Laser radar adaptive volume scanning strategy generation method for complex mountainous wind field

Abstract

The present application relates to the technical field of laser radar adaptive volume scanning, and particularly relates to a laser radar adaptive volume scanning strategy generation method for complex mountain wind fields. In view of the problem that the traditional fixed volume scanning strategy is difficult to adapt to the spatiotemporal heterogeneity of mountain flow fields, a three-dimensional digital scene model is constructed based on the terrain data of the survey area, a three-dimensional flow field database is generated through CFD numerical simulation, a virtual laser radar model is constructed and the flow field data is mapped to generate a simulation observation data set, a volume scanning strategy space is constructed and a candidate strategy set is generated, a multi-strategy simulation detection and wind field reconstruction are performed, the performance of each strategy is evaluated based on true value comparison and a ranking table is generated, the terrain and incoming flow characteristics are extracted to train an adaptive matching model, and the measured parameters are input into the model to output the optimal volume scanning strategy. The present application realizes dynamic adaptive matching of the scanning mode and the spatiotemporal characteristics of complex mountain flow fields, and improves the integrity, physical consistency and timeliness of wind field perception.

Description

Technical Field

[0001] This invention relates to the field of adaptive volume scanning technology for lidar, and in particular to a method for generating adaptive volume scanning strategies for lidar in complex mountainous wind fields. Background Technology

[0002] In engineering and scientific fields such as wind energy resource assessment, wind farm micro-site selection, turbine operation control, and early warning of meteorological disasters in mountainous areas, accurately perceiving and predicting the spatiotemporal distribution characteristics of wind fields under complex mountainous terrain is a key prerequisite for improving wind resource utilization and ensuring meteorological safety. However, the dramatic undulations and spatial variations in surface roughness of mountainous terrain cause complex flow phenomena such as flow separation, local acceleration, backflow vortices, and wake interference when airflow passes through ridges, steep slopes, and canyons, resulting in strong non-uniformity, unsteadiness, and spatial heterogeneity in the wind field. These flow characteristics not only make wind field modeling extremely difficult but also pose a severe challenge to its short-term (e.g., 0–6 hours) dynamic prediction.

[0003] Currently, the monitoring and prediction of wind fields in complex terrain mainly relies on two types of technical approaches: Field observation methods: On-site measurements are conducted using sparsely distributed ground-based meteorological stations, wind towers, or Doppler lidar. These methods can acquire high-precision time-series data on wind speed and direction at local locations, but are limited by equipment procurement and maintenance costs, terrain accessibility, and power and communication conditions, making it difficult to form a large-scale, high-spatial-resolution monitoring network. Even with scanning lidar, its fixed-volume scanning strategies (such as PPI and RHI) are difficult to adapt to the spatially heterogeneous characteristics of flow structures in complex terrain, easily missing key flow areas (such as separation zones and wake zones), resulting in incomplete overall flow field perception.

[0004] Numerical simulation methods employ computational fluid dynamics (CFD) or mesoscale meteorological models (such as WRF) to numerically solve the global wind field. These methods can provide complete spatial flow field information and support the construction of wind resource maps under complex terrain. However, their drawbacks are also obvious: they are highly sensitive to the accuracy of boundary conditions and terrain representation, consume large amounts of computational resources, have long simulation times, and are difficult to update or assimilate in real time. Therefore, in short-term dynamic prediction scenarios, they face the fundamental contradiction of "difficulty in balancing accuracy and timeliness".

[0005] To address the shortcomings of the aforementioned methods, researchers have recently attempted to introduce artificial intelligence (AI) technology to train wind speed prediction models using historical observation data. However, most existing AI models focus on single-site wind speed time series prediction, lacking physical modeling of topographic constraints and spatial correlations. These models exhibit poor generalization ability, making it difficult to achieve spatial reconstruction of the entire wind field from sparse observation points. More critically, current technological systems generally treat observation, simulation, and prediction as separate processes—observational data is not effectively assimilated for simulation calibration, and simulation results are not used to guide the optimization of observation strategies, resulting in a fragmented approach that fails to leverage the synergistic advantages of multi-source information.

[0006] In summary, existing technologies for sensing and predicting complex mountain wind fields suffer from a significant contradiction between spatial coverage integrity, prediction timeliness, and consistency with flow physics. Overcoming the limitations of fixed scanning modes and achieving adaptive matching between lidar volume scanning strategies and the spatiotemporal characteristics of complex flow fields has become a key scientific problem for improving wind field sensing capabilities. Summary of the Invention

[0007] The purpose of this invention is to provide a method for generating adaptive volume scanning strategies for lidar in complex mountainous wind fields. By constructing a high-fidelity virtual wind field environment, the method automatically generates the optimal volume scanning strategy for different terrains and wind field conditions, thereby achieving adaptive matching between lidar scanning modes and the spatiotemporal characteristics of the actual flow field.

[0008] To achieve the above objectives, this invention provides a method for generating adaptive volume scanning strategies for lidar in complex mountainous wind fields, comprising the following steps: S1. Construct a three-dimensional digital scene model based on the digital elevation model data and land cover data of the survey area, perform fluid dynamics numerical simulation on the three-dimensional digital scene model, and generate a three-dimensional flow field database containing velocity components, turbulent kinetic energy and dissipation rate parameters under multiple working conditions. S2. Construct a virtual lidar sensor model, map the three-dimensional flow field database to the sampling points of the virtual lidar sensor, calculate the radial velocity of each sampling point and add noise to generate a simulation observation dataset. S3. Define a volume scan strategy space consisting of scanning mode type, azimuth scanning range, elevation scanning range, azimuth step size, elevation step size, dwell time per point, number of sampling repetitions per point, and region priority weight parameters. Generate a candidate strategy set based on Latin hypercube sampling and prior knowledge screening. S4. Call the virtual lidar sensor model, perform simulation detection experiments on each volume scan strategy in the candidate strategy set, reconstruct the wind field based on the simulation observation dataset, compare the reconstructed wind field with the three-dimensional flow field database, calculate the root mean square error, correlation coefficient, key area coverage and data efficiency, and generate a strategy performance ranking table through multi-objective weighted scoring. S5. Extract the static terrain features and dynamic inflow features corresponding to the three-dimensional flow field database, train the adaptive matching model with the best strategy in the strategy performance ranking table as the label, input the measured terrain and meteorological parameters into the adaptive matching model, and output the optimal volume scanning strategy.

[0009] Preferably, in S1, the step of constructing a three-dimensional digital scene model based on the digital elevation model data and land cover data of the survey area includes: The digital elevation model data of the survey area is resampled to a preset resolution, and the geometric figures of buildings are established by combining the surface cover vector data. Spatial normalization is performed according to a unified geographic coordinate system. The underlying surface is divided into four types of areas according to land use: water surface, grassland, low-density urban area and high-density urban area. Corresponding roughness length parameters are assigned to construct a three-dimensional digital scene model that includes surface undulation, building entities and roughness attributes. The three-dimensional digital scene model is subjected to a hybrid structured and unstructured mesh generation. Local densification is performed around buildings and in areas with abrupt terrain changes. Prism layer meshes are set on the ground and building surfaces to meet the wall function requirements, and the overall mesh growth rate and total mesh volume are controlled. The turbulence model was selected as the core physical model. The discretization scheme of the convection term, the pressure-velocity coupling algorithm and the convergence criterion were set. The logarithmic wind speed profile and turbulence intensity were set at the inlet boundary. The rough wall function was applied to the ground and building walls. Batch numerical solutions were performed for multiple wind directions and multiple wind speeds. Spatial flow field data were extracted and exported as standard format files to construct a three-dimensional flow field database.

[0010] Preferably, in S2, the step of constructing the virtual lidar sensor model includes: Configure the physical parameters of the virtual lidar sensor, including its operating wavelength, pulse energy, pulse repetition frequency, beam divergence angle, range resolution, maximum detection range, and receiving aperture area. An exponential decay model is used to characterize the signal transmission process, and a ray tracing algorithm is used to perform occlusion discrimination. The intersection operation between the radar emitted ray and the three-dimensional digital scene model is performed to mark the occluded sampling points. For each radar attitude and sampling distance, the spatial position of the sampling point is determined by coordinate system transformation. The flow field velocity component is extracted from the three-dimensional flow field database by interpolation algorithm, the radial velocity is calculated and Gaussian white noise is superimposed, and invalid values ​​are assigned to sampling points with obstruction, sampling points with signal-to-noise ratio below the threshold and sampling points exceeding the maximum detection distance, thus generating a simulation observation dataset.

[0011] Preferably, in S3, the step of defining the volume scan strategy space includes: The scanning modes are classified into six standard forms: planar position display mode, distance and height display mode, velocity and azimuth display mode, Doppler beam swing mode, grid scanning mode, and adaptive trajectory mode. Each volume scan strategy is parameterized into an eight-dimensional feature vector consisting of scan mode type, azimuth scan range, elevation scan range, azimuth step size, elevation step size, dwell time per point, number of sampling repetitions per point, and region priority weight. By utilizing the Latin hypercube sampling technique to perform uniform sampling in the eight-dimensional feature vector space, invalid parameter combinations are eliminated based on the physical detection characteristics of lidar, generating a candidate strategy set that is controllable in scale and physically feasible.

[0012] Preferably, in S4, the step of reconstructing the wind field based on the simulation observation dataset includes: The wind field reconstruction algorithm is automatically matched according to the scanning mode type. Among them, the single elevation angle plane position display mode adopts the velocity azimuth display fitting algorithm, the multi elevation angle plane position display mode adopts the three-dimensional variational inversion algorithm, the Doppler beam swing mode adopts the analytical synthesis method based on beam geometry relationship, and the grid scanning mode adopts the Kriging interpolation method. The reconstructed wind field is compared point by point with the three-dimensional flow field database. The root mean square error of the radial velocity is calculated to assess the deviation. The correlation coefficient between the reconstructed value and the true value is calculated to assess the linear reconstruction capability. The proportion of effective sampling points covering the key detection area is calculated to determine the coverage of the key area. The proportion of effective data points to the total sampling points is calculated to determine the data efficiency.

[0013] Preferably, in step S4, the step of generating a strategy performance ranking table through multi-objective weighted scoring includes: The root mean square error, correlation coefficient, key area coverage, and data efficiency are normalized to a preset interval and then summed with weights. For high-precision reconstruction tasks, the weights of root mean square error, key area coverage, time efficiency, and data efficiency are configured as a first preset value combination. For rapid emergency detection tasks, the weights of time efficiency, key area coverage, root mean square error, and data efficiency are configured as a second preset value combination. Based on the weighted summation results, the volume scan strategies in the candidate strategy set are sorted in descending order to generate a strategy performance ranking table containing the values ​​of each indicator and the overall score.

[0014] Preferably, in S5, the steps for training the adaptive matching model include: Static terrain features, dynamic inflow features, and measurement point attribute features are extracted from the 3D flow field database. The volume scan strategy with the highest comprehensive score in the strategy performance ranking table is used as the label to construct a terrain-wind field-strategy mapping training dataset. Offline training is performed using the random forest algorithm. Normalization mapping, categorical variable encoding, and mutual information feature selection are applied to the training features. The number and maximum depth of decision trees are configured. Cross-validation is used to evaluate the generalization error and generate an adaptive matching model. The measured terrain and meteorological parameters are input into the adaptive matching model for inference. The output includes the optimal volume scanning strategy, which includes strategy identifier, mode type, parameter configuration and expected performance indicators, and is accompanied by model confidence score. If the confidence score is lower than the preset threshold, a secondary calibration is triggered or the model is switched to the default scanning mode.

[0015] Preferably, static terrain features include average elevation, elevation standard deviation, roughness distribution, building density, and average building height; dynamic inflow features include wind direction, wind speed, turbulence intensity, and atmospheric stability; and measuring point attribute features include the absolute coordinates of the measuring point, surrounding environmental obstruction information, and the range of accessible viewing angles.

[0016] The advantages and beneficial effects of this invention compared to the prior art are: 1. It breaks through the static limitations of the traditional fixed scanning mode of lidar. By constructing a terrain-wind field-strategy mapping relationship, it achieves dynamic adaptive matching between the volume scanning strategy and the spatiotemporal characteristics of complex mountain flow fields. This allows the detection perspective to focus on high-value areas such as flow separation and backflow vortices, significantly improving the completeness and targeting of key flow structures.

[0017] 2. Ensures physical consistency and high fidelity of the detection strategy. Unlike methods that rely solely on experience or random trial and error, this invention uses numerical simulation results from computational fluid dynamics (CFD) as the basis for the training database. High-precision flow field simulation pre-identifies the physical constraints of terrain on airflow, ensuring that the selection process for the optimal strategy always follows fluid dynamics principles. This effectively avoids detection blind spots caused by inappropriate strategies, thereby guaranteeing the reliability of the wind field reconstruction results in terms of physical mechanisms.

[0018] 3. By adopting an operation mode that combines offline simulation pre-training with online second-level inference, the time-consuming numerical simulation calculations are moved to the model building stage. In the online application stage, only measured terrain and meteorological parameters need to be input to quickly output the optimal strategy. This achieves the optimal balance between high-precision reconstruction and high-time-efficiency response in complex mountain wind field detection, and meets the real-time requirements of short-term dynamic prediction and emergency meteorological early warning.

[0019] 4. By using a unified data structure, parameterized strategy expression, and quantitative multi-objective comprehensive scoring function, the subjective uncertainty brought about by human experience is eliminated, making the method highly standardized and universal. It can be easily adapted to complex terrain scenarios in different geographical locations, reducing the technical threshold for engineering deployment and operation and maintenance.

[0020] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0021] Figure 1 This is a flowchart of a lidar adaptive volume scanning strategy generation method for complex mountain wind fields, as described in an embodiment of the present invention. Figure 2 This is a technical roadmap for a lidar adaptive volume scanning strategy generation method for complex mountain wind fields, according to an embodiment of the present invention. Detailed Implementation

[0022] In the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product is in use. They are used only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," and "connect" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0023] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0024] like Figure 1 As shown, this invention provides a method for generating adaptive volume scanning strategies for lidar in complex mountainous wind fields, including the following steps: S1. Construct a three-dimensional digital scene model based on the digital elevation model data and land cover data of the survey area, perform fluid dynamics numerical simulation on the three-dimensional digital scene model, and generate a three-dimensional flow field database containing velocity components, turbulent kinetic energy and dissipation rate parameters under multiple working conditions. S2. Construct a virtual lidar sensor model, map the three-dimensional flow field database to the sampling points of the virtual lidar sensor, calculate the radial velocity of each sampling point and add noise to generate a simulation observation dataset. S3. Define a volume scan strategy space consisting of scanning mode type, azimuth scanning range, elevation scanning range, azimuth step size, elevation step size, dwell time per point, number of sampling repetitions per point, and region priority weight parameters. Generate a candidate strategy set based on Latin hypercube sampling and prior knowledge screening. S4. Call the virtual lidar sensor model, perform simulation detection experiments on each volume scan strategy in the candidate strategy set, reconstruct the wind field based on the simulation observation dataset, compare the reconstructed wind field with the three-dimensional flow field database, calculate the root mean square error, correlation coefficient, key area coverage and data efficiency, and generate a strategy performance ranking table through multi-objective weighted scoring. S5. Extract the static terrain features and dynamic inflow features corresponding to the three-dimensional flow field database, train the adaptive matching model with the best strategy in the strategy performance ranking table as the label, input the measured terrain and meteorological parameters into the adaptive matching model, and output the optimal volume scanning strategy.

[0025] Preferably, in S1, the step of constructing a three-dimensional digital scene model based on the digital elevation model data and land cover data of the survey area includes: The digital elevation model data of the survey area is resampled to a preset resolution, and the geometric figures of buildings are established by combining the surface cover vector data. Spatial normalization is performed according to a unified geographic coordinate system. The underlying surface is divided into four types of areas according to land use: water surface, grassland, low-density urban area and high-density urban area. Corresponding roughness length parameters are assigned to construct a three-dimensional digital scene model that includes surface undulation, building entities and roughness attributes. The three-dimensional digital scene model is subjected to a hybrid structured and unstructured mesh generation. Local densification is performed around buildings and in areas with abrupt terrain changes. Prism layer meshes are set on the ground and building surfaces to meet the wall function requirements, and the overall mesh growth rate and total mesh volume are controlled. The turbulence model was selected as the core physical model. The discretization scheme of the convection term, the pressure-velocity coupling algorithm and the convergence criterion were set. The logarithmic wind speed profile and turbulence intensity were set at the inlet boundary. The rough wall function was applied to the ground and building walls. Batch numerical solutions were performed for multiple wind directions and multiple wind speeds. Spatial flow field data were extracted and exported as standard format files to construct a three-dimensional flow field database.

[0026] Preferably, in S2, the step of constructing the virtual lidar sensor model includes: Configure the physical parameters of the virtual lidar sensor, including its operating wavelength, pulse energy, pulse repetition frequency, beam divergence angle, range resolution, maximum detection range, and receiving aperture area. An exponential decay model is used to characterize the signal transmission process, and a ray tracing algorithm is used to perform occlusion discrimination. The intersection operation between the radar emitted ray and the three-dimensional digital scene model is performed to mark the occluded sampling points. For each radar attitude and sampling distance, the spatial position of the sampling point is determined by coordinate system transformation. The flow field velocity component is extracted from the three-dimensional flow field database by interpolation algorithm, the radial velocity is calculated and Gaussian white noise is superimposed, and invalid values ​​are assigned to sampling points with obstruction, sampling points with signal-to-noise ratio below the threshold and sampling points exceeding the maximum detection distance, thus generating a simulation observation dataset.

[0027] Preferably, in S3, the step of defining the volume scan strategy space includes: The scanning modes are classified into six standard forms: planar position display mode, distance and height display mode, velocity and azimuth display mode, Doppler beam swing mode, grid scanning mode, and adaptive trajectory mode. Each volume scan strategy is parameterized into an eight-dimensional feature vector consisting of scan mode type, azimuth scan range, elevation scan range, azimuth step size, elevation step size, dwell time per point, number of sampling repetitions per point, and region priority weight. By utilizing the Latin hypercube sampling technique to perform uniform sampling in the eight-dimensional feature vector space, invalid parameter combinations are eliminated based on the physical detection characteristics of lidar, generating a candidate strategy set that is controllable in scale and physically feasible.

[0028] Preferably, in S4, the step of reconstructing the wind field based on the simulation observation dataset includes: The wind field reconstruction algorithm is automatically matched according to the scanning mode type. Among them, the single elevation angle plane position display mode adopts the velocity azimuth display fitting algorithm, the multi elevation angle plane position display mode adopts the three-dimensional variational inversion algorithm, the Doppler beam swing mode adopts the analytical synthesis method based on beam geometry relationship, and the grid scanning mode adopts the Kriging interpolation method. The reconstructed wind field is compared point by point with the three-dimensional flow field database. The root mean square error of the radial velocity is calculated to assess the deviation. The correlation coefficient between the reconstructed value and the true value is calculated to assess the linear reconstruction capability. The proportion of effective sampling points covering the key detection area is calculated to determine the coverage of the key area. The proportion of effective data points to the total sampling points is calculated to determine the data efficiency.

[0029] Preferably, in step S4, the step of generating a strategy performance ranking table through multi-objective weighted scoring includes: The root mean square error, correlation coefficient, key area coverage, and data efficiency are normalized to a preset interval and then summed with weights. For high-precision reconstruction tasks, the weights of root mean square error, key area coverage, time efficiency, and data efficiency are configured as a first preset value combination. For rapid emergency detection tasks, the weights of time efficiency, key area coverage, root mean square error, and data efficiency are configured as a second preset value combination. Based on the weighted summation results, the volume scan strategies in the candidate strategy set are sorted in descending order to generate a strategy performance ranking table containing the values ​​of each indicator and the overall score.

[0030] Preferably, in S5, the steps for training the adaptive matching model include: Static terrain features, dynamic inflow features, and measurement point attribute features are extracted from the 3D flow field database. The volume scan strategy with the highest comprehensive score in the strategy performance ranking table is used as the label to construct a terrain-wind field-strategy mapping training dataset. Offline training is performed using the random forest algorithm. Normalization mapping, categorical variable encoding, and mutual information feature selection are applied to the training features. The number and maximum depth of decision trees are configured. Cross-validation is used to evaluate the generalization error and generate an adaptive matching model. The measured terrain and meteorological parameters are input into the adaptive matching model for inference. The output includes the optimal volume scanning strategy, which includes strategy identifier, mode type, parameter configuration and expected performance indicators, and is accompanied by model confidence score. If the confidence score is lower than the preset threshold, a secondary calibration is triggered or the model is switched to the default scanning mode.

[0031] Preferably, static terrain features include average elevation, elevation standard deviation, roughness distribution, building density, and average building height; dynamic inflow features include wind direction, wind speed, turbulence intensity, and atmospheric stability; and measuring point attribute features include the absolute coordinates of the measuring point, surrounding environmental obstruction information, and the range of accessible viewing angles.

[0032] The following verification is illustrated with a specific embodiment, such as... Figure 2 The diagram shown is a technical roadmap for a lidar adaptive volume scanning strategy generation method for complex mountain wind fields according to an embodiment of the present invention.

[0033] Step 1: Generate a high-precision three-dimensional flow field using CFD numerical simulation.

[0034] First, a high-fidelity virtual wind field environment was constructed under a unified geographic coordinate system. Spatial data processing for the survey area employed a standardized data fusion process. Based on ALOS 12.5-meter resolution DEM data as the base topographic map, it was resampled to 1-meter resolution using bilinear interpolation, and then combined with urban GIS vector data or measured BIM models to establish accurate building geometry. On this basis, spatial normalization was performed using the CGCS2000 coordinate system, dividing the underlying surface into four categories: water surface, grassland, low-density urban area, and high-density urban area. Corresponding roughness length parameters were fixedly assigned according to land use type, thereby constructing a complete 3D digital scene model including surface undulations, building entities, and roughness attributes. For impassable and restricted areas within the survey area, occlusion objects without physical calculation attributes were set to ensure the accuracy of CFD calculation boundaries.

[0035] During the mesh generation phase, a hybrid structured and unstructured meshing technique was employed to balance computational efficiency and boundary resolution. For areas surrounding buildings and regions with abrupt terrain changes, a localized mesh refinement strategy was enforced, fixing the mesh size to 0.5 meters in these areas, while the mesh size in open areas far from buildings smoothly transitioned to 1.0 meter. To ensure accurate capture of fluid flow characteristics near walls, a uniform 5-layer prism mesh was applied to both the ground and building surfaces, ensuring that the first layer of mesh satisfied the logarithmic law for wall functions. y + Value requirement, that is, satisfying: ; in, The wall friction speed is... y This represents the vertical distance from the center point of the first layer of the grid to the wall. The kinematic viscosity of the fluid. y + The value is a dimensionless parameter used in computational fluid dynamics (CFD) to characterize the resolution of the first layer of mesh near the wall.

[0036] By applying an octree-based grid growth algorithm, the overall grid growth rate is strictly controlled within 1.1, ensuring the numerical stability of the calculation process. The total number of grid cells in the entire domain is controlled at the level of 20 million to 30 million, achieving the optimal balance between computational resource consumption and flow field reconstruction accuracy.

[0037] The numerical solution process for the flow field establishes unified physical and numerical standards to eliminate uncertainties. The k-ωSST turbulence model is consistently used as the core physical model, effectively capturing common flow separation, shear layers, and backflow vortices in mountainous terrain. For the numerical discretization scheme, a second-order upwind scheme is uniformly adopted for the convection term, and the SIMPLE algorithm is used for pressure-velocity coupling to ensure convergence and numerical accuracy. The convergence criteria are set as continuity and a reduction in momentum equation residuals to 10. -4 The following measures are taken, and by monitoring the speed values ​​of key measuring points in real time, it is ensured that the wind speed fluctuation range remains within 0.5% within 100 consecutive iterations. Regarding the boundary condition input, a logarithmic wind speed profile conforming to the characteristics of the atmospheric boundary layer is uniformly set at the inlet, as shown in the following formula: ; Where: u ∗ The friction velocity is represented by z; k represents the von Kármán constant, which is 0.4 in this embodiment; z represents the ground clearance; and z0 represents the surface roughness.

[0038] Set a constant turbulence intensity of 10% and apply a rough wall function to all ground and building surfaces.

[0039] Finally, a flow field database is constructed by executing a standardized automated calculation matrix. The simulation calculations cover 12 discrete wind directions at 30-degree intervals and four fixed wind speed levels of 2 m / s, 5 m / s, 8 m / s, and 12 m / s, for a total of 48 calculation cases. After the numerical solution for each case is completed, the system automatically extracts the spatial flow field data and exports it in the standard NetCDF file format. The exported data structure is strictly standardized to physical parameters such as velocity components (u, v, w), turbulent kinetic energy (k), and dissipation rate (ω) corresponding to spatial coordinate points (x, y, z), thus forming a high-precision, multi-case, and physically consistent virtual wind field database, providing a unique and repeatable source of truth for subsequent virtual lidar detection simulation and strategy evaluation.

[0040] Step 2: Construct a virtual lidar detection environment.

[0041] To achieve quantification and standardization of detection and evaluation, a high-fidelity virtual lidar sensor model needs to be established. The sensor's physical parameters are uniformly configured as follows: operating wavelength set to 1.55 μm, pulse energy fixed at 5 mJ, pulse repetition frequency set to 5 kHz, beam divergence angle precisely controlled at 0.5 mrad, radial sampling interval (distance resolution) fixed at 20 m, maximum detection distance set to 5 km, and receiving aperture area set to 0.05 m². 2 This parameter set forms the physical baseline for the virtual sensor, ensuring a consistent hardware performance baseline for subsequent detection and evaluation under different terrain conditions.

[0042] In terms of environmental effect modeling, a standard exponential decay model is used to characterize the signal transmission process. The extinction coefficient is fixed and calculated based on a visibility of 10km. The signal power decreases exponentially with distance, and a geometric attenuation effect is added, making the signal strength inversely proportional to the square of the distance. The obstruction detection logic uses a ray tracing algorithm to perform intersection calculations between the radar-transmitted ray and the 3D scene model constructed in step 1. When the calculated distance between the intersection point of the ray and the geometric object is less than the coordinate distance of the sampling point, the system directly marks the sampling point as "invalid obstruction," thus accurately eliminating false signals caused by terrain obstruction. The signal-to-noise ratio model is calculated by integrating pulse energy, atmospheric backscattering coefficient, receiving system efficiency, and noise equivalent power to ensure that each simulated signal conforms to the laws of physical detection.

[0043] In the flow field data mapping stage, a standardized interpolation algorithm is used to achieve a smooth conversion of CFD flow field data to lidar sampling points. For each set radar attitude (azimuth angle θ and elevation angle...) The sampling distance R is determined by transforming the polar coordinate system to the Cartesian coordinate system to determine the spatial location of the sampling points. For structured grid data, a trilinear interpolation algorithm is uniformly used to extract the velocity components of the flow field; if transient CFD data is involved, linear interpolation is uniformly used on the time axis to ensure the continuity of velocity changes between different time steps. This process ensures a lossless mapping from physical field data to probe point data and eliminates uncertainties caused by differences in data reading methods.

[0044] Finally, the radial velocity is calculated and noise processing is performed to achieve closed-loop generation of the observation data. Based on the Doppler detection principle, the formula for calculating the radial velocity is: ; Where: u represents the wind speed component in the x-direction; v represents the wind speed component in the y-direction; w represents the wind speed component in the z-direction; θ represents the azimuth angle, the angle between the horizontal projection of the beam and the x-axis; φ represents the elevation angle, the angle between the beam and the horizontal plane.

[0045] Building upon this foundation, random interference in actual detection is simulated by superimposing Gaussian white noise with a standard deviation of 0.3 m / s. All measurement points deemed obstructed, sampling points with a signal-to-noise ratio below a preset threshold (e.g., -10 dB), and sampling points exceeding the maximum detection distance are forcibly marked as invalid. Finally, this is encapsulated into a standardized VirtualLidar function interface, taking a CFD flow field database, scan control parameters, sensor physical configuration, and a 3D scene model as deterministic inputs, and outputting simulation observation data in a unified format, achieving high repeatability and automated deployment of the detection environment.

[0046] Step 3: Define the volume scan strategy space.

[0047] The definition of the scanning strategy space aims to construct a parameterized set that comprehensively covers the detection capabilities of lidar. Basic scanning mode types are systematically divided into six standard forms: Planar Position Display (PPI) mode for large-scale horizontal wind field monitoring; Range and Height Display (RHI) mode focusing on fine detection of vertical profiles; Velocity and Azimuth Display (VAD) mode and Doppler Beam Swing (DBS) mode specifically for efficient wind profile inversion; Grid Scan mode for 3D wind field reconstruction in complex terrain; and Adaptive Trajectory mode, which, by introducing regional priority weight parameters, achieves dynamic focusing and guided detection of high-value targets. These mode definitions cover the complete operational needs from routine meteorological monitoring to fine reconstruction of complex terrain.

[0048] To ensure the executability and uniqueness of the scanning strategy, this method imposes strict parameter constraints and definitions on each strategy. Each scanning strategy π is uniquely determined by an eight-dimensional feature vector, specifically including: scanning mode type (PPI / RHI / VAD / DBS / grid / adaptive), azimuth scanning range (0°~360°), elevation scanning range (-10°~90°), azimuth step size (0.1°~5°), elevation step size (0.1°~5°), dwell time per point (0.05s~1s), number of single-point sampling repetitions (1~10 times), and region priority weight (0~1) specific to the adaptive mode. By standardizing the above parameters, the complex and varied scanning operation can be transformed into a precise numerical instruction set, providing a consistent data foundation for subsequent simulation evaluation and strategy selection.

[0049] To generate the policy space, this method abandons the discrete Cartesian product method, which leads to combinatorial explosion, and adopts a hybrid policy generation process based on "Latin Hypercube Sampling (LHS) + Prior Knowledge Filtering". First, the Latin hypercube sampling technique is used to perform efficient and uniform sampling within the aforementioned eight-dimensional parameter space, strictly controlling the size of the generated candidate set to between 100 and 200 samples to ensure uniform coverage and computational controllability of the search space. Then, engineering-based prior knowledge is introduced for rule-based filtering, that is, invalid combinations are eliminated based on the physical detection characteristics of the LiDAR. For example, policies that cannot meet the reconstruction spatial resolution requirements due to excessively large angular step size, and policies whose signal-to-noise ratio is lower than the detection threshold due to excessively short dwell time, are eliminated. Through this process, a moderately sized, efficient, and physically feasible candidate policy set is finally formed, achieving efficient construction and optimization of the scanning policy space.

[0050] Step 4: Multi-strategy simulation detection and performance evaluation.

[0051] For each policy in the candidate policy set, the system executes a full-process simulation and detection experiment by calling the predefined VirtualLidar function interface, generating an observation point dataset corresponding to that policy. The dataset includes the spatial coordinates (x, y, z), radial velocity observations, and signal-to-noise ratio for each sampling point. During this process, the system accurately calculates the total execution time for each policy, using the following formula: ; in, Indicates the total number of sampling points; This indicates the dwell time at each sampling point; Indicates the number of times the measurement was repeated.

[0052] To ensure the determinism of time cost calculations, the system automatically matches the appropriate wind field reconstruction algorithm based on the scanning mode type: for single-elevation PPI mode, the velocity-azimuth display (VAD) fitting algorithm is used to extract the horizontal wind field components; for multi-elevation PPI mode, a three-dimensional variational (3D-Var) inversion algorithm is used to solve the three-dimensional vector field by minimizing the cost function containing observation error terms and background constraint terms; for DBS mode, an analytical synthesis method based on beam geometry is used to directly calculate the (u,v,w) components; and for grid scanning mode, the Kriging interpolation method is used to construct the global flow field, ensuring that the inversion process has a clear mathematical basis.

[0053] The performance evaluation process employs multi-dimensional quantitative indicators to score the reconstruction results. The system automatically compares the reconstructed wind field with the true CFD data from step 1, calculates the root mean square error (RMSE) of radial velocity to assess deviation, and calculates the correlation coefficient between the reconstructed and true values ​​to evaluate linear reconstruction capability. For complex mountainous terrain, the system pre-identifies and marks high-velocity gradient, high-turbulent kinetic energy, and backflow vortex regions in the CFD flow field as "key detection areas," determining the coverage of these key areas by calculating the proportion of effective sampling points covering these regions.

[0054] Finally, the system ranks all candidate strategies using a multi-objective comprehensive scoring function. This function normalizes the four indicators to the [0,1] interval and then performs a weighted sum. Based on business requirements, the system has two built-in fixed weight configurations: for high-precision reconstruction tasks, the weights for root mean square error, coverage, time efficiency, and data efficiency are fixed at 0.5, 0.2, 0.2, and 0.1, respectively; for rapid emergency detection tasks, the weights for time efficiency, coverage, root mean square error, and data efficiency are adjusted to 0.5, 0.3, 0.1, and 0.1, respectively. Through weighted calculation, the system sorts the strategies in the π set in descending order and finally outputs a standardized "strategy performance ranking table." The table details the values ​​of each indicator and the comprehensive score for each strategy, providing clear optimal strategy labels for the adaptive model training in step 5.

[0055] Step 5: Optimal strategy generation and adaptive matching.

[0056] The core of this step lies in establishing a deterministic mapping model from "environmental physical characteristics" to "optimal control strategies," enabling second-level adaptive switching of detection modes. First, using the simulation data generated in steps 1 to 4, a standardized terrain-wind field-strategy mapping database is constructed. Database samples are rigorously processed and associated with three types of information: the first type is static terrain features, including average elevation, elevation standard deviation, roughness distribution, building density, and average building height; the second type is dynamic incoming flow features, covering boundary parameters such as wind direction, wind speed, turbulence intensity, and atmospheric stability; the third type is measuring point attributes, including the absolute coordinates of the measuring point, surrounding environmental obstruction information, and the range of accessible views. Using the strategy with the highest comprehensive score obtained from CFD simulation evaluation as the label, a standard sample set of several thousand data points is formed through combined calculations of 12 wind directions, 4 wind speed levels, and multiple measuring point layouts, providing a high-quality source of ground truth for training machine learning models.

[0057] The model training employs a random forest algorithm combining offline classification and regression to ensure the repeatability of the training process. During data preprocessing, normalization mapping and categorical variable encoding are uniformly performed, and feature selection is conducted based on mutual information to remove redundant variables with weak impact on the policy. The model architecture is fixed at 100 decision trees with a maximum depth of 15 layers. A 5-fold cross-validation mechanism is used to evaluate the model's generalization error, ensuring robustness in predicting unknown scenarios. In engineering applications, by deploying this pre-trained model, the field system only needs to input measured terrain and meteorological parameters to complete the optimal policy inference output within seconds. For higher-order dynamic scenarios, the system supports an online learning mode. Through a reinforcement learning agent, the real-time detected information gain is used as a reward signal to continuously optimize policy parameters, thereby achieving intelligent progression from "initial recommendation" to "optimal closed loop."

[0058] Ultimately, the output optimal strategy strictly adheres to a unified executable instruction format, including a unique strategy identifier, mode type, detailed parameter configuration (such as elevation angle, azimuth range, and dwell time for the PPI mode), and expected performance indicators. Each output instruction is accompanied by a model confidence score; if the confidence score falls below a preset threshold, the system automatically triggers secondary calibration or switches to the default safe scanning mode. This mechanism constructs a complete technical closed loop from environmental perception and feature mapping to action execution, ensuring that the LiDAR system maintains the most efficient and accurate operation in complex mountainous environments.

[0059] In this application, unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. In case of any inconsistency, the meaning set forth in this specification or derived from the content described herein shall prevail. Furthermore, the terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit the scope of this application.

[0060] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for generating adaptive volume scanning strategies for lidar in complex mountainous wind fields, characterized in that, Includes the following steps: S1. Construct a three-dimensional digital scene model based on the digital elevation model data and land cover data of the survey area, perform fluid dynamics numerical simulation on the three-dimensional digital scene model, and generate a three-dimensional flow field database containing velocity components, turbulent kinetic energy and dissipation rate parameters under multiple working conditions. S2. Construct a virtual lidar sensor model, map the three-dimensional flow field database to the sampling points of the virtual lidar sensor, calculate the radial velocity of each sampling point and add noise to generate a simulation observation dataset. S3. Define a volume scan strategy space consisting of scanning mode type, azimuth scanning range, elevation scanning range, azimuth step size, elevation step size, dwell time per point, number of sampling repetitions per point, and region priority weight parameters. Generate a candidate strategy set based on Latin hypercube sampling and prior knowledge screening. S4. Call the virtual lidar sensor model, perform simulation detection experiments on each volume scan strategy in the candidate strategy set, reconstruct the wind field based on the simulation observation dataset, compare the reconstructed wind field with the three-dimensional flow field database, calculate the root mean square error, correlation coefficient, key area coverage and data efficiency, and generate a strategy performance ranking table through multi-objective weighted scoring. In step S4, the wind field reconstruction based on the simulation observation dataset includes: The wind field reconstruction algorithm is automatically matched according to the scanning mode type. Among them, the single elevation angle plane position display mode adopts the velocity azimuth display fitting algorithm, the multi elevation angle plane position display mode adopts the three-dimensional variational inversion algorithm, the Doppler beam swing mode adopts the analytical synthesis method based on beam geometry relationship, and the grid scanning mode adopts the Kriging interpolation method. The reconstructed wind field is compared point by point with the three-dimensional flow field database. The root mean square error of radial velocity is calculated to evaluate the deviation. The correlation coefficient between the reconstructed value and the true value is calculated to evaluate the linear reconstruction capability. The proportion of effective sampling points covering the key detection area is calculated to determine the coverage of the key area. The proportion of effective data points to the total sampling points is calculated to determine the data efficiency. In step S4, the step of generating a strategy performance ranking table through multi-objective weighted scoring includes: The root mean square error, correlation coefficient, key area coverage, and data efficiency are normalized to a preset interval and then summed with weights. For high-precision reconstruction tasks, the weights of root mean square error, key area coverage, time efficiency, and data efficiency are configured as a first preset value combination. For rapid emergency detection tasks, the weights of time efficiency, key area coverage, root mean square error, and data efficiency are configured as a second preset value combination. Based on the weighted summation result, the volume scan strategies in the candidate strategy set are sorted in descending order to generate a strategy performance ranking table containing the values ​​of each indicator and the comprehensive score; S5. Extract the static terrain features and dynamic inflow features corresponding to the three-dimensional flow field database, train the adaptive matching model with the best strategy in the strategy performance ranking table as the label, input the measured terrain and meteorological parameters into the adaptive matching model, and output the optimal volume scanning strategy. In step S5, the steps for training the adaptive matching model include: Static terrain features, dynamic inflow features, and measurement point attribute features are extracted from the 3D flow field database. The volume scan strategy with the highest comprehensive score in the strategy performance ranking table is used as the label to construct a terrain-wind field-strategy mapping training dataset. Offline training is performed using the random forest algorithm. Normalization mapping, categorical variable encoding, and mutual information feature selection are applied to the training features. The number and maximum depth of decision trees are configured. Cross-validation is used to evaluate the generalization error and generate an adaptive matching model. The measured terrain and meteorological parameters are input into the adaptive matching model for inference. The output includes the optimal volume scanning strategy, which includes strategy identifier, mode type, parameter configuration and expected performance indicators, and is accompanied by model confidence score. If the confidence score is lower than the preset threshold, a secondary calibration is triggered or the model is switched to the default scanning mode.

2. The method for generating adaptive volume scanning strategy for lidar in complex mountainous wind fields according to claim 1, characterized in that, In step S1, the steps of constructing a three-dimensional digital scene model based on the digital elevation model data and land cover data of the survey area include: The digital elevation model data of the survey area is resampled to a preset resolution, and the geometric figures of buildings are established by combining the surface cover vector data. Spatial normalization is performed according to a unified geographic coordinate system. The underlying surface is divided into four types of areas according to land use: water surface, grassland, low-density urban area and high-density urban area. Corresponding roughness length parameters are assigned to construct a three-dimensional digital scene model that includes surface undulation, building entities and roughness attributes. The three-dimensional digital scene model is subjected to a hybrid structured and unstructured mesh generation. Local densification is performed around buildings and in areas with abrupt terrain changes. Prism layer meshes are set on the ground and building surfaces to meet the wall function requirements, and the overall mesh growth rate and total mesh volume are controlled. The turbulence model was selected as the core physical model. The discretization scheme of the convection term, the pressure-velocity coupling algorithm and the convergence criterion were set. The logarithmic wind speed profile and turbulence intensity were set at the inlet boundary. The rough wall function was applied to the ground and building walls. Batch numerical solutions were performed for multiple wind directions and multiple wind speeds. Spatial flow field data were extracted and exported as standard format files to construct a three-dimensional flow field database.

3. The method for generating adaptive volume scanning strategy for lidar in complex mountainous wind fields according to claim 1, characterized in that, In step S2, the steps for constructing the virtual lidar sensor model include: Configure the physical parameters of the virtual lidar sensor, including its operating wavelength, pulse energy, pulse repetition frequency, beam divergence angle, range resolution, maximum detection range, and receiving aperture area. An exponential decay model is used to characterize the signal transmission process, and a ray tracing algorithm is used to perform occlusion discrimination. The intersection operation between the radar emitted ray and the three-dimensional digital scene model is performed to mark the occluded sampling points. For each radar attitude and sampling distance, the spatial position of the sampling point is determined by coordinate system transformation. The flow field velocity component is extracted from the three-dimensional flow field database by interpolation algorithm, the radial velocity is calculated and Gaussian white noise is superimposed, and invalid values ​​are assigned to sampling points with obstruction, sampling points with signal-to-noise ratio below the threshold and sampling points exceeding the maximum detection distance, thus generating a simulation observation dataset.

4. The method for generating adaptive volume scanning strategy for lidar in complex mountain wind fields according to claim 1, characterized in that, In step S3, the step of defining the volume scan strategy space includes: The scanning modes are classified into six standard forms: planar position display mode, distance and height display mode, velocity and azimuth display mode, Doppler beam swing mode, grid scanning mode, and adaptive trajectory mode. Each volume scan strategy is parameterized into an eight-dimensional feature vector consisting of scan mode type, azimuth scan range, elevation scan range, azimuth step size, elevation step size, dwell time per point, number of sampling repetitions per point, and region priority weight. By utilizing the Latin hypercube sampling technique to perform uniform sampling in the eight-dimensional feature vector space, invalid parameter combinations are eliminated based on the physical detection characteristics of lidar, generating a candidate strategy set that is controllable in scale and physically feasible.

5. The method for generating adaptive volume scanning strategy for lidar in complex mountain wind fields according to claim 4, characterized in that, The static terrain features include average elevation, elevation standard deviation, roughness distribution, building density, and average building height; the dynamic inflow features include wind direction, wind speed, turbulence intensity, and atmospheric stability; and the measurement point attribute features include the absolute coordinates of the measurement point, surrounding environmental obstruction information, and the range of accessible viewing angles.

Images