A plateau mountainous area low-altitude wind field modeling method based on a terrain response relationship library

Abstract

The application discloses a kind of plateau mountain area low-altitude wind field modeling method based on terrain response relationship library, it is related to complex terrain meteorological modeling, low-altitude airspace environmental perception and computational fluid dynamics technical field.Digital elevation model, surface roughness, meteorological background field and multi-source observation data are obtained, the research area is divided into multiple typical terrain units by terrain feature clustering, and the terrain-wind field response relationship database is calculated and constructed for each unit offline;Based on the background wind field, the database is searched for local dynamic correction, the terrain thermal correction is carried out by superimposing valley wind and slope thermal convection correction term, the isolated forest algorithm is used for anomaly detection and combined with observation data fusion correction, and the spatial and temporal continuous gridded low-altitude wind field and credibility identification are output.The physical accuracy of complex terrain wind field and the business computing efficiency are considered, the wind speed and direction simulation error can be significantly reduced, the local thermal circulation can be effectively captured, and it is suitable for low-altitude flight support and route planning.

Description

Technical Field

[0001] This invention relates to the fields of complex terrain meteorological modeling, low-altitude airspace environmental perception, and computational fluid dynamics. Specifically, it relates to a wind field modeling method applicable to low-altitude airspace of 0–500m above ground in high-altitude, dramatically undulating plateau and mountainous areas. This application can be used for low-altitude flight support, UAV route planning, low-altitude traffic operation risk assessment, meteorological support for mountain emergency rescue, and refined environmental forecasting in plateau and mountainous areas. Background Technology

[0002] With the rapid development of the low-altitude economy and low-altitude transportation, there is a growing demand for refined, continuous, and rapid modeling of wind speed, direction, vertical velocity, wind shear, and turbulence intensity in the near-surface layer, especially in the 0–500m altitude range. High-altitude plateaus and mountainous regions have high average elevations, low air density, strong topographic dissection, intersecting valleys, and significant differences in underlying surface thermals. The low-altitude atmospheric boundary layer is influenced by a large-scale weather background field, valley winds, slope thermal convection, accelerated valley contraction, ridge flow around mountains, and leeward slope separation vortices. Consequently, the wind field exhibits significant non-uniformity, abrupt changes, and diurnal variation characteristics in both the horizontal and vertical directions.

[0003] Existing methods for low-level wind field modeling mainly include: downscaling methods based on mesoscale numerical weather prediction models, spatial interpolation methods based on a small number of surface meteorological stations, local simulation methods based on computational fluid dynamics, and statistical correction methods based on machine learning. While these methods are practical in plains or gently rolling hills, they still have the following technical limitations when used for low-level wind field modeling in complex mountainous areas and deep canyons above 3000m: The simulation of low-altitude wind fields in complex terrains has large biases. Mesoscale models are limited by grid resolution and terrain smoothing, making it difficult to accurately represent the dynamic forcing effect of terrain units such as narrow valleys, steep slopes, ridges, and passes on the near-ground wind field. They tend to underestimate canyon acceleration, ridge bypass, leeward side recirculation, and local wind shear.

[0004] The ability to capture local thermal circulation is insufficient. Significant diurnal radiation differences exist in plateau and mountainous areas, and the upslope winds caused by slope heating, the downslope winds formed by the sinking of cold air at night, and the transition between valley and mountain winds have a prominent impact on low-altitude flight. Traditional interpolation or simple shape correction methods often only consider distance and altitude differences, failing to explicitly characterize the coupling mechanism between topographic dynamics and thermal effects.

[0005] The ability to fuse multi-source observational data is limited. Reanalysis data, automatic weather stations, wind lidar, UAV radiosondes, and remote sensing surface parameters differ significantly in spatiotemporal resolution, error characteristics, and coverage. Existing methods typically rely heavily on single data sources and lack quality control, spatiotemporal consistency, and dynamic weighting fusion mechanisms tailored to the complex terrain of plateaus.

[0006] The conflict between high-resolution computing efficiency and engineering deployment is prominent. Directly conducting meter- or ten-meter-level computational fluid dynamics simulations across the entire region requires enormous computational resources, making it difficult to meet the needs of rapid operational modeling and real-time updates; while purely data-driven models are prone to extrapolation distortion and insufficient physical consistency in sparsely observed regions.

[0007] The ability to identify anomalous wind fields and observed anomalies is insufficient. Low-altitude wind fields in plateau and mountainous areas often exhibit real extreme phenomena such as sudden gusts, local vortices, and strong wind shear. Simultaneously, observation equipment may experience distortion, drift, or short-term data gaps. Existing methods lack effective anomaly detection mechanisms, making it difficult to distinguish between real, locally abrupt wind field changes and observed anomalies, leading to a decrease in the continuity and reliability of the output gridded wind field.

[0008] Therefore, there is an urgent need to propose a modeling method for low-altitude wind fields in plateau and mountainous areas that integrates multi-source data, mesoscale background field, small-scale topographic response relationship, topographic dynamic and thermodynamic parameterization correction and anomaly detection mechanism, so as to improve the spatial resolution, local circulation expression ability, physical consistency and engineering calculation efficiency of low-altitude wind fields at 0-500m above the ground in complex terrain. Summary of the Invention

[0009] The purpose of this invention is to address the aforementioned problems by providing a method for modeling low-altitude wind fields in plateau and mountainous areas based on a terrain response relation database. This method acquires high-resolution terrain, surface roughness, multi-source meteorological background fields, and local observation data. Complex terrain is clustered and partitioned using indicators such as terrain undulation, slope aspect, and valley axis direction. Computational fluid dynamics pre-calculation is performed on typical terrain units, and a terrain-wind field response relation database is constructed. Using reanalysis data or mesoscale model output as the background field, and combining virtual station interpolation, local dynamic correction, terrain thermal correction, deep learning super-resolution reconstruction, and anomaly detection and correction, a spatiotemporally continuous, physically consistent, and gridded wind field suitable for low-altitude airspace in plateau and mountainous areas is obtained.

[0010] The technical solution of the present invention is as follows: A method for modeling low-altitude wind fields in plateau and mountainous areas based on a terrain response relation database includes the following steps: High-resolution digital elevation models, surface roughness, land use, meteorological background field and low-altitude observation data of the study area were acquired, and coordinate, time and height layer consistency preprocessing was performed. Based on the digital elevation model, terrain feature parameters are calculated, and the study area is divided into multiple typical terrain units using a clustering algorithm; Virtual stations were set up for each typical terrain unit, and computational fluid dynamics pre-calculations were performed under different inflow conditions to build a database of terrain-wind field response relationships. Using the meteorological background field as the background wind field, interpolation is performed to the target grid point, and similar samples are retrieved from the database based on the typical terrain unit to which the target grid point belongs, to perform local dynamic correction of the background wind field; Based on the surface temperature change rate, solar radiation, slope aspect, valley axis direction, valley depth and atmospheric stability, valley wind correction terms and slope thermal convection correction terms are constructed to perform topographic thermal correction on the wind field after dynamic correction. Anomaly detection feature vectors are constructed for the thermally corrected wind field. Anomaly detection is performed using the isolated forest algorithm, and low-altitude observation data is combined for fusion correction to output a gridded low-altitude wind field within the range of 0–500 m above the ground.

[0011] By using terrain clustering and CFD pre-built libraries, the contradiction between large simulation deviations and computational efficiency is resolved; by using dynamic and thermal correction, the problem of insufficient thermal circulation capture is solved; and by using anomaly detection and multi-source fusion, the problems of weak data fusion and poor anomaly identification are solved, thus achieving high-precision, high-efficiency, and high-reliability modeling of low-altitude wind fields in plateau and mountainous areas is solved.

[0012] Furthermore, the valley wind correction term is applied along the valley axis, and its intensity is related to the rate of change of surface temperature, valley depth, and atmospheric stability, as expressed as: , in, The empirical coefficient for valley winds. Because the valley is deep, For stability, The vertical attenuation height of valley winds. The unit vector along the valley axis. For surface temperature, To prevent stable terms with a denominator of zero, This refers to the height above the ground.

[0013] The valley wind correction formula in this application explicitly quantifies the influence of surface warming rate, valley depth and atmospheric stability on low-altitude wind fields. It is applied along the valley axis and decays exponentially with height, which can accurately simulate the vertical structure and conversion process of upslope winds during the day and downslope winds at night. Compared with traditional interpolation methods without physical constraints, the accuracy of wind direction simulation is improved by more than 20%.

[0014] Furthermore, the slope heat convection correction term is applied along the uphill or downhill direction of the slope, and its intensity is related to the surface sensible heat flux, slope, aspect heat transfer factor, and atmospheric stability, and is expressed as: , in, This is the empirical coefficient for slope wind. For surface sensible heat flux, For slope, Slope direction, The azimuth of the sun. air density, The specific heat capacity of air at constant pressure. For reference potential temperature, The vertical attenuation height of the slope wind. This is the unit vector along the slope direction.

[0015] The slope wind correction formula in this application is the first to couple surface sensible heat flux, slope heating degree and atmospheric stability into low-altitude wind field modeling, which can accurately characterize the thermal driving differences of different slope aspects at different times of day and night, and effectively solve the defect of existing methods that cannot distinguish the thermal effects of windward and leeward slopes.

[0016] Furthermore, the anomaly detection feature vector includes: east-west wind speed. North-south wind speed Vertical velocity Wind speed ,wind direction Vertical wind shear Turbulent kinetic energy, Richardson number Topographic relief ,slope and surface roughness Anomaly score threshold for the Isolation Forest algorithm Using a value of 0.60 to 0.75, for points identified as anomalous, a secondary discrimination is performed by combining the consistency of neighboring observations and terrain triggering conditions, preserving the true abrupt wind field and outputting risk indicators.

[0017] The constructed anomaly detection feature vector integrates multi-dimensional physical information such as wind shear, turbulent kinetic energy, Richardson number, and topographic relief. After initial screening through isolated forests, it is further judged by combining the consistency of neighboring observations and topographic triggering conditions. It can accurately distinguish between observation equipment anomalies, interpolation errors, and real local sudden wind fields (such as canyon gusts and ridge vortices), significantly reducing the false alarm rate and the missed alarm rate.

[0018] Furthermore, the fusion correction employs optimal interpolation or ensemble Kalman filtering methods, and the correction formula is as follows:

[0019] in, The revised forecast wind field, To analyze the wind field, For background error covariance, For observation error covariance, Here is the gain matrix. For the observation operator, This is the transpose of the observation operator.

[0020] By employing optimal interpolation or ensemble Kalman filtering, and dynamically calculating the gain matrix based on the background error covariance and observation error covariance, a weighted fusion of multi-source observations (automatic weather stations, lidar, and UAVs) and corrected wind fields is achieved. In densely observed areas, the observation constraint weights are automatically increased; in sparsely observed areas, the physical constraints of the terrain response database are maintained, effectively solving the problem of the difficulty in uniformly utilizing multi-source heterogeneous data.

[0021] Furthermore, the background wind field interpolation employs an inverse distance weighted method, with interpolation weights... ,in For the grid points to be estimated, For background grid points or virtual sites, For spatial distance, As a stable term, The distance attenuation index is 1.5–3; the local dynamic correction uses a kernel-weighted method to obtain the correction from the database. , in, Let this be the current state vector. For the first in the database The state vector of a pre-computed sample For kernel width parameter, To pre-calculate the total sample size, This is the kernel weight function.

[0022] The inverse distance weighted interpolation method in this application ensures a smooth transition of the background wind field from coarse to fine grids; the kernel weight retrieval method dynamically weights the data based on the Euclidean distance between the current state vector (terrain type, inflow wind direction and speed, stability, heat flux, etc.) and the database samples, avoiding the skipping errors of simple nearest neighbor matching. This mechanism replaces real-time CFD calculation with table lookup, reducing the modeling time per time interval from several hours to minutes, while maintaining physical consistency.

[0023] Furthermore, this also includes wind field super-resolution reconstruction based on deep learning: Using coarse-resolution background wind field, digital elevation model terrain features, surface roughness, terrain category coding, and historical corrected wind field as inputs, U-Net, residual convolutional network, or ConvLSTM are used to output high-resolution 3D wind field; the training loss function includes wind speed reconstruction error, wind direction error, mass conservation constraint, and observation consistency constraint.

[0024] The optional deep learning module of this invention learns the mapping relationship from coarse-grid wind field to high-resolution wind field. The loss function simultaneously constrains wind speed error, wind direction error, mass conservation and observation consistency, and can achieve millisecond-level super-resolution output while maintaining physical constraints, further meeting the business requirements of 5-15 minute rolling updates.

[0025] Furthermore, each entry in the topography-wind field response database includes: typical topographic unit characteristics, inflow wind profile, inflow wind direction, atmospheric stability, surface heat flux, roughness length, and three-dimensional wind field response increment relative to the background wind field; computational fluid dynamics pre-calculation employs Reynolds-averaged Navier-Stokes equations combined with... Turbulence model.

[0026] This application employs the RANS equations and a k-ε turbulence model for CFD pre-calculation. For each typical terrain unit, wind field response increments are generated under various inflow conditions (wind direction 0–360°, wind speed 2–20 m / s, etc.). The database stores labels such as terrain features, inflow profiles, stability, and heat flux, along with the corresponding three-dimensional wind field increments. This database transforms complex terrain dynamic processes into reusable response relationships, fundamentally solving the problem of excessive computational cost in high-resolution CFD across the entire region.

[0027] Furthermore, the terrain feature parameters include elevation. Topographic relief ,slope Slope aspect sinusoidal component cosine component of slope curvature Valley Ridge Index Sine component along the valley axis, cosine component along the valley axis, and roughness length Clustering algorithm is Number of clusters It is determined by both the profile coefficient and the elbow method.

[0028] The 11-dimensional terrain feature vector extracted in this application covers elevation, relief, slope, aspect, curvature, valley depth, ridge index, valley axis direction, and roughness. By automatically dividing terrain units through K-means clustering and contour coefficient / elbow method, it can effectively distinguish complex landforms (deep canyons, ridges, passes, slopes, etc.) in plateau and mountainous areas, providing an accurate terrain category basis for subsequent targeted CFD pre-calculation and dynamic correction.

[0029] Furthermore, the gridded low-altitude wind field within the output range of 0–500m above the ground includes east-west wind speeds at multiple altitude levels within this range. North-south wind speed Vertical velocity Wind speed ,wind direction Vertical wind shear, turbulence intensity, and reliability indicators; low-altitude observation data includes one or more of the following: automatic weather station observation data, wind profile data from wind-measuring lidar, and UAV observation data.

[0030] The low-altitude wind field output by this application includes horizontal wind components, vertical velocity, wind speed, wind direction, vertical wind shear, turbulence intensity, and confidence indicators, which can meet the comprehensive needs of low-altitude flight path planning, aircraft attitude control, and risk assessment. It also supports the fusion of three types of observation data: automatic weather station, lidar, and UAV, and has good data adaptability.

[0031] Compared with existing technologies, the advantages of this invention are: 1. Significantly reduces the simulation bias of low-altitude wind fields in complex terrain. Multi-dimensional terrain features such as topographic relief, slope, aspect, valley depth, ridge index, and valley axis direction are extracted using a high-resolution digital elevation model. K-means clustering is then used to divide the study area into typical terrain units such as deep canyons, ridges, passes, and windward slopes. For each typical unit, high-precision computational fluid dynamics wind field response relationships are pre-calculated offline, constructing a terrain-wind field response database. During operational use, dynamic corrections are quickly retrieved based on the terrain category and inflow conditions of the target grid point, enabling precise representation of dynamic forcing effects such as canyon acceleration, ridge bypass, leeward recirculation, and local wind shear.

[0032] 2. Effectively capturing local thermal circulation and thermodynamic coupling effects. Based on dynamic correction, this application innovatively introduces valley wind correction terms and slope thermal convection correction terms related to terrain type. The valley wind correction is applied along the valley axis, and its intensity is related to the surface temperature change rate, valley depth, and atmospheric stability; the slope wind correction is applied along the slope aspect, and its intensity is related to sensible heat flux, slope, and aspect heat transfer factor. Both corrections decay exponentially with height and are stability-suppressed based on the Richardson number. This application can explicitly characterize the impact of daytime upslope winds, nighttime downslope winds, and the valley wind transition process on the low-altitude wind field, overcoming the deficiency of traditional parameterization schemes that lack a terrain thermodynamic coupling mechanism.

[0033] 3. Achieving deep fusion and consistency correction of multi-source heterogeneous observation data. This application performs unified coordinate, time, and altitude layer standardization preprocessing on reanalysis data, mesoscale model output, automatic weather stations, wind lidar, UAV radiosondes, and remote sensing surface parameters. After anomaly detection, optimal interpolation or ensemble Kalman filtering methods are used to dynamically fuse multi-source observations into the corrected wind field. This mechanism can automatically adjust the fusion weights according to observation density and error covariance, maintaining the physical constraints of the topographic response library in sparsely observed areas and achieving high-precision correction in densely observed areas, significantly improving the spatial continuity and temporal consistency of the low-altitude wind field.

[0034] 4. Balancing physical accuracy with engineering computational efficiency, resolving the accuracy-speed contradiction. This application employs a modeling strategy combining offline budgeting and online retrieval, performing high-resolution CFD pre-computation only on a limited number of typical terrain units to construct a reusable terrain-wind field response database. During operational runs, only terrain category matching, similar sample retrieval, and a small amount of parametric calculations are required, eliminating the need for real-time CFD simulation of the entire region. Simultaneously, optional deep learning super-resolution modules (such as U-Net) can further rapidly map the coarse-resolution background field to a high-resolution wind field. This architecture reduces the modeling time per session from several hours to minutes or even seconds, meeting the real-time requirement of 5-15 minute rolling updates for low-altitude traffic support while maintaining physical consistency.

[0035] 5. It possesses intelligent identification capabilities for anomaly observations and real abrupt wind field changes. This application constructs a multi-dimensional feature vector including wind speed, wind direction, vertical wind shear, turbulence intensity, Richardson number, topographic relief, slope, and surface roughness, and employs the isolated forest algorithm to detect anomalies in the corrected wind field. For points identified as anomalous, a secondary discrimination is performed based on the consistency of neighboring observations, topographic triggering conditions, and physical continuity: if an anomalous point meets the conditions of strong wind shear or topographic flow around the terrain and is supported by multi-source observations, it is retained as a real local abrupt wind field and a risk label is output; otherwise, it is labeled as an observational anomaly or interpolation error and is removed or replaced. The method of this application effectively reduces the interference of equipment drift, short-term missing measurements, and interpolation distortion on the final gridded wind field, improving the reliability and usability of the output product.

[0036] By employing an integrated technical framework encompassing "terrain clustering, response library pre-calculation, dynamic and thermal correction, anomaly detection, and multi-source fusion," this invention addresses long-standing challenges in low-altitude wind field modeling (0-500m above ground) in complex mountainous regions, including low accuracy, missing thermal effects, data silos, excessive computational load, and poor anomaly robustness. Compared to existing technologies, this invention demonstrates significant advancements in wind speed / direction prediction accuracy, local circulation representation capabilities, computational timeliness, and output stability, providing reliable meteorological support for ensuring low-altitude flight safety. Attached Figure Description

[0037] Figure 1 This is a flowchart of the method described in this application.

[0038] Figure 2 A schematic diagram illustrating the principle of constructing a topography-wind field response relationship database.

[0039] Figure 3 A schematic diagram illustrating the principle of topographic dynamic thermodynamic correction.

[0040] Figure 4 This is a framework diagram for multi-source observation fusion and anomaly detection and correction. Detailed Implementation

[0041] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0042] The features and performance of the present invention will be further described in detail below with reference to embodiments.

[0043] Please see Figure 1-4 A method for modeling low-altitude wind fields in plateau and mountainous areas based on a terrain response relation database, such as... Figure 1 As shown, the method first constructs a unified spatiotemporal dataset using data from digital elevation, surface roughness, land use, reanalysis or mesoscale model wind fields, and observations from automatic weather stations, wind-measuring lidar, and UAVs. Then, it identifies typical terrain units through terrain feature clustering and pre-calculates a local wind field response relation database for each terrain unit. During operational use, it performs rapid retrieval and correction based on the background wind field, overlaying valley winds, slope thermal convection, stability, and observational fusion corrections to output a three-dimensional low-altitude wind field with a credibility indicator. Its technical contribution lies in replacing real-time CFD calculations across the entire region with a reusable terrain-wind field response database, ensuring that the low-altitude wind field in complex plateau terrain has both physical constraints and meets the requirements of rolling updates for engineering timeliness.

[0044] Specifically, the following steps are included: S1 Multi-source Data Acquisition and Spatiotemporal Consistency Preprocessing in Plateau and Mountainous Areas: Acquire high-resolution digital elevation model (DEM), land use / land cover data, and roughness length of the study area. The study area includes vegetation index, surface temperature, underlying surface type, reanalysis data or mesoscale numerical weather prediction data, automatic weather station observation data, wind corridor data from wind-measuring lidar, and UAV observation data. The study area is spatially defined. The low-altitude modeling altitude range is the height above the ground. The time range is .

[0045] All types of data are projected onto a unified coordinate system and unified to the target time step. and target horizontal grid resolution , Quality control was performed on the observation data, removing wind speed, wind direction, temperature, and air pressure observations that were significantly outside the physical range; short-term missing data were supplemented using linear interpolation, spline interpolation, or Kalman smoothing; and wind direction variables were converted into east-west wind speed vectors using vector decomposition. and north-south wind speed vector Then interpolation is performed to avoid errors caused by jumps in the 0° / 360° direction.

[0046] , in, For wind speed, The wind direction angle is based on due north.

[0047] For lidar and UAV wind corridor data, according to altitude layer Perform vertical interpolation to form an observation wind corridor set. ,in, For the east-west wind speed component, This represents the north-south wind speed component. This represents the vertical wind speed component.

[0048] S2-based terrain unit clustering partitioning based on topographic relief and aspect: The terrain feature parameters of each grid cell in the study area are calculated based on the DEM. The terrain feature parameters include at least elevation. ,slope Slope aspect Topographic relief curvature Valley Ridge Index Valley axis direction and terrain shading factor Topographic relief Within radius The calculation is performed within the moving window, and the formula is: , in, This is the set of all grid points.

[0049] slope and slope Determined based on the first derivative of the DEM, the formula is: , , To avoid discontinuities in the slope aspect angle near 0° and 360°, the slope aspect... Convert to and Two components. Constructing a terrain feature vector. : , After standardizing the terrain feature vectors, the study area was divided into groups using the K-means clustering algorithm. Typical terrain units, The clustering objective function was determined jointly by contour coefficients, the elbow method, and interpretation of artificial landforms. The objective function is:

[0050] in, For the first Terrain-like units, For the first Cluster centers for topographic units. After clustering, each topographic unit is labeled with its landform type, including but not limited to deep canyons, open valleys, windward slopes, leeward slopes, ridges, passes, plateaus, and basin edges.

[0051] S3 Virtual Site Layout and Wind Field Pre-calculation for Typical Topographic Units: like Figure 2 As shown, for each type of typical terrain unit At least one virtual site is deployed based on terrain representativeness, observational accessibility, and wind field sensitivity. Virtual stations are deployed near the cluster center of the terrain unit and in locations where the local slope and undulation are close to the category mean; for areas with drastic wind field changes, such as canyons, passes, and ridges, the density of virtual stations is increased.

[0052] A high-resolution computational fluid dynamics model was constructed using each virtual site and its neighborhood as a computational subdomain. The network resolution ranged from 5 to 50 meters, with vertical densification ranging from 0 to 100 meters above the ground. The computational boundary conditions were determined using background wind profiles provided by reanalysis data or mesoscale models. Temperature profile Atmospheric stability Surface heat flux and roughness length Sure.

[0053] Computational fluid dynamics models can use Reynolds-averaged values. Equations, large eddy simulations, or diagnostic wind field models are used. For scenarios with high real-time requirements in engineering, incompressible RANS equations combined with k-ε or k-ω turbulence closure models are employed; for localized strong thermal convection or typical experimental verification scenarios, LES can be used for high-fidelity pre-calculation.

[0054] , , in, It is a three-dimensional wind speed vector. air density, This is the pressure disturbance term. Molecular viscosity, For turbulent viscosity, For buoyancy, This is a topographic dynamic forcing term.

[0055] By varying conditions such as inflow wind direction, inflow wind speed, vertical wind shear, stability, surface heat flux, and roughness, multiple sets of pre-calculated samples are generated, and the local response increment between the background wind field and the high-resolution wind field is extracted. Each sample is written into the terrain-wind field response database. Database entries can be represented as: , in, Indicates the first Topographic features of terrain-like units, Indicates the inflow profile. Indicates the direction of the incoming wind. This represents the increase in wind field caused by topographic dynamics and thermal effects.

[0056] S4 background wind field interpolation and local dynamic correction: Use reanalysis data or mesoscale model output as background wind field Background wind field according to time and height Interpolate to virtual stations and target grid points. For horizontal interpolation, the inverse distance weighted method is preferred; for vertical interpolation, a combination of logarithmic wind profiles, power-law wind profiles, or spline interpolation is preferred. The inverse distance weighted interpolation formula is: , , in, For the grid points to be estimated, For background grid points or virtual sites, For spatial distance, To prevent stable terms with a denominator of zero, The distance attenuation index ranges from 1.5 to 3.

[0057] Based on the terrain category of the grid points to be estimated Background inflow direction Inflow velocity Stability and surface heat flux From the topography-wind field response database The most similar pre-computed sample is retrieved; when multiple similar samples are retrieved, the local dynamic correction is obtained by weighting with kernel weights. : , , in, Let this be the current state vector. For the first in the database A pre-computed sample state vector, For kernel width parameter, To pre-calculate the total sample size, The kernel weight function is used. The background wind field is added to the local dynamic correction to obtain the dynamically corrected wind field: , By using the above method, high-resolution CFD calculations can be performed on the entire region in real time. High-resolution local dynamic correction results can be obtained by pre-calculating typical terrain units and searching the database.

[0058] S5 considers topographic thermodynamic parameterization corrections for valley winds and slope thermal convection: Based on the dynamically corrected wind field, a parameterized scheme for thermal circulation in complex plateau terrain is introduced to correct the vertical distribution of low-altitude wind speeds caused by valley winds, slope thermal convection, and stability variations. This is based on surface temperature. Solar radiation Near-surface temperature gradient and terrain slope To determine the direction and intensity of thermal forcing on the slope.

[0059] Valley wind correction term along valley axis direction The intensity of the application is related to the surface warming rate, valley depth, and atmospheric stability, and can be expressed as:

[0060] in, The empirical coefficient for valley winds. Because the valley is deep, For stability, The vertical attenuation height of valley winds. This is the unit vector along the valley axis. During the day, when the surface warms up, valley winds tend to develop from the valley mouth or lower areas to higher areas; at night, when the surface cools down, cold air flows down the valley floor.

[0061] Slope heat convection correction terms along the uphill or downhill direction of the slope. The intensity of the application, which is related to slope aspect heating, slope gradient, and surface heat flux, can be expressed as: , in, This is the empirical coefficient for slope wind. For surface sensible heat flux, For slope, Slope direction, The azimuth of the sun. air density, The specific heat capacity of air at constant pressure. For reference potential temperature, The vertical attenuation height of the slope wind.

[0062] like Figure 3 As shown, coupling the dynamic correction term with the thermal correction term yields the thermodynamically corrected wind field: , For nighttime stable boundary layers or strong temperature inversion conditions, a stability suppression factor is introduced. Correcting vertical mixing intensity: , in, For stability coefficient, The Richardson number represents the volume. When... As the temperature increases, vertical mixing is suppressed, and low-level wind shear is enhanced.

[0063] S6 Wind Field Super-Resolution Reconstruction and Fast Prediction Based on Deep Learning: To improve the efficiency of rapid operational modeling, a deep learning-based wind field super-resolution reconstruction module can be built on the wind fields obtained from S4 and S5. The super-resolution reconstruction module takes the coarse-resolution background wind field, DEM topographic features, surface roughness, surface heat flux, topographic category coding, and historically corrected wind field as input, and uses high-resolution CFD pre-calculation results or high-resolution wind fields after fusion observations as training targets to learn the mapping relationship from coarse-grid wind fields to fine-grid low-altitude wind fields.

[0064] The super-resolution reconstruction module employs U-Net, residual convolutional networks, ConvLSTM, or Transformer structures. Its loss function consists of wind speed reconstruction error, wind direction error, mass conservation constraints, and observation consistency constraints. , in, To predict wind field for the model, For reference to high-resolution wind fields, and These are the predicted wind direction and the reference wind direction, respectively. For the observation operator, For observation data, , , , This represents the weighting coefficient. This module enables the rapid output of high-resolution wind fields while maintaining physical consistency, further meeting the real-time support needs for low-altitude traffic.

[0065] S7 outlier detection, observation fusion correction, and gridded output: like Figure 4 As shown, outlier detection is performed on thermally corrected wind fields or super-resolution reconstructed wind fields. The feature vector of each grid point within a given time window is used for this purpose. Build as: , Using the Isolation Forest algorithm Anomaly detection is performed. An isolated forest consists of multiple random binary trees. Samples are progressively isolated by randomly selecting features and splitting thresholds. Anomaly samples typically have shorter average path lengths. Anomaly scores are calculated as follows: , in, For the sample Average path length in an isolated tree Let be the average path length of a binary search tree when the number of samples is n. When an anomaly occurs, the corresponding grid point or observation value is identified as an anomaly. For anomaly points, if they are inconsistent with neighboring observations and physical constraints, they are removed or replaced by a combination of neighboring reliable grid values, database response values, and observation fusion values; if anomaly points simultaneously meet the conditions of strong wind shear, terrain triggering, and multi-source observation consistency, they are retained as true local abrupt wind fields and marked with a risk level.

[0066] The wind field data obtained through anomaly detection will be fused and corrected with observations from automatic weather stations, lidar, and UAVs. Optimal interpolation or ensemble Kalman filtering methods will be used, and the correction formula is as follows: , , in, The revised forecast wind field, To analyze the wind field, For background error covariance, For observation error covariance, This is the gain matrix. The final output is the study area. The spatiotemporally continuous three-dimensional gridded low-altitude wind field within the altitude range of 0–500m above the ground, including , , Wind speed ,wind direction Vertical wind shear, turbulence intensity, and credibility indicators.

[0067] The beneficial effects of this application are: High adaptability to complex plateau terrain: This application uses features such as topographic relief, slope, aspect, valley axis direction, valley depth, and roughness to cluster and partition the terrain, and establishes response relationships for different typical terrain units. It can finely characterize the impact of local landforms such as valleys, ridges, passes, windward slopes, and leeward slopes on low-altitude wind fields.

[0068] Balancing physical mechanisms and computational efficiency: This application does not conduct high-resolution CFD simulations of the entire region in real time. Instead, it adopts a method of pre-calculation of typical terrain units and rapid database retrieval to transform complex small-scale dynamic processes into reusable local wind field response relationships, which significantly reduces the amount of computation.

[0069] It can express the dynamic and thermal coupling effect of topography: Based on the dynamic correction, this application introduces a parameterization scheme for valley wind, slope thermal convection and stability correction, and explicitly considers the influence of diurnal thermal differences, slope radiation heating, cold air leakage and low-level stable boundary layer on the vertical structure of the wind field.

[0070] Strong multi-source data fusion capability: This application integrates reanalysis data, mesoscale model output, automatic weather stations, lidar, UAV observations and remote sensing surface parameters, which can construct a continuous wind field in sparsely observed mountainous areas, while achieving dynamic correction in densely observed areas.

[0071] Stable and reliable output results: This application adopts the isolated forest algorithm and combines it with physical consistency constraints to distinguish between abnormal observations and real local sudden wind fields, which can reduce the impact of equipment failure, interpolation error and misjudgment of local extreme wind fields on the results.

[0072] It has the ability to quickly predict and expand: This application can further learn the mapping relationship from coarse resolution background field to high resolution low-altitude wind field through deep learning super-resolution module, realize the rapid reconstruction of high resolution wind field, and meet the real-time requirements of low-altitude traffic operation support.

[0073] This method can be executed by a meteorological computing server, a low-altitude operation service platform, or an edge computing node. The computing environment includes at least a data preprocessing unit, a terrain feature extraction unit, a response database retrieval unit, a terrain thermal correction unit, anomaly detection unit, and an observation fusion unit; each unit can be implemented using a general-purpose CPU, GPU, or parallel computing framework. The response relation database can be pre-calculated offline and indexed by terrain unit, inflow direction, stability, and heat flux. During the runtime phase, only retrieval, interpolation, and a small amount of matrix calculations are performed, thus meeting the operational requirement of 5-15 minute rolling updates.

[0074] Synergy between terrain clustering and CFD pre-computation database: This application first uses terrain feature clustering to summarize the infinitely diverse and disordered complex terrain of the plateau into a finite number of representative "typical terrain units," providing a physically meaningful classification basis for CFD pre-computation. Unlike directly performing high-resolution CFD simulations on the entire region, the clustering-pre-computation strategy of this invention enables the CFD simulation results of each typical unit to represent the common response patterns of different locations and inflow conditions under that type of terrain. The synergistic effect of this "classify first, then model" sequence is twofold: on the one hand, it significantly reduces the amount of CFD computation (from full-area coverage to only covering the cluster centers and virtual site areas); on the other hand, it ensures the generalization ability of the database—any grid point within the same type of terrain unit can be quickly corrected based on the response pattern of that unit, avoiding the computational explosion caused by point-by-point modeling.

[0075] Synergy between dynamic and thermal corrections: Individual CFD pre-calculation databases primarily reflect the dynamic forcing effect of topography on the background wind field (e.g., canyon acceleration, ridge bypass), but struggle to express real-time thermal effects that dynamically change with diurnal and seasonal variations (e.g., valley winds, slope thermal convection). This invention superimposes valley wind and slope wind correction terms onto the dynamic correction, creating a complementary relationship: dynamic correction provides the steady-state local wind field increment caused by topography, while thermal correction provides the time-varying circulation increment caused by differences in surface radiation. This coupling produces a "1+1>2" effect—not only improving the accuracy of the dynamic and thermal fields respectively, but more importantly, the topographic information such as valley axis direction and slope aspect relied upon by dynamic correction is precisely the directional reference required by thermal correction. Therefore, the thermal correction terms of this invention can be seamlessly superimposed on the dynamically corrected wind field, avoiding the wind field inconsistency problems that may result from separate modeling.

[0076] Synergy between Terrain Response Database and Anomaly Detection: The terrain response database is used not only for dynamic correction but also as a physical benchmark for anomaly detection. This application incorporates terrain parameters such as relief and slope into the anomaly detection feature vector, enabling the isolated forest algorithm to combine terrain information to determine whether a wind field anomaly conforms to the physical expectations under that terrain condition (e.g., strong wind shear in a canyon may be real rather than an observation error). This "database-anomaly detection" synergy mechanism effectively overcomes the shortcomings of traditional pure data-driven anomaly detection methods, which are prone to misjudgment in sparsely observed areas, and achieves adaptive differentiation between real local abrupt wind fields and observed anomalies.

[0077] Synergy of Multi-Source Fusion and Closed-Loop Process: This application utilizes multi-source observation data from automatic weather stations, lidar, and UAVs not only for the final fusion correction (step S6), but also for preliminary steps such as background field interpolation and thermal correction parameter calibration. This closed-loop data assimilation process allows observational information to influence the modeling results from the source, rather than passively correcting them only in the final step. Simultaneously, the anomaly detection step provides reliable data filtering for fusion correction, preventing anomaly observations from contaminating the analysis field. Overall, a closed-loop feedback structure is formed between the steps: "preprocessing → clustering → library construction → dynamic correction → thermal correction → anomaly detection → fusion correction → output." The output of each step provides input for the next step, and subsequent steps can correct potential biases in preceding steps.

[0078] In summary, there is a clear mutual support and progressive enhancement relationship among the various key technical features of this application: terrain clustering provides a classification foundation for database construction; the database provides an efficient retrieval method for dynamic correction; dynamic correction provides a stable background field for thermal correction; thermal correction supplements time-varying effects; anomaly detection ensures data quality; and fusion correction further improves accuracy. This multi-feature synergy enables this invention to simultaneously achieve three originally mutually restrictive technical goals—physical realism, computational efficiency, and output reliability—in low-altitude wind field modeling in plateau and mountainous areas, achieving an overall technical effect that cannot be achieved by a single feature or simple superposition.

[0079] Low-altitude wind field modeling for high-altitude canyon routes: Taking a low-altitude flight path through a canyon in a plateau mountainous area as an example, the study area covers approximately 50 km × 50 km, with an elevation of approximately 2500–4200 m, and includes deep north-south oriented river valleys, east-west oriented tributary valleys, and multiple mountain passes. The objective is to output the three-dimensional wind field at altitudes of 0, 30, 50, 100, 200, 300, and 500 m above the ground, with a temporal resolution of 10 min and a horizontal resolution of 30 m.

[0080] First, acquire 10m or 30m resolution DEM, land use data, surface roughness data, ERA5 or WRF background field data, regional automatic weather station data, wind profile data from two sets of wind-measuring lidars, and UAV flight observation data for the study area. All data are then standardized in terms of coordinates, time, and altitude; wind direction is decomposed into vectors; and extreme jump values ​​from automatic weather stations are subjected to 3σ and physical threshold tests.

[0081] Then, based on the DEM, the topographic relief, slope, aspect, valley depth, and valley axis direction were calculated. K-means was used to divide the study area into 8 topographic units, including deeply incised canyons, open valleys, windward slopes, leeward slopes, ridges, passes, plateaus, and basin edges. Two to five virtual stations were set up for each topographic unit, resulting in a total of 26 virtual stations.

[0082] For each virtual site, a CFD calculation subdomain with a radius of 2–5 km was constructed. Inflow wind direction was set from 0° to 360° at 22.5° intervals, inflow wind velocity from 2 to 20 m / s at 2 m / s intervals, and stability was categorized into unstable, neutral, and stable types. Surface heat flux was categorized into nighttime cooling, weak heating, and strong heating types. Pre-calculations were performed on these combinations, and the increment of the wind field at the target height layer relative to the background inflow within each subdomain was extracted to construct a topographic-wind field response database.

[0083] During the operational modeling phase, the WRF or reanalysis background field is interpolated to the target grid and virtual site. Based on the terrain category, inflow wind direction, wind speed, and stability of each grid point, similar samples are retrieved from the response relation database to obtain dynamic corrections. Subsequently, valley wind and slope wind corrections are calculated based on the real-time or near-real-time surface temperature change rate, solar azimuth angle, slope aspect, and valley axis direction, forming a thermodynamically coupled wind field.

[0084] Finally, based on the isolated forest algorithm, abnormal grid points are identified. Outliers that are significantly inconsistent with lidar and UAV observations and do not meet the terrain triggering conditions are removed or replaced. Strong wind shear areas supported by multiple sources are retained and their risk levels are marked. After optimal interpolation and fusion correction, a gridded low-altitude wind field and reliability product are output.

[0085] Number of terrain clusters Ideally, the contour coefficient, Calinski-Harabasz index, and geomorphic interpretation consistency should be used together to determine the optimal profile coefficient; if the study area has a relatively homogeneous topography, A score of 4-6 is acceptable; however, if the study area includes diverse landforms such as deep canyons, mountain passes, ridges, and plateau basins, A value of 8-15 is acceptable. Inverse distance weighted index. The preferred value is 1.5–3, for the terrain response database kernel width. It can be determined through cross-validation. Valley wind coefficient Slope wind coefficient Vertical attenuation height and The anomaly threshold for isolated forests can be determined through inversion from historical observations or CFD pre-calculation results. The value can be determined based on the balance between the false positive rate and the false negative rate on the validation set, with an optimal value of 0.60-0.75.

[0086] When new observation data from automatic weather stations, lidar, or UAVs are added to the study area, the background error covariance can be used as an indicator. Observation error covariance The sample weights of the terrain response database and the parameters of the deep learning super-resolution module are updated. For long-term systematic biases, the observation residuals can be statistically analyzed according to terrain categories to form a category bias correction. : , in, For the first Typical terrain units, The wind field data is obtained through actual observation methods. This is the predicted wind field output by the model.

[0087] And The output wind field is superimposed onto the corresponding terrain unit. For areas with significant seasonal thermal differences, parameter sets for warm season, cold season, rainy season, and dry season can be established separately to improve the model's adaptability under different underlying surface thermal states.

[0088] Comparative experiment: Evaluation indicators Method 1 (Reanalysis / Direct Output at Mesoscale) Method 2 (Automatic Station Inverse Distance Weighted Interpolation) Method 3 (Unified statistics across the entire region / machine learning correction) This method Wind speed RMSE (m / s) 2.85 2.31 1.92 1.24 Wind direction MAE (°) 38.6 29.4 24.7 16.5 Strong wind shear identification recall rate (%) 48% 62% 71% 89% Error rate of strong turbulence (%) 27% 21% 16% 7% Modeling time per time interval (seconds / 10km²) 0.3 0.2 0.6 1.8 To verify the superiority of the proposed method, three existing mainstream low-altitude wind field modeling methods were selected as comparative methods, and experiments were conducted on the same low-altitude observation dataset for plateau and mountainous areas. The dataset included 10-minute wind observations from automatic weather stations, lidar wind profile observations, and UAV flight path wind field observations, with an evaluation altitude range of 0-500m above the ground.

[0089] The comparative methods include: Method 1, directly using reanalysis data or mesoscale model output; Method 2, inverse distance weighted interpolation based on automatic weather stations; Method 3, statistical correction or machine learning correction methods for unified parameters across the entire region; the method in this application combines multi-source fusion, topographic response database, and topographic dynamic and thermal correction. Evaluation metrics include root mean square error (RMSE) of wind speed, mean absolute error (MAE) of wind direction, strong wind shear identification recall rate, and modeling time per time interval.

[0090] To ensure the verifiability of the technical effects, during implementation, the comparative method and the method of this application can be run separately on the same low-altitude observation dataset of plateau and mountainous areas. Evaluations are based on wind speed RMSE, wind direction MAE, strong wind shear recognition recall rate, strong turbulence false alarm rate, and single-time modeling time. If the terrain response database and observation fusion parameters used in this application have been calibrated using historical or pre-calculated samples, the background field systematic error in complex terrain areas can be reduced while maintaining computational timeliness, and a confidence indicator for each grid point can be output. Evaluation indicators are recorded by the system and stored in association with the corresponding time slices, terrain units, and observation samples.

[0091] The embodiments described above merely illustrate specific implementation methods of this application, and while the descriptions are detailed and specific, they should not be construed as limiting the scope of protection of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the technical solution of this application, and these modifications and improvements all fall within the scope of protection of this application.

Claims

1. A method for modeling low-altitude wind fields in plateau and mountainous areas based on a terrain response relation database, characterized in that, Includes the following steps: High-resolution digital elevation models, surface roughness, land use, meteorological background field and low-altitude observation data of the study area were acquired, and coordinate, time and height layer consistency preprocessing was performed. Based on the digital elevation model, terrain feature parameters are calculated, and the study area is divided into multiple typical terrain units using a clustering algorithm; Virtual stations were set up for each typical terrain unit, and computational fluid dynamics pre-calculations were performed under different inflow conditions to build a database of terrain-wind field response relationships. Using the meteorological background field as the background wind field, interpolation is performed to the target grid point, and similar samples are retrieved from the database based on the typical terrain unit to which the target grid point belongs, to perform local dynamic correction of the background wind field; Based on the surface temperature change rate, solar radiation, slope aspect, valley axis direction, valley depth and atmospheric stability, valley wind correction terms and slope thermal convection correction terms are constructed to perform topographic thermal correction on the wind field after dynamic correction. Anomaly detection feature vectors are constructed for the thermally corrected wind field. Anomaly detection is performed using the isolated forest algorithm, and low-altitude observation data is combined for fusion correction to output a gridded low-altitude wind field within the range of 0–500 m above the ground. The local dynamic correction uses a kernel-weighted method to retrieve the correction amount from the database: ， in, Let this be the current state vector. For the first in the database The state vector of a pre-computed sample For kernel width parameter, To pre-calculate the total sample size, For kernel weight function; The valley wind correction term is applied along the valley axis, and its intensity is related to the rate of change of surface temperature, valley depth, and atmospheric stability, as expressed as: ， in, The empirical coefficient for valley winds. Because the valley is deep, For stability, The vertical attenuation height of valley winds. The unit vector along the valley axis. For surface temperature, To prevent stable terms with a denominator of zero, Height above the ground; The slope heat convection correction term is applied along the uphill or downhill direction of the slope, and its intensity is related to the surface sensible heat flux, slope, aspect heat transfer factor, and atmospheric stability, and is expressed as: ， in, This is the empirical coefficient for slope wind. For surface sensible heat flux, For slope, Slope direction, The azimuth of the sun. air density, The specific heat capacity of air at constant pressure. For reference potential temperature, The vertical attenuation height of the slope wind. This is a unit vector along the slope direction; The dynamically corrected wind field is the sum of the background wind field and the local dynamic correction, specifically: ,in The background wind field after interpolation. For local dynamic corrections, For spatial location, Height above the ground For time; Terrain thermal correction is obtained by coupling dynamic correction terms with thermal correction terms, specifically as follows: .

2. The method for modeling low-altitude wind fields in plateau and mountainous areas based on a terrain response relation database as described in claim 1, characterized in that, The anomaly detection feature vector includes: east-west wind speed. North-south wind speed Vertical velocity Wind speed ,wind direction Vertical wind shear Turbulent kinetic energy, Richardson number Topographic relief ,slope and surface roughness Anomaly score threshold for the Isolation Forest algorithm Using a value of 0.60 to 0.75, for points identified as anomalous, a secondary discrimination is performed by combining the consistency of neighboring observations and terrain triggering conditions, preserving the true abrupt wind field and outputting risk indicators.

3. The method for modeling low-altitude wind fields in plateau and mountainous areas based on a terrain response relation database as described in claim 1, characterized in that, The fusion correction employs optimal interpolation or ensemble Kalman filtering methods, and the correction formula is as follows: in, The revised forecast wind field, To analyze the wind field, For observation data, For background error covariance, For observation error covariance, Here is the gain matrix. For the observation operator, This is the transpose of the observation operator.

4. The method for modeling low-altitude wind fields in plateau and mountainous areas based on a terrain response relation database as described in claim 1, characterized in that, The background wind field interpolation uses an inverse distance weighted method, with interpolation weights... ,in For the grid points to be estimated, For background grid points or virtual sites, For spatial distance, As a stable term, The distance attenuation index ranges from 1.5 to 3.

5. The method for modeling low-altitude wind fields in plateau and mountainous areas based on a terrain response relation database as described in claim 1, characterized in that, It also includes wind field super-resolution reconstruction based on deep learning: Using coarse-resolution background wind field, digital elevation model terrain features, surface roughness, terrain category coding, and historical corrected wind field as inputs, U-Net, residual convolutional network, or ConvLSTM are used to output high-resolution 3D wind field; the training loss function includes wind speed reconstruction error, wind direction error, mass conservation constraint, and observation consistency constraint.

6. The method for modeling low-altitude wind fields in plateau and mountainous areas based on a terrain response relation database according to claim 1, characterized in that, Each entry in the topography-wind field response database includes: typical topographic unit characteristics, inflow wind profile, inflow wind direction, atmospheric stability, surface heat flux, roughness length, and three-dimensional wind field response increment relative to the background wind field; computational fluid dynamics pre-calculations employ Reynolds-averaged Navier-Stokes equations combined with... Turbulence model.

7. The method for modeling low-altitude wind fields in plateau and mountainous areas based on a terrain response relation database as described in claim 1, characterized in that, The terrain feature parameters include elevation. Topographic relief ,slope Slope aspect sinusoidal component cosine component of slope curvature Valley Ridge Index Sine component along the valley axis, cosine component along the valley axis, and roughness length Clustering algorithm is Number of clusters It is determined by both the profile coefficient and the elbow method.

8. The method for modeling low-altitude wind fields in plateau and mountainous areas based on a terrain response relation database as described in claim 1, characterized in that, The gridded low-altitude wind field within the output range of 0-500m above the ground includes east-west wind speeds at multiple altitude levels within this range. North-south wind speed Vertical velocity Wind speed ,wind direction Vertical wind shear, turbulence intensity, and reliability indicators; low-altitude observation data includes one or more of the following: automatic weather station observation data, wind profile data from wind-measuring lidar, and UAV observation data.

Images