transmitted to the central node through a self-organizing network.

[0099] Multiple key features corresponding to each UAV are mapped to the corresponding 3D mesh nodes according to preset rules.

[0100] Perception domain matching rules: The UAV collects its own position coordinates in real time, calculates the distance between itself and the coordinates of all 3D grid nodes, and matches it to the nearest 3D grid node that is within its perception domain. This node is the UAV's primary assigned grid node, and the UAV's perception data is preferentially assigned to this node.

[0101] Overlapping area data weighting rule: If the UAV is located in the overlapping area of ​​the perception domain of multiple 3D grid nodes, calculate the 3D spatial distance from the UAV to each overlapping grid node, and allocate data weights according to the inverse distance weighting method (the weight is inversely proportional to the distance, and the sum of the weights is 1). The perception data is then allocated to each overlapping grid node according to the weights.

[0102] Empty node data completion rule: If no drone enters the perception domain of a fixed grid node within a set time, the wind field perception data of its adjacent grid nodes are fused by bilinear interpolation to complete the wind field data of the node, ensuring the continuity of wind field data of all grid nodes.

[0103] Global wind field feature map: Using multiple 3D grid nodes as network nodes and multiple key features of each 3D grid node as node features, the edge weights between network nodes are obtained through Gaussian kernel function based on the 3D spatial distance between different 3D grid nodes. The network nodes, node features and edge weights are input into graph neural network to construct a real-time global wind field feature map of the target area.

[0104] Key features (horizontal wind speed gradient, vertical wind shear, turbulence intensity, and wind speed time-varying rate of change) are input into a graph neural network (GNN), with nodes in a 100m×100m×25-layer grid in a UTM coordinate system (25 layers in the z-direction correspond to different altitudes). Edge weights are calculated based on a Gaussian kernel function.

[0105] ;

[0106] In the formula, For nodes i and nodes j Edge weights between them Spatial distance between nodes (including) x , y , z(3D components), outputting a global feature map with dimensions of 100×100×25×16 (16 feature channels corresponding to the above key features and derived quantities).

[0107] S5: Wind field prediction and dynamic updates.

[0108] Model inputs: Global wind field feature map (including x, y, z 3D spatial features, used as real-time sensing input), atmospheric turbulence theoretical model (combined with terrain z-coordinate correction, used as physical constraint input, based on terrain z-coordinate to correct the theoretical thresholds of turbulence intensity and wind speed gradient, ensuring that the prediction results conform to the physical laws of atmospheric motion), historical wind field data ("season-time period-wind field features" time series prior library, used as time series prior input, learning the seasonal and time period periodicity of the wind field during the model pre-training stage, and matching the historical features of the current time period during the prediction stage to correct the deviation of real-time prediction);

[0109] CNN-LSTM prediction model:

[0110] CNN layer: When extracting xyz 3D spatial features, the physical constraint parameters of the atmospheric turbulence theoretical model are used as regularization terms and embedded in the weight update process of the convolution kernel to ensure that the extracted spatial features such as wind speed gradient and vertical wind shear do not exceed the theoretically reasonable range under the corresponding terrain z coordinate; 3 layers of 3D convolution (kernel size 3×3×3, stride 1) extract xyz 3D spatial features and output a 50×50×25×32 tensor;

[0111] LSTM layer: Loads a historical meteorological time series prior library of "seasonal-time period-wind field characteristics", adjusts the forget gate weights of LSTM units, so that the model prioritizes learning the wind field change trend of the corresponding season and time period (such as the wind speed stability from 0-6 in winter), and consolidates the regularity foundation of time series prediction; bidirectional LSTM (64 hidden layer dimensions) captures time series patterns, inputs 10-minute sliding time window data, and outputs the three-dimensional wind field prediction value for the next 10 minutes. .

[0112] Dynamic update: Adaptive Kalman filtering is used to correct model parameters, and the prediction residual is calculated every second.

[0113] ;

[0114] In the formula, for k Predict residuals at all times. To measure the wind speed, To predict wind speed.

[0115] The updated formula is:

[0116] ;

[0117] In the formula, For model parameters, With Kalman gain, the response time to sudden wind fields is less than 2 seconds.

[0118] Example

[0119] This embodiment details the entire process of the system from hardware deployment to outputting prediction results, covering five core stages: initialization, data acquisition, collaborative processing, model computation, and application feedback. The specific steps are as follows:

[0120] I. Experimental Environment and Equipment Configuration.

[0121] (a) Hardware equipment

[0122] Drone swarm: 30 DJI M300RTK quadcopter drones were selected, each with a maximum flight time of 40 minutes and a maximum flight altitude of 500m. The core sensors included are as follows:

[0123] Ultrasonic airspeed meter: CUAVSkye2 model, sampling frequency 100Hz, wind speed measurement accuracy ±0.1m / s, supports low-temperature heating and de-icing function.

[0124] Nine-axis IMU: Model BMI088, sampling frequency 100Hz, roll / pitch / yaw measurement error ≤0.5°.

[0125] Dual-frequency GPS: ubloxF9P model, sampling frequency 10Hz, latitude and longitude positioning accuracy ±1m, supports RTK differential positioning.

[0126] MEMS barometric pressure sensor: Model BMP390, sampling frequency 50Hz, altitude measurement accuracy ≤1m, with temperature compensation correction.

[0127] Communication and computing equipment:

[0128] Cluster communication module: Supports 802.11p protocol and TDMA scheduling, communication radius ≥600m, data transmission latency <50ms, adaptable to low-latency data interaction in mobile Ad-Hoc network scenarios;

[0129] Central node server: configured with an Intel Xeon Gold 6330 processor and an NVIDIA A100 GPU (40GB VRAM) for data aggregation, model inference, and dynamic updates;

[0130] Ground reference equipment: 10 Airmar220WX ultrasonic weather sensors were deployed as true references, with a sampling frequency of 10Hz, wind speed measurement accuracy of ±0.5m / s, and wind direction accuracy of ±2°.

[0131] (II) Software and Data Preparation

[0132] Software environment:

[0133] Unmanned aerial vehicle (UAV) flight control software: Based on secondary development of the PX4 open-source platform, it supports automatic flight of preset tracks and real-time parameter acquisition;

[0134] Model training framework: TensorFlow 2.10, CUDA 11.7 for accelerated computation, using the built-in Adam optimizer (exponential smoothing parameter). μ =0.9, weight decay coefficient λ =1e-5).

[0135] Data processing tools: Python 3.9, integrating NumPy, SciPy (for filtering and feature calculation), and OpenCV (for feature map processing).

[0136] Basic data:

[0137] Topographic and building data: DEM topographic data of the target area (30m resolution), building distribution vector data (including height and outline information), and vegetation roughness parameters (0.15~0.35).

[0138] Historical wind field data: 3D wind field data of the target area for the same period over 3 years, hourly (sourced from local meteorological stations and the National Meteorological Data Center), including multi-dimensional parameters such as wind speed, wind direction, temperature and humidity.

[0139] Atmospheric turbulence theoretical model: constructed based on the Reynolds-averaged Navier-Stokes equation, and the turbulence intensity threshold is corrected by combining the topographic features of the target area.

[0140] II. Implementation Steps of Wind Field Prediction Methods

[0141] Step 1: Cluster Deployment and Data Acquisition

[0142] The target area is divided into 100×100×25 three-dimensional grid nodes at intervals of 100m×100m×20m. 30 UAVs are assigned to sub-regions according to the DARP region partitioning algorithm to ensure that the area difference of each sub-region is ≤15%, the horizontal distance between adjacent UAVs is ≤300m, the vertical height covers 0~500m (25 height layers), and the spatial data overlap rate is ≥20%.

[0143] The drones fly in a grid-like, fully covered flight path (50m spacing between paths, 8m / s flight speed), and data collection is achieved through a master-slave collaborative architecture: one master node drone is responsible for clock synchronization (Kalman filter time synchronization error ≤1ms), command issuance and data aggregation, while the remaining 29 slave node drones perform local parameter collection, including scalar airspeed, flight attitude angle, ground speed vector, latitude and longitude, and altitude.

[0144] The collected data is transmitted to the central node in real time via an 802.11p+TDMA communication link. Before transmission, the data is encapsulated in the format of "timestamp + device ID + parameter type" and the feature vector is compressed to <200B / time to ensure low bandwidth usage.

[0145] Step 2: Data preprocessing and wind speed inversion.

[0146] Time synchronization: Kalman filtering is used to align the time of all data collected by the UAVs, and the state vector... It updates every 10ms with a synchronization error of ≤1ms.

[0147] Ground speed calculation: By fusing GPS and IMU data through extended Kalman filter (EKF), the ground speed vector (eastward, northward, and vertical components) is calculated with a calculation error ≤0.1m / s.

[0148] Wind speed inversion: Based on the flight attitude angle, a rotation matrix is ​​constructed from the body coordinate system to the UTM inertial coordinate system. The scalar airspeed is converted into an airspeed vector in the inertial coordinate system. The real-time three-dimensional wind speed vector of each grid node is obtained by subtracting the ground speed vector from the airspeed vector. u Eastward wind speed, v Northbound wind speed, w (Vertical wind speed).

[0149] Step 3: Feature extraction.

[0150] Noise filtering: A 5th-order Butterworth low-pass filter (cutoff frequency 3Hz) is used for the three-dimensional wind speed vector. This frequency is determined based on the characteristic that the main energy of the low-altitude wind field is concentrated in the low-frequency band, which can effectively filter out high-frequency noise from the sensor and interference from flight vibration.

[0151] Feature calculation:

[0152] Horizontal wind speed gradient: The central difference method is used for calculation.

[0153] Vertical wind shear: ∂w / ∂z, calculated based on the wind speed difference in the same vertical column of grid.

[0154] Turbulence intensity: N=1000 (number of sampling points per 10 seconds). The wind speed is the time average of eastward winds.

[0155] Wind speed change over time: Δt=0.01s (corresponding to a 100Hz sampling period), W is the magnitude of the three-dimensional wind speed vector.

[0156] Feature vector generation: by format [timestamp, x (Eastward coordinates in UTM inertial coordinate system) y (North coordinates in UTM inertial coordinate system) z (UTM inertial coordinate system vertical coordinates), wind speed amplitude, horizontal wind speed gradient, vertical wind shear, turbulence intensity, wind speed time change rate] generate a standardized feature vector and transmit it to the central node.

[0157] Step 4: Global wind field feature fusion.

[0158] The feature vectors are input into a graph neural network (GNN), with nodes in a 100m×100m×20m 3D grid, and processed using a Gaussian kernel function. Calculate the edge weights between nodes ( (where i is the three-dimensional spatial distance between nodes i and j), and after fusing the local features of neighboring nodes, a global wind field feature map with dimensions of 100×100×25×16 is output (16 channels correspond to core features and spatial correlation derivatives).

[0159] Step 5: Multi-source fusion wind field prediction.

[0160] Multi-source input preprocessing:

[0161] Atmospheric turbulence theoretical model correction: Based on DEM topographic data and building height, the turbulence intensity threshold was corrected. The turbulence intensity threshold was increased by 20% in areas with building height > 30m, and in valley topography ( z (Coordinate gradient > 5°) is corrected to 1.2 to 1.5 times that of the plain area.

[0162] Historical wind field data clustering: Clustering is performed by "season-time period" (12-18:00 in summer is the target cluster), and the historical wind field characteristics of the cluster are extracted (average turbulence intensity 0.18, average northerly wind speed 3.2m / s) to generate a time-series prior library;

[0163] Model training and inference:

[0164] Training data partitioning: 3 years of historical data were partitioned into training set (864,000 sets), validation set (108,000 sets), and test set (108,000 sets) in a ratio of 8:1:1. A 10-minute sliding window (5-minute step) was used to expand the sample.

[0165] Model configuration: 3 layers of 3D convolutional layers (kernel size 3×3×3, stride 1, activation function ReLU, number of output channels 16, 32, 64) + 1 layer of bidirectional LSTM layer (hidden layer dimension 64, dropout rate 0.2) + fully connected layer;

[0166] Training parameters: initial learning rate 0.001, cosine annealing decay strategy, batch size 32, 100 training epochs, early stopping mechanism (stop if the validation set loss does not decrease for 10 consecutive epochs).

[0167] Multi-source fusion inference: Input the global wind field feature map, the corrected atmospheric turbulence theoretical model, and the historical wind field data of the target cluster into the pre-trained model, and output the three-dimensional wind field prediction value for the next 10 minutes.

[0168] Step 6: Dynamic update.

[0169] An adaptive Kalman filter is used to update the model parameters every second, and the prediction residuals are calculated based on the measurement update equation of the Kalman filter. ( The wind speed is measured by a ground-based reference sensor. (To predict wind speed for the model), through parameter update formulas ( The model weights are dynamically adjusted (using Kalman gain) so that the response time to sudden wind fields is less than 2 seconds.

[0170] This invention achieves efficient perception and accurate prediction of wide-area wind fields through a distributed perception framework of UAV swarms, combined with multi-source data fusion and deep learning technology.

[0171] (1) Construct a wide-area high-resolution dynamic sensing network to overcome the limitations of single-point sensing.

[0172] By distributing heterogeneous sensors (such as wind speed sensors, barometric pressure sensors, and lidar) in a drone swarm, each node autonomously collects local wind field data and extracts key features (such as wind speed amplitude, gradient, and frequency). This information is shared in real time through a swarm communication network, forming a collaborative sensing network covering a wide low-altitude area. Compared to traditional single-point or single-platform sensing, this method expands the wind field sensing range by 5-10 times, improves spatiotemporal resolution by more than 30%, effectively captures microscale meteorological changes (such as sudden anomalies like turbulence and downbursts), and solves the problem that existing technologies cannot cover dynamic wind fields with localized sensing.

[0173] (2) Integrate multi-source data with deep learning algorithms to improve the accuracy and timeliness of wind field prediction.

[0174] This system integrates atmospheric turbulence theoretical models (including physical features such as topography, vegetation, and buildings), historical meteorological data (such as multi-year wind speed and air pressure time-series data), and real-time UAV perception data. It uses a CNN-LSTM hybrid neural network to fuse the spatiotemporal features of multi-source data: CNN extracts spatial distribution features influenced by topography, while LSTM captures the correlation of wind field time series. Combined with dynamic update mechanisms such as adaptive Kalman filtering, it absorbs new data in real time to correct the prediction model, reducing wind speed prediction errors by 20%-25% and shortening the response time to sudden wind field events to the second level. This solves the problems of insufficient prediction accuracy and lag behind actual wind field changes in traditional single-data-source models.

[0175] (3) Design a distributed collaboration and dynamic update mechanism to enhance the system's environmental adaptability.

[0176] Through lightweight communication protocols and a distributed processing architecture, efficient data sharing and collaborative processing among UAV nodes are achieved, with the central node or distributed network fusing global wind field feature maps in real time. For different scenarios such as urban areas, mountainous regions, and oceans, transfer learning or parameter fine-tuning enables the system to quickly adapt to wind field characteristics in specific environments (such as the nonlinear effects of urban canyon winds and the periodic fluctuations of mountain airflow), avoiding prediction biases in complex terrains using traditional fixed models. Simultaneously, the modular design supports flexible expansion of sensor payloads and computing units, allowing the cluster size to seamlessly scale from 5 UAVs to hundreds, adapting to the different mission requirements for sensing range and accuracy, and addressing the issues of weak collaborative capabilities and poor scenario adaptability in existing technologies.

[0177] (4) Establish a closed loop of "perception-prediction-correction" to ensure the safety and efficiency of UAV flight.

[0178] By driving dynamic model updates through real-time sensing data, a closed loop of "prediction-feedback-optimization" is formed, ensuring that wind field prediction results reflect current environmental changes in real time. For example, in tasks such as drone logistics and emergency rescue, the system can predict the wind field distribution along the path 10 minutes in advance, assisting drones in optimizing their flight trajectories, reducing flight path deviations or increased energy consumption caused by wind shear and turbulence, improving flight safety by more than 25% in complex wind field environments, and increasing mission execution efficiency by 15%, providing core technological support for the low-altitude economy.

[0179] Improve the range and accuracy of wind field sensing:

[0180] (1) Distributed sensing framework: Through the distributed sensing of the UAV cluster, each UAV node can independently sense the real-time wind field information of its own location, and realize information sharing and collaborative processing through cluster communication, which significantly expands the spatiotemporal resolution of wide-area wind field sensing.

[0181] (2) Multi-source and multi-modal data fusion: By integrating atmospheric turbulence theoretical model data, historical meteorological data and real-time distributed perception data of UAV clusters, the data is integrated and processed through deep learning methods, which further improves the accuracy and reliability of wind field perception.

[0182] Enhance wind field forecasting capabilities:

[0183] (1) Dynamic update mechanism: The introduction of dynamic update mechanism enables the wind field prediction benchmark model to absorb real-time sensing data and correct and adjust the prediction results, which significantly improves the timeliness and accuracy of wind field prediction.

[0184] (2) Deep learning model: Using deep learning models (such as convolutional neural networks, long short-term memory networks, etc.) to extract and fuse features from multi-source data can better capture the spatiotemporal variation of wind fields and improve prediction accuracy.

[0185] Improve system scalability:

[0186] (1) Modular design: The modular design is adopted to make the function of each node modular, which can realize the allocation and use of different types of sensor data of the node, and facilitates expansion and maintenance.

[0187] (2) Distributed management: Through the distributed management mechanism, resource allocation is dynamically adjusted to adapt to clusters of different sizes.

[0188] Adaptable to complex weather conditions:

[0189] (1) Sudden wind field response: Through real-time distributed wind field perception and prediction, it can effectively respond to sudden changes in low-altitude wind fields, such as gusts and downbursts, and significantly improve the flight safety and stability of UAVs under complex weather conditions.

[0190] (2) Multi-factor influence: Taking into account the influence of various factors such as topography, vegetation, buildings, atmospheric pressure and temperature on the wind field environment, the UAV can better adapt to the complex and ever-changing low-altitude flight environment.

[0191] Improve task execution efficiency:

[0192] (1) Path optimization: By real-time wind field perception and prediction, the flight path of the UAV is optimized to reduce energy consumption and improve mission execution efficiency.

[0193] (2) Task adaptability: Design a task-driven perception strategy based on specific task requirements to ensure that the perception results can meet the task requirements and improve the system's flexibility and user satisfaction.

[0194] Based on the same concept, the present invention also provides a wind field prediction system based on UAV swarm collaboration, including a deployment module, a data acquisition module, an extraction module and a prediction module.

[0195] The deployment module is used to divide the target area into multiple three-dimensional grid nodes and corresponding perception domains, and deploy a drone swarm in the target area. The drones fly in multiple perception domains according to preset trajectories.

[0196] The extraction module is used to collect the flight parameters of each UAV in real time, and to invert the three-dimensional wind speed vector of the area where the UAV is located in real time based on the UAV flight parameters; to extract features from each three-dimensional wind speed vector to obtain multiple key features for characterizing the uncertainty of the wind field; and to map the multiple key features corresponding to each UAV to the corresponding three-dimensional grid nodes according to preset rules.

[0197] The construction module is used to construct a real-time global wind field feature map of the target area by using multiple 3D grid nodes as network nodes and multiple key features of each 3D grid node as node features. Based on the 3D spatial distance between different 3D grid nodes, the edge weights between network nodes are obtained through a Gaussian kernel function. The network nodes, node features and edge weights are then input into a graph neural network.

[0198] The prediction module is used to input the global wind field feature map within the current set time period into the pre-trained wind field prediction model to obtain the three-dimensional wind field prediction value of the target area in the future set time period.

[0199] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0200] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A wind field prediction method based on UAV swarm collaboration, characterized in that, Includes the following steps: The target area is divided into multiple three-dimensional grid nodes and corresponding perception domains. A drone swarm is deployed in the target area, and the drones fly in multiple perception domains according to a preset trajectory. Real-time acquisition of flight parameters for each drone; real-time inversion of the three-dimensional wind speed vector of the area where the drone is currently located based on the drone flight parameters; Feature extraction is performed on each three-dimensional wind speed vector to obtain several key features used to characterize wind field uncertainty; Map multiple key features corresponding to each UAV to the corresponding 3D mesh nodes according to preset rules; Using multiple 3D grid nodes as network nodes and multiple key features of each 3D grid node as node features, the edge weights between network nodes are obtained through a Gaussian kernel function based on the 3D spatial distance between different 3D grid nodes. The network nodes, node features, and edge weights are then input into a graph neural network to construct a real-time global wind field feature map of the target area. Input the global wind field feature map within the current set time period into the pre-trained wind field prediction model to obtain the three-dimensional wind field prediction value of the target area in the future set time period. The wind field prediction model is a CNN-LSTM model, including CNN layers and LSTM layers; the step of inputting the global wind field feature map within the current set time period into the pre-trained wind field prediction model to obtain the three-dimensional wind field prediction value for each location in the future set time period includes the following steps: Acquire theoretical models of atmospheric turbulence and historical wind field data for the target area; The CNN layer extracts features from multiple global wind field feature maps within the current time period to obtain multiple corresponding three-dimensional spatial features. During the feature extraction process, the physical constraint parameters of the atmospheric turbulence theoretical model are used as regularization terms and embedded into the weight update process of the convolution kernel. The LSTM layer captures multiple three-dimensional spatial features in a temporal sequence to obtain the predicted three-dimensional wind field value for a future time period. During the temporal capture process, the forget gate weights of the LSTM unit are adjusted using historical wind field data.

2. The wind field prediction method based on UAV swarm collaboration as described in claim 1, characterized in that, The UAV flight parameters include scalar airspeed, flight attitude angle, and ground speed vector; the real-time inversion of the three-dimensional wind speed vector based on the UAV flight parameters specifically includes the following steps: Construct the rotation matrix from the body coordinate system to the inertial coordinate system based on the flight attitude angle; Input the scalar airspeed at each position into the rotation matrix to obtain the airspeed vector in the inertial coordinate system; The difference between the ground speed vector and the air speed vector is used to obtain the real-time three-dimensional wind speed vector at each location.

3. The wind field prediction method based on UAV swarm collaboration as described in claim 2, characterized in that, The airspeed vector in the inertial coordinate system is shown below: ； In the formula, The airspeed vector in the inertial coordinate system. For rotation matrix, The airspeed vector in the body coordinate system. b Refers to the body coordinate system. i Refers to the inertial coordinate system; The specific three-dimensional wind speed vector is shown below: ； In the formula, , and These represent the eastward, northward, and vertical components of airspeed in the inertial coordinate system. , and These represent the eastward, northward, and vertical components of the three-dimensional wind speed vector in the inertial coordinate system. , and These are the eastward, northward, and vertical components of the ground velocity.

4. The wind field prediction method based on UAV swarm collaboration as described in claim 2, characterized in that, The UAV flight parameters also include latitude, longitude, and altitude; feature extraction is performed on the real-time three-dimensional wind speed vector at each location to obtain several key features for characterizing wind field uncertainty, specifically including the following steps: The latitude and longitude are projected onto the east and north coordinates in the inertial coordinate system, and the altitude is used as the vertical coordinate in the inertial coordinate system. The horizontal wind speed gradient is obtained by using the eastward and northward components of the three-dimensional wind speed vector in the inertial coordinate system, as well as the eastward and northward coordinates. Vertical wind shear is obtained by using the vertical component and vertical coordinate of the three-dimensional wind speed vector in the inertial coordinate system. Turbulence intensity is obtained by the eastward component of the three-dimensional wind speed vector in the inertial coordinate system; The rate of change of wind speed over time is obtained by using the modulus of the three-dimensional wind speed vector.

5. The wind field prediction method based on UAV swarm collaboration as described in claim 1, characterized in that, The preset rules specifically include: The UAV collects its own position coordinates in real time and obtains its actual distance to the coordinates of all three-dimensional grid nodes. Based on the actual distance, it matches the nearest three-dimensional grid node that is within its perception domain, and the UAV's multiple key features are preferentially assigned to that three-dimensional grid node. If the UAV is located in the overlapping area of ​​the perception domain of multiple 3D grid nodes, obtain the 3D spatial distance from the UAV to each overlapping 3D grid node, allocate data weights according to the inverse distance weighting method, and distribute multiple key features to each overlapping 3D grid node according to the weights.

6. The wind field prediction method based on UAV swarm collaboration as described in claim 1, characterized in that, The parameters of the wind field prediction model are updated using adaptive Kalman filtering.

7. The wind field prediction method based on UAV swarm collaboration as described in claim 1, characterized in that, The flight parameters of the UAV at each location are synchronized in time using Kalman filtering.

8. A prediction system for a wind field prediction method based on UAV swarm collaboration according to claim 1, characterized in that, include: The deployment module is used to divide the target area into multiple three-dimensional grid nodes and corresponding perception domains, and deploy a drone swarm in the target area. The drones fly in multiple perception domains according to preset trajectories. The extraction module is used to collect the flight parameters of each UAV in real time and to invert the three-dimensional wind speed vector of the area where the UAV is currently located in real time based on the UAV flight parameters; Feature extraction is performed on each three-dimensional wind speed vector to obtain several key features used to characterize wind field uncertainty; Map multiple key features corresponding to each UAV to the corresponding 3D mesh nodes according to preset rules; The module is used to construct a real-time global wind field feature map of the target area by using multiple 3D mesh nodes as network nodes and multiple key features of each 3D mesh node as node features. Based on the 3D spatial distance between different 3D mesh nodes, the edge weights between network nodes are obtained through a Gaussian kernel function. The network nodes, node features and edge weights are then input into the graph neural network. The prediction module is used to input the global wind field feature map within the current set time period into the pre-trained wind field prediction model to obtain the three-dimensional wind field prediction value of the target area in the future set time period.

Images