A Method for Optimizing the Layout of Measuring Points for 3D Wind Field Inversion Based on CFD and Artificial Intelligence
By combining CFD and artificial intelligence, a lightweight proxy model and adaptive optimization algorithm were constructed, which solved the problems of ill-conditioned inversion and large computational load in traditional radar spotting methods, and achieved accurate inversion and stable observation of three-dimensional wind fields in complex environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUANGZHOU UNIVERSITY
- Filing Date
- 2026-06-29
- Publication Date
- 2026-07-31
AI Technical Summary
In complex terrain and dense urban building complexes, traditional radar deployment methods rely excessively on geometric coverage, leading to ill-conditioned inversion equations, omission of core aerodynamic features, excessive computational load, and reduced engineering value, while lacking an adaptive online closed-loop mechanism.
A lightweight surrogate model with physical constraints is constructed by combining CFD and artificial intelligence. The optimal point combination is found by using a divergence penalty term and a regularized inversion cost function, combined with an intelligent optimization algorithm, and then online adaptive correction is performed during field implementation.
It achieves stable reconstruction of three-dimensional wind fields in complex environments, enhances the engineering value of observation data, reduces computational costs, avoids matrix ill-conditioning and observation blind spots, and improves inversion accuracy.
Smart Images
Figure CN122490754A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wind engineering technology, and in particular to a method for optimizing the layout of three-dimensional wind field inversion measurement points based on CFD and artificial intelligence. Background Technology
[0002] With the continuous development of structural wind engineering and wind disaster prevention technologies, research on the three-dimensional wind field characteristics of complex terrains (such as mountains and canyons) and high-density urban building clusters is becoming increasingly in-depth. Accurately acquiring detailed three-dimensional wind field data within complex underlying surfaces is of paramount practical significance for the wind-resistant design of high-rise buildings, wind field assessment, and the exploration of near-surface flow field evolution patterns during extreme wind disasters (such as strong typhoons). Currently, coordinated observation by multiple Doppler laser wind radars has become a key method for obtaining large-scale three-dimensional wind fields.
[0003] However, when deploying radar networks and inverting wind fields under complex three-dimensional spatial constraints, the existing technologies have the following technical bottlenecks that are difficult to overcome: (1) There is a serious bias in the evaluation orientation. The traditional point deployment method excessively pursues spatial geometric coverage and ignores the stability of the inversion mathematical model, which easily induces ill-conditioned distortion of the inversion equation; (2) Core aerodynamic features are easily missed. The traditional method does not consider key flow field areas such as the backflow zone and acceleration zone, which are crucial to engineering analysis, thus reducing the engineering value of the observation data; (3) The traditional computing power bottleneck completely restricts the global joint optimization of "inversion orientation". Each iteration requires calling high-precision CFD simulation, which makes the computational load unbearable; (4) There is a lack of an adaptive online closed-loop mechanism to cope with the real engineering environment of the field. There is a deviation between the ideal model and the actual environment, which leads to a decrease in observation efficiency after deployment.
[0004] In summary, there is an urgent need in this field to propose a novel optimization scheme for wind measurement radar deployment in order to break through the computing power bottleneck, solve the ill-conditioned inversion problem under complex obstruction, and achieve a technological leap from the traditional "spatial coverage-oriented" approach to the modern "inversion accuracy and matrix stability-oriented" approach. Summary of the Invention
[0005] The purpose of this invention is to provide a method for optimizing the layout of measurement points for three-dimensional wind field inversion based on CFD and artificial intelligence, so as to overcome the shortcomings of traditional radar point layout methods that rely too much on geometric coverage and face the bottleneck of computing power in forward modeling optimization, and solve the ill-conditioned problem of inversion equations caused by complex terrain obstruction.
[0006] To achieve the above objectives, this invention provides a method for optimizing the layout of three-dimensional wind field inversion measurement points based on CFD and artificial intelligence, comprising the following steps: S1: Establish a three-dimensional digital scene model of the survey area and generate a candidate survey point library with attribute tags.
[0007] Specifically, the basic spatial information of the area to be measured is first acquired, including topographic elevation data, building outlines and height data, surface roughness type, and distribution of vegetation and obstacles. A three-dimensional digital scene model of the area is then established in a unified coordinate system. After the three-dimensional scene model is established, the locations where laser wind radars can theoretically be deployed within the area are initially discretized to form a set of candidate measurement points. For each candidate measurement point, its engineering and geometric attributes are further recorded, resulting in a candidate measurement point library with complete attribute labels, which serves as input for subsequent optimization.
[0008] S2: Use CFD calculations to build a wind field sample library containing different incoming flow conditions, and extract key flow field regions such as the recirculation zone and acceleration zone.
[0009] Specifically, computational fluid dynamics (CFD) numerical simulations are conducted on the three-dimensional digital scene model of the survey area established in step S1 to construct a wind field sample library containing complex flow characteristics. Based on long-term meteorological statistics or design conditions of the survey area, the incoming flow condition parameter space is determined, including the inlet wind direction (e.g., discretized at 10° or 15° intervals), inlet wind speed level, turbulence intensity level, and atmospheric boundary layer wind speed profile parameters. CFD calculations are then performed sequentially on multiple sets of typical operating conditions after discretization. The computational platform can be Fluent or OpenFOAM, and the Realizable k-ε model or SST k-ω model is preferred for the turbulence model. After the calculation is completed, the three-dimensional wind speed vector within the grid cell is output (…). , , Subsequently, based on threshold determination and flow feature extraction algorithms, areas with increased local wind speeds are marked as acceleration zones, areas with large velocity gradients are marked as shear layers, and areas with changes in the sign of near-wall backflow velocity are marked as backflow zones. After these processes, a high-fidelity CFD sample library is formed, encompassing "incoming flow conditions—three-dimensional wind field ground truth—key region markings."
[0010] S3: Combine equipment parameters to perform ray tracing and establish observation geometric models and visible field models for each candidate measurement point.
[0011] Specifically, based on the candidate measurement point set obtained in step S1, and combined with the radar equipment performance parameters, an observation geometry model and a visible field model for each candidate measurement point in the 3D scene are established. The maximum / minimum detection range, azimuth and elevation scanning range, and range gate length of the radar are set. Then, in the 3D digital scene of the survey area, a ray tracing algorithm is used to spatially traverse the laser beams emitted from each candidate measurement point to determine whether the line of sight is obstructed by terrain or buildings, thereby outputting the set of unobstructed effective lines of sight (i.e., the visible field result) for that measurement point to the spatial grid cells. Based on this, for different scanning modes such as VAD, DBS, PPI, or RHI, the angle between the laser scanning direction and the 3D wind speed vector on the effective observation grid is calculated, generating the corresponding direction cosine vector and local observation matrix.
[0012] S4: Use a CFD sample library to train a lightweight wind field proxy model with physical constraints to achieve rapid prediction of wind speed field and key area labels.
[0013] Since using high-precision CFD in the site optimization iteration would lead to excessive computation, this step constructs a fast wind field proxy model with physical constraints to replace part of the CFD calculations. Using the CFD sample library constructed in step S2 as training data, a mapping model combining intrinsic orthogonal decomposition and neural networks is established. To ensure that the predicted wind field satisfies the law of mass conservation in hydrodynamics, a loss function incorporating divergence physical constraints is designed during model training. : ; In the formula, This is the predicted three-dimensional wind speed field vector; For CFD wind field true values; This is the divergence term, used for physical consistency penalty; The classification cross-entropy loss function is used for key regions of the flow field (such as the separation zone and the recirculation zone). , , These are the weighting coefficients. This model can quickly output three-dimensional wind field prediction results and labels for key flow field regions under given inflow conditions.
[0014] S5: Generation of simulated observations based on combinations of candidate measurement points.
[0015] For any given combination of measuring points and incoming flow conditions, the proxy model constructed in step S4 is invoked to quickly generate a three-dimensional wind field background. The radial velocity on each observed grid is calculated using the observation geometry model from step S3, and a measurement error model including random noise and pointing error is superimposed to generate a simulated radial velocity observation set consistent with the actual instrument output. .
[0016] S6: Construct a complex three-dimensional wind field inversion model that includes background field constraints and divergence regularization terms, and perform inversion calculations on the point layout scheme.
[0017] For complex terrain and high-density building complexes, observation blind spots can lead to ill-conditioned inversion equations. This step constructs a complex 3D wind field inversion model based on variational assimilation / Tikhonov regularization, whose cost function... The expression is as follows: ; In the formula, This is the local observation matrix determined by the scanning azimuth and elevation angles; The simulated radial velocity obtained in step S5; The background flow field output by the surrogate model in step S4 is used as prior compensation for the under-observed region. and For regularization parameters; This is the weight matrix determined by the observed signal-to-noise ratio. Solve this equation and output the 3D inverted wind field results corresponding to the current point layout scheme.
[0018] S7: Establish a comprehensive objective function that includes inversion weighting error, local observation matrix condition number, and engineering feasibility.
[0019] Establish an evaluation index system for complex 3D wind field inversion, evaluating not only the coverage area but also the direct contribution of the sampling point layout to the inversion task. Construct a comprehensive fitness function (objective function). Comprehensive evaluation of candidate solutions Quality: ; In the formula, As the inversion error term, the error between the inverted wind field and the true value is calculated for the whole field, and high weights are assigned to key areas such as strong shear layer and recirculation zone; It is the condition number or singular value stability index of the local observation matrix, used to penalize ill-conditioned point placement with collinear observation directions or excessively small intersection angles; These are engineering constraints and penalties that include the number of equipment, site accessibility, and power and communication levels. , , The weighting coefficients are assigned.
[0020] S8: A discrete intelligent optimization algorithm is used for two-level search to output the optimal point combination and scanning parameters.
[0021] Discrete intelligent optimization algorithms (such as NSGA-II or Discrete Particle Swarm Optimization) are employed for optimization. A two-stage search strategy is designed: the first stage uses a simplified evaluation function for rapid coarse screening of a large number of candidate combinations; the second stage calls the complete simulation observation and regularized inversion process (steps S5-S6) to perform high-precision fine evaluation on the preferred solution. The algorithm encoding includes "radar number combination + scan mode number combination", and outputs a non-dominated solution set or a single optimal solution after iteration.
[0022] S9: Utilize high-precision CFD to perform fine verification and parameter solidification of the optimal solution.
[0023] After obtaining the optimal or suboptimal combination of measurement points through intelligent optimization algorithms, high-precision verification is essential to ensure engineering reliability. This step departs from the rapid proxy model and re-invokes the high-precision CFD ground truth from step S2 and the complete regularized inversion process from step S6 to finely verify the selected preferred scheme. During the verification process, a more refined wind field grid and a more stringent instrument measurement error model (including signal-to-noise ratio attenuation and distance pointing error) are used to repeatedly execute simulation observations and inversions under prevailing wind directions and extremely unfavorable conditions. If the preferred scheme meets the preset standards for inversion error in key areas, overall root mean square error, and matrix condition number, the parameters of the scheme are solidified.
[0024] S10: In the initial stage of the experiment, the surrogate model was calibrated online using measured data, and the scanning parameters were adaptively re-optimized.
[0025] After the radar equipment completes its physical deployment in the actual field and begins initial observations, an online dynamic correction mechanism is activated to compensate for the deviation between the ideal model and the real environment. In the initial stage of the observation experiment, real radial velocity echo data of a certain period are collected and compared with the predicted observations of the surrogate model under the same inflow conditions as described in the above implementation steps. If the actual effective data acquisition rate at a certain measuring point is found to be significantly lower than expected, or the inversion residual is too large, then this batch of measured small sample data is used to perform transfer learning and parameter fine-tuning on the wind field surrogate model based on physical constraints. Simultaneously, based on the corrected flow field and the actual obstruction situation, while keeping the radar's physical position unchanged, a lightweight re-optimization calculation is initiated to adaptively adjust the scanning sector, elevation range, and scanning period of each radar, enabling the system to regain optimal observation performance under real field conditions.
[0026] Preferably, the discrete intelligent optimization algorithm in step S8 adopts a two-level search strategy: the first level uses a simplified evaluation function to quickly screen the candidate measurement point combinations; the second level calls the complete simulation observation and regularized inversion process to perform high-precision evaluation on the selected scheme after screening.
[0027] Preferably, the key flow field regions in step S2 include the acceleration zone, shear layer, and recirculation zone marked by threshold determination and flow feature extraction algorithms; the parameter space of the CFD numerical simulation for various incoming flow conditions includes inlet wind direction, inlet wind speed level, turbulence intensity level, and atmospheric boundary layer wind speed profile parameters.
[0028] Preferably, the online adaptive correction in step S10 specifically includes: collecting radial velocity echo data measured in the field, comparing it with the predicted values of the lightweight wind field proxy model under the same incoming flow conditions, performing transfer learning and parameter fine-tuning on the proxy model based on the comparison deviation, and re-optimizing the radar's scanning sector, elevation range and scanning period based on the corrected proxy model while keeping the radar's physical position unchanged.
[0029] Preferably, the process of establishing the observation geometry model and the visible field model in step S3 includes: setting the maximum / minimum detection range, azimuth and elevation scanning range, and range gate length of the radar; using a ray tracing algorithm to traverse the laser beams emitted from the candidate measurement point, determining whether the line of sight is blocked by terrain or buildings, and outputting the set of unobstructed effective lines of sight for the measurement point; and calculating the direction cosine vector and local observation matrix on the effective observation grid according to the scanning mode.
[0030] Preferably, the fine verification in step S9 uses a more refined wind field grid and a more stringent instrument measurement error model than the CFD sample library. The simulation observation and inversion are repeatedly performed under the prevailing wind direction and extreme unfavorable conditions. When the inversion error and matrix condition number meet the preset standards, the point layout scheme is solidified.
[0031] Preferably, after step S10, the method further includes: encapsulating the calculation, analysis and correction results of the entire process into an engineering implementation data package, which includes a three-dimensional visualization model of the survey area, the final coordinates of the radar distribution points, an adaptive scanning control script, and an evaluation report on the expected coverage and inversion accuracy of key complex flow areas.
[0032] Therefore, the present invention employs a three-dimensional wind field inversion measurement point optimization arrangement method based on CFD and artificial intelligence, which has the following beneficial effects: (1) This invention breaks through the limitation of traditional radar inversion relying solely on measured data. It innovatively uses the full-field predicted wind speed generated by the physical constraint proxy model as the prior background field and introduces it into the regularized inversion cost function with a divergence penalty term. This mechanism suppresses the ill-conditioned amplification effect of the matrix induced by the limited radar line of sight and the small angle between multiple source beams from a mathematical perspective. It successfully achieves stable reconstruction of three-dimensional wind speed in the blind spot and under-observed areas, and solves the problem of inversion distortion and unsolvable problems caused by complex and towering obstacles.
[0033] (2) This invention abandons the single "maximizing line-of-sight coverage" evaluation standard commonly used in the industry and constructs a multi-objective optimization fitness function for inversion stability. By forcibly incorporating the condition number index into the local observation matrix, it effectively avoids the poor placement of multiple radars in collinear detection; at the same time, it assigns high weight penalties to key aerodynamic regions such as separation zone, recirculation zone and strong shear layer, ensuring that limited radar resources are prioritized for capturing the core features of the flow field, and significantly improving the engineering value of the observation data.
[0034] (3) In the training stage of the proxy model, the present invention deeply integrates the three-dimensional wind speed field divergence penalty term and the key area classification label to construct a lightweight wind field prediction model with embedded fluid dynamics constraints. While maintaining a similar accuracy to high-precision CFD simulation, the model reduces the time of a single wind field forward modeling by several orders of magnitude, making it possible to perform global intelligent optimization of more than 100,000 discrete point combinations under limited computing power; and the predicted flow field strictly satisfies the law of mass conservation, avoiding the non-physical "source and sink point" fallacy that is easily generated by pure data-driven models.
[0035] 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
[0036] Figure 1 This is an overall flowchart of a three-dimensional wind field inversion measurement point optimization layout method based on CFD and artificial intelligence according to the present invention. Detailed Implementation
[0037] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.
[0038] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.
[0039] Example Please see Figure 1This embodiment provides a method for optimizing the layout of measuring points for three-dimensional wind field inversion based on CFD and artificial intelligence, which includes the following detailed steps: S1. Construction of 3D Scene of Survey Area and Initial Generation of Candidate Measurement Points: First, basic spatial information of the area to be measured is acquired, including topographic elevation data, building outlines and height data, surface roughness type, and distribution of vegetation and obstacles. A three-dimensional digital scene model of the area is then established in a unified coordinate system. After the three-dimensional scene model is established, the locations where laser wind radar can theoretically be deployed within the area are initially discretized to form a set of candidate measurement points. For each candidate measurement point, its engineering attributes (such as ease of power supply and communication, site use rights, etc.) and geometric attributes (such as coordinates and altitude) are further recorded, resulting in a candidate measurement point library with complete attribute labels, which serves as input for subsequent optimization.
[0040] S2. Construction of a complex 3D wind field CFD sample library: For the three-dimensional digital scene model of the survey area established in step S1, computational fluid dynamics (CFD) numerical simulations are conducted to construct a wind field sample library containing complex flow characteristics. Specifically, based on long-term meteorological statistics or design conditions of the survey area, the incoming flow condition parameter space is first determined. This parameter space includes the inlet wind direction (e.g., discretized at 10° or 15° intervals), inlet wind speed level, turbulence intensity level, and atmospheric boundary layer wind speed profile parameters. CFD calculations are then performed sequentially on multiple sets of typical operating conditions after discretization. The computational platform can be Fluent or OpenFOAM, and the turbulence model preferably uses the Realizable k-ε model or the SST k-ω model. After the calculation is completed, the three-dimensional wind speed vector within the grid cell is output. , , Subsequently, based on threshold determination and flow feature extraction algorithms, regions with local wind speeds greater than 1.2 times the incoming wind speed are marked as acceleration zones, regions with velocity gradients greater than a specific threshold are marked as shear layers, and regions with near-wall velocity vectors opposite to the mainstream direction are marked as recirculation zones. After these processes, a high-fidelity CFD sample library is formed, covering "incoming flow conditions—three-dimensional wind field ground truth—key region markings," providing a data foundation for subsequent surrogate model training.
[0041] S3. Establishment of the geometric model and field-of-view model for laser wind measurement radar observation: Based on the candidate measurement point set obtained in step S1, and combined with radar equipment performance parameters such as laser wavelength, pulse repetition frequency, and beam divergence angle, an observation geometric model and a visible field model for each candidate measurement point in the three-dimensional scene are established. First, the maximum / minimum detection range of the radar (e.g., 50m-5km), the azimuth and elevation scanning range (e.g., azimuth 0-360°, elevation 0-30°), and the range gate length (e.g., 50m) are set. Then, in the three-dimensional digital scene of the measurement area, a ray tracing algorithm is used to spatially traverse the laser beams emitted from each candidate measurement point, determining whether the line of sight intersects with the terrain or building surface, thereby outputting the set of unobstructed effective lines of sight (i.e., the visible field result) for that measurement point to the spatial grid cells. Based on this, for different scanning modes such as VAD, DBS, PPI, or RHI, the angle between the laser scanning direction and the three-dimensional wind speed vector on the effective observation grid is calculated, generating the corresponding direction cosine vector and local observation matrix. This step outputs the observable grid list, the visible proportion and occlusion proportion of key areas for each candidate measurement point, providing geometric input for subsequent evaluation of the stability of the observation matrix.
[0042] S4. Construction of a lightweight wind field proxy model based on physical information constraints: Since calling high-precision CFD in the site optimization iteration would lead to excessive computation, this embodiment constructs a fast wind field proxy model with physical constraints to replace part of the CFD calculations. Using the CFD sample library constructed in step S2 as training data, a mapping model combining intrinsic orthogonal decomposition and a neural network is established. First, the CFD samples are decomposed into POD to extract the main modes. Then, a neural network is trained to learn the mapping from incoming flow conditions to mode coefficients.
[0043] To ensure that the predicted wind field satisfies the law of mass conservation in hydrodynamics, a loss function L containing divergence physical constraints is designed during model training: ; In the formula, This is the predicted three-dimensional wind speed field vector; For CFD wind field true values; This is the divergence term, used for physical consistency penalty; The classification cross-entropy loss function is used for key regions of the flow field (such as the separation zone and the recirculation zone). , , where is the weighting coefficient. By minimizing this loss function, the model not only learns to fit the CFD data during training, but also forces the prediction results to satisfy the law of conservation of mass ( ). (Or it may conform to the divergence characteristics of compressible fluids). This model can quickly output three-dimensional wind field prediction results and labels of key flow field regions under given inflow conditions.
[0044] S5. Generation of simulated observations based on candidate measurement point combinations: For any given combination of measuring points and incoming flow conditions, the proxy model constructed in step S4 is invoked to quickly generate a three-dimensional wind field background. The radial velocity on each observed grid is calculated using the observation geometry model from step S3, and a measurement error model including random noise and pointing error is superimposed to generate a simulated radial velocity observation set consistent with the actual instrument output. .
[0045] S6. Construction and computation of regularized inversion model with prior background field: For complex terrain and high-density building complexes, observation blind spots can lead to ill-conditioned inversion equations. This embodiment constructs a complex three-dimensional wind field inversion model based on variational assimilation / Tikhonov regularization, whose cost function... The expression is as follows: ; In the formula, This is the local observation matrix determined by the scanning azimuth and elevation angles; The simulated radial velocity obtained in step S5; The background flow field output by the surrogate model in step S4 is used as prior compensation for the under-observed region. and For regularization parameters; This is the weight matrix determined by the observed signal-to-noise ratio. Solve this equation and output the 3D inverted wind field results corresponding to the current point layout scheme.
[0046] S7. Construction of a multi-objective optimization fitness function for inversion stability: Establish an evaluation index system for complex 3D wind field inversion, evaluating not only the coverage area but also the direct contribution of the sampling point layout to the inversion task. Construct a comprehensive fitness function (objective function). Comprehensive evaluation of candidate solutions Quality: ; In the formula, As the inversion error term, the error between the inverted wind field and the true value is calculated for the whole field, and high weights are assigned to key areas such as strong shear layer and recirculation zone; It is the condition number or singular value stability index of the local observation matrix, used to penalize ill-conditioned point placement with collinear observation directions or excessively small intersection angles; These are engineering constraints and penalties that include the number of equipment, site accessibility, and power and communication levels. , , These are the weighting coefficients for allocation. Simultaneously, the performance of indicators under multiple operating conditions is statistically analyzed to ensure the robustness of the solution.
[0047] S8. Measurement point combination search based on intelligent optimization algorithm: Discrete intelligent optimization algorithms (such as NSGA-II or Discrete Particle Swarm Optimization) are employed for optimization. A two-level search strategy is designed: the first level uses a simplified evaluation function for rapid coarse screening of a large number of candidate combinations; the second level calls the complete simulation observation and regularized inversion process (steps S5-S6) to perform high-precision fine evaluation on the preferred solution. The algorithm encoding includes "radar number combination + scan mode number combination", and outputs a non-dominated solution set or a single optimal solution after iteration.
[0048] S9. Detailed verification and parameter solidification of the optimal point layout scheme: After obtaining the optimal or suboptimal measurement point combination scheme through the intelligent optimization algorithm (step S8), high-precision verification must be performed to ensure engineering reliability. This step departs from the rapid proxy model and re-invokes the high-precision CFD ground truth from step S2 and the complete regularized inversion process from step S6 to finely verify the selected preferred scheme. During the verification process, a more refined wind field grid and a more stringent instrument measurement error model (including signal-to-noise ratio attenuation and range pointing error) are used to repeatedly execute simulated observations and inversions under the prevailing wind direction and extremely unfavorable conditions. If the preferred scheme meets the preset standards for inversion error in key areas, overall field mean square error, and matrix condition number, the parameters of the scheme are solidified. A detailed point deployment implementation table is output, specifying in detail the precise installation coordinates, initial orientation, scanning mode, azimuth / elevation range, range gate setting, and expected inversion accuracy level for each radar, and a backup plan can be included to deal with unforeseen circumstances on site.
[0049] S10. Online adaptive correction and final optimization result output in the field: After the radar equipment completes its actual field physical deployment and begins initial observations, an online dynamic correction mechanism is activated to compensate for the deviation between the ideal model and the real environment. In the early stages of the observation experiment, real radial velocity echo data of a certain period are collected and compared with the predicted observations of the surrogate model under the same incoming flow conditions in the above implementation steps.
[0050] If the actual effective data acquisition rate at a certain measuring point is found to be significantly lower than expected (e.g., due to the presence of unmodeled temporary obstructions), or the inversion residual is too large, then this batch of measured small sample data is used to perform transfer learning and parameter fine-tuning on the wind field proxy model based on physical constraints. Simultaneously, based on the corrected flow field and the actual obstruction situation, while keeping the radar's physical location unchanged, a lightweight re-optimization calculation is initiated to adaptively adjust the scanning sector, elevation angle range, and scanning period of each radar, enabling the system to regain optimal observation performance under real field conditions.
[0051] After the above online adaptive correction, the calculation, analysis, and correction results of the entire process are uniformly packaged and output as an engineering implementation data package that can directly guide field testing. The output comprehensively covers: a 3D visualization model of the test area, the final coordinates of the radar distribution points, a refined configuration of the radar adaptive scanning control script, the expected coverage and inversion accuracy assessment of key complex flow areas (such as the wake region and separation region), and a robustness evaluation report for multiple operating conditions.
[0052] Through this complete closed loop, the present invention realizes a technological paradigm shift from simply relying on human experience and pursuing spatial coverage to "addressing the inversion accuracy and matrix stability of complex wind fields," effectively solving the industry pain points of observation distortion and ill-conditioned equations under complex three-dimensional terrain with limited equipment.
[0053] Working Principle: This invention first constructs a high-fidelity wind field sample library through CFD simulation and trains a lightweight surrogate model with embedded divergence physical constraints from it. This model can quickly predict the three-dimensional wind field under different incoming flow conditions with extremely low computational cost. In the site selection optimization stage, this invention abandons the traditional "maximum coverage" approach and instead constructs a comprehensive objective function directly oriented towards inversion stability. This function simultaneously considers inversion error, observation matrix condition number (avoiding ill-conditioned site selection), and engineering implementation costs. A two-level intelligent search algorithm is used to efficiently find the optimal solution within a vast candidate point combination space. After obtaining the optimal site selection scheme, it is verified and solidified using high-precision CFD. After actual field deployment, the system can also use measured data to perform online transfer learning on the surrogate model and adaptively adjust radar scanning parameters, forming a closed-loop optimization process. The entire solution fundamentally solves the ill-conditioned problem of radar inversion in complex environments and overcomes computational limitations.
[0054] Therefore, the present invention adopts the above-mentioned method for optimizing the layout of three-dimensional wind field inversion measurement points based on CFD and artificial intelligence, realizing a technological leap from "blind spatial coverage" to "precise inversion guidance" in radar measurement point configuration.
[0055] 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 optimizing the layout of measuring points for three-dimensional wind field inversion based on CFD and artificial intelligence, characterized in that, Includes the following steps: S1: Obtain basic information of the survey area, establish a three-dimensional digital scene model, and generate a candidate survey point library with attribute tags discretely in the three-dimensional digital scene model; S2: Based on a three-dimensional digital scene model, CFD numerical simulations are performed on various incoming flow conditions to construct a CFD sample library containing a three-dimensional wind speed vector field and key flow field region markings. S3: Based on the candidate measurement point library and combined with the parameters of the laser wind radar equipment, the observation geometric model and the visible field model of each candidate measurement point are established through the ray tracing algorithm. S4: Train a lightweight wind field surrogate model with embedded physical constraints using a CFD sample library. The physical constraints include a divergence penalty term, which is used to ensure that the predicted wind field output by the surrogate model satisfies the law of mass conservation. S5: For any given combination of candidate measurement points and incoming flow conditions, call the lightweight wind field proxy model to generate the background flow field, and combine it with the observation geometry model to generate a simulated radial velocity observation set; S6: Construct a three-dimensional wind field inversion model with prior background field constraints and divergence regularization term. Use the simulated radial velocity observation set and background flow field to perform inversion calculations on the current candidate measurement point combination to obtain the inverted wind field. S7: Construct a comprehensive objective function for inversion stability, which includes an inversion error term, a local observation matrix condition term, and an engineering constraint penalty term; S8: A discrete intelligent optimization algorithm is adopted, with the comprehensive objective function as the fitness function, to perform a global optimization search on the combination of measurement points in the candidate measurement point library and output the optimal combination of points; S9: Use a high-precision CFD model to perform a fine verification of the optimal point combination, and solidify the point layout scheme and radar scanning parameters after the verification is passed; S10: In the initial stage of field deployment, the lightweight wind field proxy model is adaptively corrected online using measured data, and the radar scanning parameters are re-optimized based on the corrected model.
2. The method for optimizing the layout of three-dimensional wind field inversion measuring points based on CFD and artificial intelligence according to claim 1, characterized in that, The lightweight wind field proxy model constructed in step S4 is a mapping model combining intrinsic orthogonal decomposition and neural networks. Its training loss function L is designed as a combination of data fitting terms, divergence physical constraint terms, and key region classification terms. ； in, This is the predicted three-dimensional wind speed field vector. For the true value of the CFD wind field, To predict wind field divergence, The classification cross-entropy loss function for key regions of the flow field. , , These are the weighting coefficients. Mark the key areas for prediction. Mark the real key areas.
3. The method for optimizing the layout of measuring points for three-dimensional wind field inversion based on CFD and artificial intelligence according to claim 1, characterized in that, The three-dimensional wind field inversion model constructed in step S6, with prior background field constraints and divergence regularization term, has a cost function... Specifically, it can be expressed as follows: ； in, This is a local observation matrix determined by the radar scanning azimuth and elevation angles. To simulate the radial velocity observation set, The background flow field is output by the lightweight wind field proxy model. and For regularization parameters, The weight matrix is determined by the observed signal-to-noise ratio. The vector is the three-dimensional wind speed field to be inverted.
4. The method for optimizing the layout of three-dimensional wind field inversion measuring points based on CFD and artificial intelligence according to claim 1, characterized in that, The comprehensive objective function for inversion stability constructed in step S7 Specifically, it can be expressed as follows: ； in, This is the inversion error term, used to calculate the error between the inverted wind field and the true value, and assigns high weights to key regions of the flow field; The condition number of the local observation matrix is used to evaluate the ill-conditioned nature of the inversion equation; This is a penalty item for engineering constraints; , , These are the weighting coefficients.
5. The method for optimizing the layout of three-dimensional wind field inversion measuring points based on CFD and artificial intelligence according to claim 1, characterized in that, The discrete intelligent optimization algorithm in step S8 adopts a two-level search strategy: the first level uses a simplified evaluation function to quickly screen the candidate measurement point combinations; the second level calls the complete simulation observation and regularized inversion process to perform high-precision evaluation on the optimized scheme after screening.
6. The method for optimizing the layout of three-dimensional wind field inversion measuring points based on CFD and artificial intelligence according to claim 1, characterized in that, The key flow field regions in step S2 include the acceleration zone, shear layer, and recirculation zone marked by threshold determination and flow feature extraction algorithms; the parameter space of CFD numerical simulation for various incoming flow conditions includes inlet wind direction, inlet wind speed level, turbulence intensity level, and atmospheric boundary layer wind speed profile parameters.
7. The method for optimizing the layout of three-dimensional wind field inversion measuring points based on CFD and artificial intelligence according to claim 1, characterized in that, The online adaptive correction in step S10 specifically includes: collecting radial velocity echo data measured in the field, comparing it with the predicted values of the lightweight wind field proxy model under the same incoming flow conditions, performing transfer learning and parameter fine-tuning on the proxy model based on the comparison deviation, and re-optimizing the radar's scanning sector, elevation range, and scanning period based on the corrected proxy model while keeping the radar's physical position unchanged.
8. The method for optimizing the layout of measuring points for three-dimensional wind field inversion based on CFD and artificial intelligence according to claim 1, characterized in that, The process of establishing the observation geometry model and the visible field model in step S3 includes: setting the maximum / minimum detection range, azimuth and elevation scanning range, and range gate length of the radar; using the ray tracing algorithm to traverse the laser beams emitted from the candidate measurement point, determining whether the line of sight is blocked by terrain or buildings, and outputting the set of unobstructed effective lines of sight for the measurement point; and calculating the direction cosine vector and local observation matrix on the effective observation grid according to the scanning mode.
9. The method for optimizing the layout of measuring points for three-dimensional wind field inversion based on CFD and artificial intelligence according to claim 1, characterized in that, The fine verification in step S9 uses a more refined wind field grid and a more stringent instrument measurement error model than the CFD sample library. Simulation observation and inversion are repeatedly performed under the prevailing wind direction and extreme unfavorable conditions. When the inversion error and matrix condition number meet the preset standards, the point layout scheme is solidified.
10. The method for optimizing the layout of measuring points for three-dimensional wind field inversion based on CFD and artificial intelligence according to claim 1, characterized in that, After step S10, the process also includes: encapsulating the calculation, analysis and correction results of the entire process into an engineering implementation data package, which includes a three-dimensional visualization model of the survey area, the final coordinates of the radar distribution points, an adaptive scanning control script, and an assessment report on the expected coverage and inversion accuracy of key complex flow areas.