Three-dimensional wind field spatial-temporal feature reconstruction and efficient prediction method

Abstract

The invention discloses a three-dimensional wind field spatial-temporal feature reconstruction and efficient prediction method, and relates to the field of spatial-temporal feature reconstruction and efficient prediction. Constructing a three-dimensional terrain computational domain based on the digital elevation model of the target mountain region, performing multi-scene numerical simulation by adopting a fluid mechanics method or a mesoscale meteorological model, generating a wind field training data set, and constructing and training a wind field spatial feature mapping model; training a wind speed and wind direction short-time prediction model based on the actually measured data set; inputting the monitoring data obtained in real time into the wind speed and wind direction short-time prediction model to obtain a future wind speed and wind direction prediction value of each monitoring station; and inputting the wind speed and direction predicted values into the wind field spatial feature mapping model to obtain the mountain overall wind field distribution of the target mountain region at the future moment. By constructing an'actual measurement-simulation-modeling-prediction-reconstruction 'integrated technical framework, high-temporal-spatial-resolution short-time prediction from observation of local wind speed and wind direction to the overall three-dimensional wind field of the mountainous region is realized.

Description

Technical Field

[0001] This invention relates to the field of spatiotemporal feature reconstruction and efficient prediction technology, and in particular to a method for spatiotemporal feature reconstruction and efficient prediction of three-dimensional wind fields. Background Technology

[0002] Engineering and scientific fields such as wind energy resource assessment, wind farm site selection and operation control, and mountain weather forecasting and disaster early warning rely heavily on the accurate perception and prediction of wind field characteristics under complex mountainous terrain. However, due to the dramatic undulations and significant variations in surface roughness of mountainous terrain, airflow is prone to complex flow phenomena such as separation, acceleration, vortices, and wakes when passing over slopes, ridges, and valleys. This results in wind fields exhibiting strong non-uniformity, unsteadiness, and spatial heterogeneity, posing a significant challenge to accurate wind field modeling and short-term forecasting.

[0003] Currently, the monitoring and prediction of wind fields in complex terrain mainly rely on two technical approaches: one is on-site observation based on sparsely distributed ground meteorological stations or lidar. Although this can obtain local high-precision measured data, it is difficult to achieve large-scale, high spatial resolution wind field coverage due to limitations in equipment cost and deployment conditions, and it cannot reflect the overall flow field structure. The other is numerical simulation using computational fluid dynamics (CFD) or mesoscale meteorological models. Although this can provide full-field wind field information, it is sensitive to boundary conditions, consumes a lot of computational resources, and is difficult to update in real time, making it difficult to meet the needs of short-term (e.g., 0-6 hours) dynamic prediction.

[0004] To improve prediction accuracy, some studies in recent years have attempted to introduce artificial intelligence and machine learning methods to build wind speed prediction models using historical data. However, existing AI models mostly focus on wind speed time series predictions at single sites, neglecting the physical constraints of topography and spatial correlation, resulting in poor generalization ability and difficulty in achieving spatial reconstruction of the global wind field from local observations. In addition, most methods treat observation, simulation, and prediction in isolation, lacking a systematic fusion mechanism and failing to fully leverage the synergistic advantages of multi-source data.

[0005] Therefore, a method for reconstructing and efficiently predicting the spatiotemporal characteristics of three-dimensional wind fields is provided to solve the above problems. Summary of the Invention

[0006] To address the aforementioned challenges, this invention provides a method for reconstructing and efficiently predicting the spatiotemporal characteristics of three-dimensional wind fields. By constructing an integrated technical framework of "measurement-simulation-modeling-prediction-reconstruction," it solves the problems of incomplete wind field perception, low forecast accuracy, and slow response speed caused by sparse observations, time-consuming simulations, and lack of physical constraints in prediction models in existing technologies. This method enables high spatiotemporal resolution short-time prediction of the overall three-dimensional wind field of mountains from local wind speed and direction observations.

[0007] To achieve the above objectives, this invention provides a method for reconstructing and efficiently predicting the spatiotemporal characteristics of three-dimensional wind fields, comprising the following steps: S1: Obtain the measured dataset. The measured dataset is obtained by setting up multiple wind field monitoring stations in the target mountainous area, collecting wind speed and direction data, and performing data preprocessing. S2: Construct a three-dimensional terrain computational domain based on the digital elevation model of the target mountainous area, and use fluid dynamics methods or mesoscale meteorological models to perform multi-scenario numerical simulations to generate a wind field training dataset. S3: Based on the wind field training dataset, construct and train a wind field spatial feature mapping model; the wind field spatial feature mapping model is a mapping of the wind speed characteristics from the reference station observations to the target station. S4: Based on the measured dataset, train a short-term wind speed and direction prediction model, which is constructed based on a long short-term memory network; S5: Input the real-time acquired monitoring data into the short-term wind speed and direction prediction model to obtain the future wind speed and direction prediction values ​​for each monitoring station; input the wind speed and direction prediction values ​​into the wind field spatial feature mapping model to obtain the overall wind field distribution of the target mountain area at future times.

[0008] Preferably, data preprocessing includes removing outliers, filling in short-term missing data, and performing terrain occlusion correction.

[0009] Preferably, wind field monitoring stations are located at the top of ridgelines, slope inflection areas, canyon entrances and narrowing sections, abrupt topographic changes, and downwind areas where wake shielding effects occur.

[0010] Preferably, S2 specifically includes: S21: Construct a three-dimensional terrain calculation domain based on a high-precision digital elevation model of the target mountainous region; the calculation domain includes the target mountainous region and its upstream and downstream extension areas; S22: High-resolution numerical simulation of wind fields is performed using computational fluid dynamics methods or mesoscale meteorological models; S23: Extract the full-field three-dimensional wind speed vector distribution data and corresponding terrain auxiliary information from each simulation result to form a wind field training dataset.

[0011] Preferably, numerical simulation in S22 is performed using computational fluid dynamics methods, specifically including: The steady-state or transient incompressible airflow field is solved by combining the Reynolds-averaged Navier-Stokes equations with a k-ε or k-ω turbulence model. The inlet boundary conditions are set as logarithmic or power-law wind profiles, the wind direction is discretized by sector, the outlet is set as a free outflow boundary, the sidewalls and top are set as symmetrical or slip walls, the ground is set as a no-slip boundary and roughness is applied, and the mesh is unstructured or multi-block structured mesh. Local refinement is carried out in areas with drastic topographic changes and near the ground layer. The solver adopts a pressure-based implicit algorithm.

[0012] Preferably, S22 uses a mesoscale meteorological model for numerical simulation, specifically including: The simulation was performed using a WRF model with nested multi-layer grids. The physical parameterization schemes used included the YSU boundary layer scheme, the Monin-Obukhov surface layer scheme, and the Noah land surface model. The driving data were NCEP FNL or ERA5 analysis data.

[0013] Preferably, the wind field spatial feature mapping model includes a first model and a second model; The first model takes synchronous observation data from ground meteorological reference stations as input, while the second model takes profile observations from vertical laser wind radar at the reference location as input. The first and second models are trained, validated, and solidified independently, and a standardized inference interface is designed.

[0014] Preferably, the training process of the first model includes: Data is extracted from the wind field training dataset. The 10-minute average wind speed and wind direction sequence of multiple reference stations are used as input features, and the wind field data of the target station is used as output to construct training sample sets and test sample sets. The candidate machine learning algorithms are trained one by one using the training sample set. The trained models are evaluated based on the root mean square error and coefficient of determination on the test set. The model whose performance meets the predetermined requirements is selected as the first model after training.

[0015] Preferably, the training process of the second model includes: Data is extracted from the wind field training dataset. The vertical height layer observed by radar, the instantaneous wind speed and wind direction angle at the corresponding height are used as inputs, and the average wind speed, fluctuating wind speed and the maximum wind speed in the next 10 minutes of the target site are used as outputs to construct training sample sets and test sample sets. The candidate machine learning algorithms are trained one by one using the training sample set. The trained models are evaluated based on the root mean square error and coefficient of determination on the test set. The model whose performance meets the predetermined requirements is selected as the second model after training.

[0016] Preferably, the short-term wind speed and direction prediction model includes a wind speed model and a wind direction model: The wind speed model is based on historical wind speed time series data and uses a long short-term memory network model to predict wind speed values ​​at future times. The wind direction model first converts historical wind speed and wind direction angle into horizontal orthogonal wind speed components U and V, then uses a long short-term memory network model to independently predict the future U and V components, and finally reconstructs the predicted U and V components into the predicted wind direction angle. The hyperparameters of the Long Short-Term Memory (LSTM) network model were determined using a Bayesian optimization algorithm.

[0017] Therefore, the present invention employs the above-mentioned method for reconstructing and efficiently predicting three-dimensional wind field spatiotemporal features, which has the following beneficial effects: (1) The present invention realizes the reconstruction of the wind field under sparse observation: By establishing a wind field spatial feature mapping model between monitoring stations and the overall wind field of the mountain, the wind speed and direction distribution of the whole field can be deduced based on local measured or predicted data when the number of monitoring points is limited, which significantly reduces the dependence on dense observation network and saves equipment deployment and maintenance costs.

[0018] (2) This invention improves the accuracy of wind field prediction in complex terrain: by integrating training data generated by high-precision numerical simulation, the wind field spatial feature mapping model can fully learn the disturbance law of airflow on the terrain (such as acceleration, separation, vortex, etc.), which has stronger physical consistency and spatial accuracy compared with pure data-driven methods.

[0019] (3) This invention takes into account both the timeliness and dynamism of prediction: it introduces a deep learning model to make short-term predictions of wind conditions at monitoring points, and combines a pre-trained wind field reconstruction model to quickly generate full-field forecasts, avoiding the high computational overhead caused by repeated solving of partial differential equations in traditional numerical simulation, achieving minute-level response, and meeting the short-term (0-6 hours) dynamic prediction requirements.

[0020] (4) This invention achieves synergistic optimization of physical models and artificial intelligence: it deeply integrates physical-driven numerical simulation with data-driven AI prediction: the simulated data is used to train the wind field spatial feature mapping model to make up for the lack of measured data; AI is used to improve the time prediction capability. The two complement each other and form a closed-loop mechanism of "physics guides AI and AI accelerates physics", which improves the overall generalization ability and robustness.

[0021] (5) This invention supports the fusion of multi-source heterogeneous data: it is compatible with multiple data sources such as ground meteorological stations, lidar, remote sensing topographic data, and mesoscale meteorological output, and has good data adaptability and system scalability, and is suitable for application scenarios with different terrain complexity and observation conditions.

[0022] (6) The present invention has good engineering application prospects: it can realize the fine perception of mountain wind fields without the need for large-scale construction of monitoring equipment, and can be widely used in wind farm site selection and power prediction, mountain flight safety assurance, local extreme wind disaster early warning, ecological environment assessment and other fields, with significant economic and social benefits.

[0023] 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

[0024] Figure 1 This is a flowchart illustrating a method for reconstructing and efficiently predicting three-dimensional wind field spatiotemporal features according to the present invention. Figure 2 This is an overall technical flowchart of an embodiment of the present invention. Detailed Implementation

[0025] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0026] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0027] The terms "comprising" or "including" as used in this invention mean that the element preceding the term encompasses the element listed after the term, and do not exclude the possibility of encompassing other elements. Terms such as "inner," "outer," "upper," and "lower" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. When the absolute position of the described object changes, the relative positional relationship may also change accordingly. In this invention, unless otherwise explicitly specified and limited, the term "attached" and similar terms should be interpreted broadly. For example, it can refer to a fixed connection, a detachable connection, or an integral part; it can refer to a direct connection or an indirect connection through an intermediate medium; it can refer to the internal communication of two elements or the interaction relationship between two elements. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0028] Example A method for reconstructing and efficiently predicting the spatiotemporal features of three-dimensional wind fields, such as Figures 1-2 As shown, it includes the following steps: S1: Obtain the measured dataset. The measured dataset is obtained by setting up multiple wind field monitoring stations in the target mountainous area, collecting wind speed and direction data, and performing data preprocessing. Specifically, within the target complex mountainous area, multiple wind field monitoring stations are deployed based on topographic features and airflow response sensitivity characteristics to effectively capture key flow structures. The deployment principles include: prioritizing monitoring points at the top of ridgelines (airflow acceleration zones), slope aspect transition areas (such as the boundary between windward and leeward slopes, where separation bubbles are easily generated), canyon entrances and contractions (where channel effects are significant), abrupt topographic changes (such as cliff edges and isolated hills), and leeward areas where wake shielding effects may occur, ensuring coverage of locations where typical complex flow phenomena occur. The wind field monitoring station consists of one or more wind measurement devices, including but not limited to: automatic weather stations: installed on the ground or on a wind tower, equipped with ultrasonic anemometers or propeller-type wind direction and speed sensors, capable of continuously measuring three-dimensional wind speed (u, v, w), wind direction, and turbulence intensity (TI); lidar (LiDAR) wind measurement equipment: employing continuous wave or Doppler pulse LiDAR systems, deployed on the ground or mobile platforms, acquiring wind speed and direction distribution data with a spatial resolution of 10-50 m within a range of 30-200 m above ground height through remote scanning modes (such as PPI, RHI, vertical profile, or virtual wind tower mode). Compared with traditional measurement equipment, lidar can obtain much richer wind field monitoring information due to its advantage of being able to measure spatial wind fields; acoustic wind profiler (SODAR) or other wind field monitoring equipment.

[0029] Each monitoring station is equipped with a GPS time synchronization module and a wireless data transmission unit (such as 4G / 5G or LoRa communication) to achieve real-time data acquisition, timestamp alignment, and centralized uploading. The system records data using a unified time base (UTC) to form a standardized measured dataset with a time resolution of 1-10 minutes, including timestamps, latitude and longitude, altitude, mean and standard deviation of instantaneous wind speed, mean wind direction, estimated turbulent kinetic energy, and equipment status information.

[0030] By removing outliers (such as sudden changes in wind speed and wind direction), filling in short-term missing data (using linear interpolation or Kriging spatial interpolation), and performing terrain occlusion correction, the accuracy and consistency of the data are ensured, providing a high-quality input foundation for subsequent numerical simulation verification, model training, and wind field reconstruction.

[0031] S2: Construct a three-dimensional terrain computational domain based on the digital elevation model of the target mountainous area, and use fluid dynamics methods or mesoscale meteorological models to perform multi-scenario numerical simulations to generate a wind field training dataset. In practical applications, a three-dimensional terrain calculation domain is constructed based on a high-precision digital elevation model of the target mountain with a resolution of no less than 10 m. The calculation domain covers the actual monitoring area and its upstream and downstream extensions, with a lateral dimension of no less than 9 km and a vertical extension to the top of the boundary layer (usually 800-1500 m) to ensure sufficient airflow development.

[0032] High-resolution numerical simulations of the wind field were then performed using computational fluid dynamics methods or mesoscale meteorological models. The Reynolds-averaged Navier-Stokes (RANS) equations were then combined with... or Turbulence model, solving for steady-state or transient incompressible airflow field, where Turbulence models can be Turbulence model. Inlet boundary conditions are set as logarithmic or power-law wind profiles, covering the local dominant wind speed range (e.g., 4-16 m / s). Wind direction is discretized into 15° sectors (24 wind directions in total), forming multiple scenario combinations. The outlet is set as a free outflow boundary, the sidewalls and top are set as symmetrical or slip walls, and the ground is set as a no-slip boundary with appropriate roughness applied. The mesh uses unstructured or multi-block structured meshes, with local refinement in areas of drastic topographic changes and near the ground layer. The height of the first mesh layer is set to meet the requirements of the target turbulence model: for Turbulence model, y + ≈1; for other cases, y + The number of elements should also be kept below 30, and the total number of grid cells should be no less than 14 million. The solver should use a pressure-based implicit algorithm (such as SIMPLE or PISO), and the convergence criterion should be a residual of less than 1 × 10⁻⁶. -4 This ensures a stable flow field.

[0033] Mesoscale meteorological model simulations were performed using the open-source simulation software WRF, specifically version CoupledWRFv4.0 or later, with nested multi-layer grids (e.g., 3 nested layers with horizontal resolutions of 9 km, 3 km, and 1 km). The innermost layer could be further downscaled using CFD. The physical parameterization schemes employed included the YSU boundary layer scheme, the Monin-Obukhov surface layer scheme, and the Noah land surface model. The driving data used were NCEP FNL or ERA5 reanalysis data with a temporal resolution of 6 hours to simulate wind field responses under different atmospheric stability conditions (neutral, stable, and unstable), thereby enhancing the generalization ability of the training data.

[0034] Simulate scenarios with multiple wind directions, wind speeds, and stability levels, such as wind directions of 0°, 15°, 30°, ... (24 directions in total); Inlet wind speed: 6 m / s, 10 m / s, 14 m / s (multiple intensities for each wind direction); Atmospheric stability: Neutral (Δθ / Δz ≈ 0), stable (temperature inversion), unstable (hyperadiabatic). Each scenario is run independently until the flow field converges or reaches statistical steady state.

[0035] Two key data points are extracted from each simulation result: instantaneous, average, fluctuating, and 10-minute maximum values ​​of the full-field three-dimensional wind speed vector distribution (u(x, y, z), v(x, y, z), w(x, y, z)) at each actual monitoring station location and under the corresponding scenario, serving as wind speed characteristic parameters. The spatial resolution is 10-50 m, typically stored as two-dimensional slices (e.g., a 10 m horizontal plane) or three-dimensional volumetric data. Terrain auxiliary information T, including elevation maps, slope maps, and aspect maps, is also recorded as static input for subsequent model training. A dataset is constructed based on the above.

[0036] S3: Based on the wind field training dataset, construct and train a wind field spatial feature mapping model; the wind field spatial feature mapping model is a mapping of the wind speed characteristics from the reference station observations to the target station. Based on high-resolution wind field data generated by numerical simulation, a spatial feature mapping model for the wind field is constructed, transforming "observable data from a reference station" into "wind speed characteristics at the target station," enabling data-driven inversion of the wind field state at the target station. The model architecture consists of two types: one uses synchronous observation data from a surface meteorological reference station as input, and the other uses profile observations from a vertical laser wind radar at the reference location as input. Both types of models are trained, validated, and solidified independently. The inference interface is standardized for easy system integration and engineering maintenance.

[0037] Taking the wind field mapping from multiple reference stations to the target station as an example, the input features are the 10-minute average wind speed and wind direction sequence simultaneously simulated at multiple reference stations, and the output is the wind speed feature parameters of the target point. The training process first divides the training and test sets into an 80% / 20% ratio, then fits and compares each model to a pre-defined candidate learner. Candidate models include linear regression, Ridge, Lasso, second- and third-order polynomial regression, RBF kernel support vector regression (RBF-SVR), Kernel Ridge Regression, k-Nearest Neighbors (kNN), Random Forest, Gradient Boosting Tree (GBRT), Gaussian Process Regression (GPR), Multilayer Perceptron (MLP), and XGBoost. Among these, tree models and Gaussian Process Regression are trained at their original scale, Ridge and Lasso models use five-fold cross-validation to select regularization strength, and the remaining models use empirical hyperparameters or lightweight grids. The model evaluation uses root mean square error (RMSE) and coefficient of determination (R²) as core metrics. The model with the smallest RMSE on the test set is selected as the optimal solution, and R² is recorded to measure stability. The optimal model and its necessary preprocessing steps are serialized and saved in a unified manner, and the model name and key performance indicators are recorded to ensure traceability and reproducibility.

[0038] Building upon this foundation, a mapping model is constructed based on vertical laser wind measurement radar reference observations to the wind field at the target site. Input features include the vertical height layers observed by the radar (e.g., 50m, 100m, ..., 1000m), the instantaneous wind speed and wind direction angle at the corresponding heights, and the outputs are the average wind speed, fluctuating wind speed, and the maximum wind speed extreme value for the target site over the next 10 minutes. The model training strategy, candidate learner set, hyperparameter tuning method, and evaluation metrics are all consistent with the multi-reference site model, ensuring methodological uniformity. Specifically, for radar vertical profile information, height encoding or wind shear gradient features can be introduced to enhance spatial extrapolation capabilities. Finally, the selected optimal model, along with input / output specifications, preprocessing logic, and performance reports, is solidified to form a standardized wind field inversion module, supporting rapid deployment and migration applications at different target sites.

[0039] S4: Based on the measured dataset, train a short-term wind speed and direction prediction model, which is constructed based on a long short-term memory network; Using historical data from meteorological stations and vertical lidar wind measurement radar, short-term forecasting models for wind speed and direction parameters were established, covering average wind speed, fluctuating wind speed, and 10-minute extreme wind speed values. Wind speed was directly modeled as a scalar; wind direction was modeled using a vector decomposition strategy, converting wind speed and wind direction angle into horizontally orthogonal wind speed components U and V, respectively, and constructing independent forecasting models for each. After forecasting, the wind direction angle was reconstructed from the wind speed components. Both models were based on a Long Short-Term Memory (LSTM) network architecture, and the hyperparameters were determined through Bayesian optimization to improve the robustness and generalization ability of the models in short-term forecasts under typical meteorological conditions.

[0040] The meteorological station section uses continuous observation time series from a single station as the modeling unit, employing a sliding window to construct input-output sample pairs: the input is a historical sequence of length L, and the output is a multi-step prediction sequence of length H. Three independent sequence prediction models are constructed for wind speed, the U component, and the V component. The network structure uses a 1-3 layer stacked LSTM, with the number of hidden units in each layer adjustable from 64 to 256. A fully connected layer at the end maps the hidden states to an H-dimensional output space. The dataset is strictly divided into training, validation, and test sets according to time order to prevent future information leakage. The loss function is mean squared error for both wind speed and component tasks. Model performance is evaluated using root mean square error (RMSE) and coefficient of determination (R²). The circumferential mean absolute error (Circular MAE) is calculated after wind direction angle reconstruction to quantify the stability and consistency of the angle prediction.

[0041] Building upon this foundation, modeling of laser wind radar data was conducted. Continuous observation sequences at different vertical heights were used as modeling units to construct LSTM prediction models for wind speed and U and V components, respectively. Two modeling strategies were employed: one was independent modeling at different heights, suitable for scenarios where the dynamic processes at each level differed significantly; the other was multi-height joint modeling, which involved concatenating multiple sequences at the input and introducing learnable height encoding to explicitly model the spatiotemporal evolution characteristics of vertical wind shear. The network structure design, data partitioning strategy, loss function, and evaluation index system were all consistent with the meteorological station models to ensure the comparability of prediction performance across platforms.

[0042] The hyperparameters of all models were uniformly determined through Bayesian optimization. The optimization search space covered key hyperparameters: historical window length, prediction step size, number of LSTM layers, hidden unit size, dropout rate, learning rate, weight decay coefficient, and loss weights for U and V components. The objective function was based on the comprehensive performance indicators on the validation set, prioritizing minimizing the RMSE of wind speed and U and V components, while controlling the circumferential error of the reconstructed wind direction angle within a preset threshold, forming a constrained multi-objective optimization problem. After optimization, the optimal hyperparameter configuration was fixed, and the random seed, search space boundary, iteration trajectory, and final parameters were fully recorded to ensure the traceability of the experimental process and the reproducibility of the results.

[0043] S5: Input the real-time acquired monitoring data into the short-term wind speed and direction prediction model to obtain the future wind speed and direction prediction values ​​for each monitoring station; input the wind speed and direction prediction values ​​into the wind field spatial feature mapping model to obtain the overall wind field distribution of the target mountain area at future times.

[0044] Specifically, during the real-time operation phase, the measured data of each monitoring point at the current moment is input into the short-term wind speed and direction prediction model to obtain the predicted wind speed and direction values ​​of each point in the future. Then, this predicted value is used as input and substituted into the pre-trained wind field spatial feature mapping model to directly output the overall wind field distribution of the mountain in the future, realizing short-term wind field forecasts with a response time of minutes.

[0045] Therefore, the present invention adopts the above-mentioned method for reconstructing and efficiently predicting the spatiotemporal characteristics of three-dimensional wind fields. By combining the physical regularity of high-fidelity numerical simulation with the data-driven advantages of deep learning models, it not only ensures the spatial and physical rationality of wind field reconstruction, but also achieves high efficiency and automation of the prediction process. It overcomes the contradiction between accuracy, speed and cost in traditional methods and has good scalability and engineering applicability.

[0046] 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 reconstructing and efficiently predicting the spatiotemporal characteristics of a three-dimensional wind field, characterized in that, Includes the following steps: S1: Obtain the measured dataset; the measured dataset is obtained by setting up multiple wind field monitoring stations in the target mountainous area, collecting wind speed and direction data, and performing data preprocessing. S2: Construct a three-dimensional terrain computational domain based on the digital elevation model of the target mountainous area, and use fluid dynamics methods or mesoscale meteorological models to perform multi-scenario numerical simulations to generate a wind field training dataset. S3: Based on the wind field training dataset, construct and train a wind field spatial feature mapping model; The wind field spatial feature mapping model is a mapping of wind speed characteristics from the reference station observations to the target station. S4: Based on the measured dataset, train a short-term wind speed and direction prediction model, which is constructed based on a long short-term memory network; S5: Input the real-time acquired monitoring data into the short-term wind speed and direction prediction model to obtain the future wind speed and direction prediction values ​​for each monitoring station; input the wind speed and direction prediction values ​​into the wind field spatial feature mapping model to obtain the overall wind field distribution of the target mountain area at future times.

2. The method for reconstructing and efficiently predicting three-dimensional wind field spatiotemporal features according to claim 1, characterized in that, Data preprocessing includes removing outliers, filling in short-term missing data, and performing terrain occlusion correction.

3. The method for reconstructing and efficiently predicting the spatiotemporal characteristics of a three-dimensional wind field according to claim 1, characterized in that, Wind field monitoring stations are set up at the top of ridgelines, slope inflection areas, canyon entrances and narrowing sections, abrupt topographic changes, and downwind areas where wake shielding effects occur.

4. The method for reconstructing and efficiently predicting the spatiotemporal characteristics of a three-dimensional wind field according to claim 1, characterized in that, S2 specifically includes: S21: Construct a three-dimensional terrain calculation domain based on a high-precision digital elevation model of the target mountainous region; the calculation domain includes the target mountainous region and its upstream and downstream extension areas; S22: High-resolution numerical simulation of wind fields is performed using computational fluid dynamics methods or mesoscale meteorological models; S23: Extract the full-field three-dimensional wind speed vector distribution data and corresponding terrain auxiliary information from each simulation result to form a wind field training dataset.

5. The method for reconstructing and efficiently predicting the spatiotemporal characteristics of a three-dimensional wind field according to claim 4, characterized in that, S22 employs computational fluid dynamics methods for numerical simulation, specifically including: The steady-state or transient incompressible airflow field is solved by combining the Reynolds-averaged Navier-Stokes equations with a k-ε or k-ω turbulence model. The inlet boundary conditions are set as logarithmic or power-law wind profiles, the wind direction is discretized by sector, the outlet is set as a free outflow boundary, the sidewalls and top are set as symmetrical or slip walls, the ground is set as a no-slip boundary and roughness is applied, and the mesh is unstructured or multi-block structured mesh. Local refinement is carried out in areas with drastic topographic changes and near the ground layer. The solver adopts a pressure-based implicit algorithm.

6. The method for reconstructing and efficiently predicting three-dimensional wind field spatiotemporal features according to claim 4, characterized in that: S22 employs a mesoscale meteorological model for numerical simulation, specifically including: The simulation was performed using a WRF model with nested multi-layer grids. The physical parameterization schemes used included the YSU boundary layer scheme, the Monin-Obukhov surface layer scheme, and the Noah land surface model. The driving data were NCEP FNL or ERA5 analysis data.

7. The method for reconstructing and efficiently predicting three-dimensional wind field spatiotemporal features according to claim 1, characterized in that: The wind field spatial feature mapping model includes a first model and a second model; The first model takes synchronous observation data from ground meteorological reference stations as input, while the second model takes profile observations from vertical laser wind radar at the reference location as input. The first and second models are trained, validated, and solidified independently, and a standardized inference interface is designed.

8. The method for reconstructing and efficiently predicting the spatiotemporal characteristics of a three-dimensional wind field according to claim 7, characterized in that, The training process for the first model includes: Data is extracted from the wind field training dataset. The 10-minute average wind speed and wind direction sequence of multiple reference stations are used as input features, and the wind field data of the target station is used as output to construct training sample sets and test sample sets. The candidate machine learning algorithms are trained one by one using the training sample set. The trained models are evaluated based on the root mean square error and coefficient of determination on the test set. The model whose performance meets the predetermined requirements is selected as the first model after training.

9. The method for reconstructing and efficiently predicting the spatiotemporal characteristics of a three-dimensional wind field according to claim 7, characterized in that, The training process for the second model includes: Data is extracted from the wind field training dataset. The vertical height layer observed by radar, the instantaneous wind speed and wind direction angle at the corresponding height are used as inputs, and the average wind speed, fluctuating wind speed and maximum wind speed in the next 10 minutes at the target site are used as outputs to construct training sample sets and test sample sets. The candidate machine learning algorithms are trained one by one using the training sample set. The trained models are evaluated based on the root mean square error and coefficient of determination on the test set. The model whose performance meets the predetermined requirements is selected as the second model after training.

10. The method for reconstructing and efficiently predicting the spatiotemporal characteristics of a three-dimensional wind field according to claim 1, characterized in that, Short-term wind speed and direction prediction models include wind speed models and wind direction models: The wind speed model is based on historical wind speed time series data and uses a long short-term memory network model to predict wind speed values ​​at future times. The wind direction model first converts historical wind speed and wind direction angle into horizontal orthogonal wind speed components U and V, then uses a long short-term memory network model to independently predict the future U and V components, and finally reconstructs the predicted U and V components into the predicted wind direction angle. The hyperparameters of the Long Short-Term Memory (LSTM) network model were determined using a Bayesian optimization algorithm.

Images