Nowcasting method and device for low-altitude three-dimensional wind field

Abstract

The invention discloses a low-altitude three-dimensional wind field nowcasting method and device, and the method comprises the steps: firstly collecting and preprocessing wind field data through an API, constructing a spatial-temporal feature project, and dividing a data set according to a time sequence; thirdly, a LightGBM framework is adopted, and a static optimal forecasting model is trained by means of Bayesian optimization; the method is characterized in that the static optimal forecasting model and an ensemble Kalman filtering framework are deeply fused, a dynamic assimilation forecasting system is constructed, forecasting output is continuously corrected and optimized, and error accumulation is effectively restrained. And finally, after performance verification, the system is operated in a business mode, and a final wind field forecasting result containing the NaN identifier is output.

Description

TECHNICAL FIELD

[0001] The present application relates to a low-altitude three-dimensional wind field nowcasting method and device, in particular to a low-altitude three-dimensional wind field nowcasting method and device based on deep learning and ensemble Kalman filtering, and relates to the fields of data processing and machine learning. BACKGROUND

[0002] Fine low-altitude wind field prediction has important application value in low-altitude economy, unmanned aerial vehicle route planning, urban air traffic, wind power generation, and aviation safety, etc. Although new detection technologies (such as laser radar, phased array radar) can obtain high spatiotemporal resolution low-altitude wind field observation data, how to effectively use these data for accurate and real-time prediction still faces the following core challenges.

[0003] Firstly, traditional prediction techniques cannot meet the fine prediction requirements. Numerical weather prediction (NWP) models are the main means of traditional prediction, but they are limited by physical parameterization schemes and computational resource consumption, and are mainly suitable for large-scale, medium and long-term weather process prediction. For local terrain, building disturbance and turbulence below three kilometers from the ground, the NWP model is insufficient in describing ability and cannot meet the needs of business fine prediction.

[0004] Secondly, the strong nonlinear characteristics of ultra-high precision wind field data pose a severe challenge to modeling methods. High spatiotemporal resolution data puts high requirements on the computational efficiency and scalability of the model. The evolution of low-altitude wind field is a complex nonlinear physical process, and traditional numerical models are difficult to completely capture its spatiotemporal dynamic correlation. Although machine learning (ML) methods provide a new solution, their inherent "black box" characteristics often lead to insufficient physical interpretation, affecting the reliability and credibility of prediction decisions.

[0005] Furthermore, the prediction model based on machine learning has the inherent defect of "open loop" prediction (error accumulation problem). The parameters of the existing data-driven model are fixed after training, and a static "open loop" mode is used for continuous time extrapolation prediction. In this mode, small deviations in the initial field or systematic errors in the model itself will be amplified step by step as the prediction time elapses. This leads to rapid decay of prediction accuracy over time, making it difficult to meet the requirements of long-term stability for business applications.

[0006] In summary, there is an urgent need in the art for a new technical solution that can effectively utilize high spatiotemporal resolution observation data, efficiently and accurately capture the nonlinear evolution of low-altitude wind field, and have a dynamic correction capability to suppress the accumulation of prediction errors, thereby realizing truly high-precision, business-oriented fine prediction of low-altitude wind field. The present application is proposed to solve the problems in the prior art. SUMMARY

[0007] To solve the above problems, the application provides a low-altitude three-dimensional wind field nowcasting method and device, which realizes dynamic prediction of three-dimensional high-precision wind field through deep learning algorithm, ensemble Kalman filter and other methods.

[0008] The application provides a low-altitude three-dimensional wind field nowcasting method, characterized by comprising the following steps: a data collection and preprocessing step, wherein observation data of a low-altitude three-dimensional wind field is collected in real time through an API interface, stored as a NetCDF (nc) file, and cleaned, standardized and interpolated to generate a spatially continuous and complete three-dimensional wind field data matrix; a feature engineering and target construction step, wherein a spatiotemporal feature engineering is constructed, key factors reflecting dynamic evolution of the wind field are extracted, and historical data is divided into a training set, a validation set and a test set according to a time sequence principle; a static model construction and optimization step, wherein a LightGBM framework is used for model training, and a Bayesian optimization algorithm is used for fine tuning of core hyperparameters, and an optimal prediction model is obtained based on an optimal parameter combination on an extended data set composed of the training set and the validation set; a dynamic assimilation system integration step, wherein the static optimal prediction model is used as a core prediction engine, and is deeply integrated with an ensemble Kalman filter framework to construct a dynamic assimilation prediction system; a system performance verification step, wherein the test set in the feature engineering and target construction step is used to evaluate the performance of the entire assimilation prediction system, and the mean absolute error (MAE) and the root mean square error (RMSE) are calculated, and the result is lower than the baseline in the static model construction and optimization step by 2; and a business operation step, wherein the dynamic assimilation prediction system is operated in business, and NaN values are used to fill and identify missing grid positions in the observation data in the final prediction product to generate a final wind field prediction result.

[0009] In some embodiments, the interpolation processing in the data collection and preprocessing step uses an inverse distance weighted interpolation method to estimate the value of a missing position marked as NaN in the three-dimensional wind field data matrix.

[0010] In some embodiments, the features extracted by the spatiotemporal feature engineering in the static model construction and optimization step include a time feature and a space feature, the time feature includes periodic time features, lag variables, sliding window statistics and wind speed change trend features, and the space feature includes static space features, neighborhood features, spatial gradient features and divergence or vertical vorticity physical features.

[0011] In some embodiments, the static model construction and optimization steps include: a step of setting initial parameters, which involves setting a set of empirically sound and computationally efficient LightGBM model parameters based on the size and feature dimensions of the dataset; a step of initial model training, which involves training the model on the training set using the aforementioned initial parameters; and a step of benchmark performance evaluation, which involves evaluating the trained preliminary model on the validation set and calculating its mean absolute error (MAE) and root mean square error (RMSE) as performance metrics. These metrics will serve as the baseline that subsequent hyperparameter tuning needs to surpass.

[0012] In some embodiments, the dynamic assimilation system integration step includes: an initialization step, which obtains the observed wind field y(0) at the initial time and adds N random perturbations to generate N set members; and a cyclic assimilation step, which, based on the static optimal forecast model, continuously corrects and optimizes the output of the deep learning wind field forecast model by periodically incorporating the latest observation data and using an ensemble Kalman filter algorithm; wherein, using the static optimal forecast model, all set members from the previous time (t-1) are pushed forward to the current time (t), and at the same time, in order to compensate for the shortcomings of a single model and prevent set collapse, a random perturbation is added to each member.

[0013] In some embodiments, the Kalman gain K of the dynamic assimilation system integration step is calculated using a set approximation, which involves the perturbation matrix of the forecast set and the matrix that maps the perturbation to the observation space.

[0014] This invention also provides a near-term forecasting device for low-altitude three-dimensional wind fields, characterized by comprising: a data collection and preprocessing unit, used to collect low-altitude three-dimensional wind field observation data in real time through an API interface, store the data as a NetCDF(nc) file, and perform preprocessing such as cleaning, standardization, and interpolation on the data; a feature engineering and target construction unit, used to construct spatiotemporal feature engineering, extract key factors that can reflect the dynamic evolution of the wind field, and generate training sets, validation sets, and test sets according to time series principles; and a static model construction and optimization unit, used to train the model using the LightGBM framework and fine-tune the core hyperparameters using a Bayesian optimization algorithm. The system employs several mechanisms: a static optimal forecast model is obtained by retraining on an extended dataset (training set + validation set) based on the optimal parameter combination; a dynamic assimilation system integration unit is used to deeply integrate the static optimal forecast model with an ensemble Kalman filter framework to construct a complete dynamic assimilation forecast system; a system performance verification unit is used to evaluate and verify the performance of the entire system on the test set to ensure that it meets the preset accuracy requirements; and an operational unit is used to operationally run the dynamic assimilation forecast system, filling and marking missing grid locations in the final forecast product with NaN values ​​to generate the final wind field forecast result.

[0015] In some embodiments, the feature engineering and target construction unit are specifically used to extract key factors including periodic time features, lagged variables, sliding window statistics, wind speed change trend features, static spatial features, neighborhood features, spatial gradient features, and physical features.

[0016] In some embodiments, the feature engineering and target construction unit are specifically used to extract key factors including periodic time features, lagged variables, sliding window statistics, wind speed change trend features, static spatial features, neighborhood features, spatial gradient features, and physical features.

[0017] In some embodiments, the static model building and optimization unit includes a hyperparameter search module, which uses a Bayesian optimization method to automatically search for the number of iterations, learning rate, number of leaf nodes, maximum tree depth, and regularization parameters of LightGBM, so as to obtain the parameter combination that minimizes the root mean square error (RMSE) on the validation set.

[0018] In some embodiments, the Kalman gain calculated by the dynamic assimilation system integration unit is based on the covariance of the perturbation matrix of the forecast set and the observation perturbation matrix, and a gain matrix approximating the set is constructed accordingly.

[0019] The beneficial effects of this invention are as follows.

[0020] 1. This invention directly utilizes high-density observed wind field data, combined with advanced machine learning methods and data assimilation techniques, to construct for the first time an operational system capable of achieving refined wind field forecasting at ultra-small scales, effectively solving the current problem of insufficient refined forecasting capabilities for low-altitude three-dimensional wind fields.

[0021] 2. This invention employs the LightGBM model, which can process massive amounts of high-resolution wind field data with extremely high computational efficiency, meeting the timeliness requirements of operational needs. Simultaneously, the model can automatically learn the nonlinear physical relationships in wind field evolution, providing a certain degree of physical interpretability for forecast results; and through regularization and anti-overfitting settings, it further enhances the model's generalization ability and stability.

[0022] 3. This invention integrates the optimal static prediction model into the Ensemble Kalman Filter (EnKF) framework, enabling dynamic correction of forecast results based on the latest observational data. This effectively suppresses model errors accumulated over time, maintaining high forecast accuracy throughout the 0-2 hour nowcast period. In summary, this invention systematically integrates efficient machine learning models and advanced data assimilation methods, fundamentally addressing the three core challenges of massive data processing, strong nonlinear dynamic feature learning, and error accumulation in low-altitude high-precision three-dimensional wind field forecasting. It provides an end-to-end, dynamically correctable, operationally refined wind field forecasting solution. Attached Figure Description

[0023] Embodiments of the present invention will be described with reference to the accompanying drawings, which are briefly described below.

[0024] Figure 1 A flowchart illustrating the nowcasting method for low-altitude three-dimensional wind fields of the present invention is shown below.

[0025] Figure 2 This diagram illustrates the process of training the preliminary prediction model based on LightGBM according to the present invention.

[0026] Figure 3 A flowchart illustrating the integration steps of the dynamic assimilation system of the present invention is provided.

[0027] Figure 4 A schematic diagram illustrating the near-term forecasting device for low-altitude three-dimensional wind fields of the present invention.

[0028] In the accompanying drawings, the same reference numerals generally denote identical, functionally similar, and / or structurally similar components. The first appearance of a component in the accompanying drawing is indicated by the leftmost digit of the corresponding reference numeral.

[0029] In the figure: 1. Low-altitude three-dimensional wind field nowcasting device; 10. Data collection and preprocessing unit; 20. Feature engineering and target construction unit; 30. Static model construction and optimization unit; 40. Dynamic assimilation system integration unit; 50. System performance verification unit; 60. Operational unit. Detailed Implementation

[0030] Example 1

[0031] like Figure 1 As shown, according to the present invention, a method for near-term forecasting of low-altitude three-dimensional wind fields is characterized by comprising the following steps: data collection and preprocessing step S1, feature engineering and target construction step S2, static model construction and optimization step S3, dynamic assimilation system integration step S4, system performance verification step S5, and operational operation step S6.

[0032] Data collection and preprocessing step S1

[0033] The observation data of the low-altitude three-dimensional wind field is collected in real time through the API interface, stored as a NetCDF(nc) file, and the observation data is cleaned, standardized and interpolated to generate a spatially continuous and complete three-dimensional wind field data matrix.

[0034] In terms of data collection, low-altitude three-dimensional wind fields for 1 to 2 years are obtained through API interfaces, such as horizontal wind speed hs, horizontal wind direction hd, and vertical wind speed w, with a time resolution of 2 minutes, a horizontal spatial resolution of 20 meters, and a vertical spatial resolution of 10 meters / 20 meters (which varies with altitude), and are saved as NetCDF(nc) files.

[0035] For data preprocessing, all historical 3D wind field data stored in NetCDF(nc) format were read in batches. All original missing or invalid values ​​in the data were uniformly filled and marked using standard NaN. Analysis confirmed that the missing data was mainly due to physical occlusion caused by fixed obstacles such as terrain and buildings. Therefore, the locations of missing data exhibited a fixed spatial distribution at each height level, and the missing area systematically decreased with increasing height. Based on this fixed pattern, the data was cleaned by checking the number and spatial distribution of NaN values ​​for each data slice at each time step. If the distribution pattern significantly differed from the expected physical occlusion pattern, the data slice was determined to be an anomaly and removed from the dataset.

[0036] The cleaned wind field vectors are transformed into horizontal wind components u and v using trigonometric functions. Inverse distance weighted interpolation is then used to interpolate the NaN positions in the u, v, and w wind components, generating a spatially continuous and complete three-dimensional wind field data matrix of u, v, and w, which serves as input for subsequent feature engineering and model training.

[0037] The horizontal wind speed hs and wind direction hd are converted into horizontal wind components u and v using trigonometric functions. Since due north (the direction from which the wind originates) is 0˚, standard trigonometric decomposition is used. , .

[0038] The interpolation process calculates weights based on the distances to nearby valid data points around the target point; data points that are closer to each other have higher weights, thereby estimating the values ​​at the missing measurement locations. The specific steps are as follows.

[0039] First, all missing measurement points (nan_indexes) are found. To improve interpolation efficiency and ensure the physical locality of the results, the maximum neighborhood range is determined for each missing measurement point (x, y, z). This neighborhood range is defined as a three-dimensional cube space with a side length of 2R+1 centered on the missing measurement point, where R is the preset maximum search radius (in units of grid points, generally R=5). The adaptive search radius starts from the smallest radius and gradually expands the search range until a valid point is found. First, the locations of all missing data points are identified and indexed, generating a list of missing point coordinates. Second, an initial search radius and a maximum search radius are set. The initial search radius generally starts from 1, and the maximum search radius is determined based on the scale of the entire wind field to prevent infinite expansion. A radius increment is also set, usually 1. Third, all missing measurement points are traversed, and an adaptive search is performed for each missing measurement point. The steps are: a. Set the current search radius as the initial search radius. b. Execute an iterative neighborhood search loop: Using the current missing point as the center, construct a two-dimensional square neighborhood with a side length of 2R+1. Within this neighborhood, traverse all non-center neighbor points. If any neighbor point has a non-empty value, use the current radius as the search radius for that missing point. At this point, terminate the iterative neighborhood search and exit the loop. c. Generate a list of search radii for each missing point for future use.

[0040] Specifically, the index range of this neighborhood in three dimensions The calculation method is as follows, taking into account the boundary conditions of the wind field grid:

[0041] x-axis range : , ;

[0042] y-axis range : , ;

[0043] z-axis range : , ;

[0044] in , , This represents the total dimension of the wind field grid.

[0045] Second, a traversal is performed within this neighborhood to identify all valid data points that are not missing tests and their corresponding values, forming a set of nearest neighbors for weighted calculation. The local search strategy ensures that the interpolation results are primarily influenced by the most relevant (i.e., spatially closest) data points, and significantly reduces the algorithm's time complexity.

[0046] Third, calculate the distance between the missing measurement points and the valid measurement points. The calculation formula is as follows:

[0047] ;

[0048] in, It is a missing measurement point. , is a valid point within the largest neighborhood of the missing measurement point.

[0049] Fourth, calculate the interpolation weights based on the distance. Calculate the average weight value And fill in the corresponding missing test points.

[0050] Feature engineering and target construction steps S2

[0051] A spatiotemporal feature engineering model was constructed to extract key factors that can reflect the dynamic evolution of the wind field, and historical data was divided into independent training, validation and test sets according to the time series principle.

[0052] In terms of feature construction, spatiotemporal features of the three-dimensional wind field are extracted. Temporal features are extracted to capture the temporal evolution of the wind field, specifically including: Periodic temporal features: Sine and cosine encoding is performed on times of day and dates within a year to characterize the diurnal cycle and seasonal variations of the wind field. Lag variables u, v, w: The u, v, and w wind components of the target point over multiple consecutive time steps (t-Δt, t-2Δt, ...) are introduced as the core dynamic inputs of the model. The number of consecutive time steps is generally three, but needs to be adjusted according to specific circumstances. Sliding window statistics: The sliding average, standard deviation, and maximum / minimum values ​​of a past time window are calculated to smooth short-term fluctuations and characterize the stability of recent trends. Wind speed change trend features: The wind speed difference and acceleration of u, v, and w are calculated to characterize the wind speed change trend.

[0053] Extracting spatial features is to describe the spatial distribution structure of the wind field and its interaction with the environment. Specifically, this includes: static spatial features: the coordinates themselves, the distance to the center, boundary features (whether it is on the boundary), and underlying surface features. Neighborhood features: incorporating wind speed values ​​from other grid points within the three-dimensional neighborhood of the target point, as well as local spatial statistics within the neighborhood (such as mean and standard deviation). Spatial gradient features: calculating the partial derivatives of u, v, and w in three directions through central difference, i.e. Physical characteristics: Divergence and vertical vorticity are calculated using u, v, and w, with the following formulas:

[0054] ;

[0055] .

[0056] Regarding dataset generation, all historical data are traversed, and a sample (row) is created for each point in the three-dimensional space and each time point. That is, each row consists of extracted features. The target of this sample is the u / v / w value of the point at time t+dt. All samples generated at all historical times are divided into mutually independent training sets, validation sets and test sets according to the time sequence division principle.

[0057] The division into independent training, validation, and test sets according to the time sequence principle requires that the timestamp of the training set must be earlier than the timestamp of the validation / test set. Furthermore, since the training set and validation set will expand into a new training set during subsequent training and optimization, the timestamp of the validation set must be earlier than the timestamp of the test set. In other words, the timestamp of the training set must be earlier than the timestamp of the validation set, which in turn must be earlier than the timestamp of the test set. To ensure the reliability of the model's long-term stability assessment, the test set needs to have at least one month's worth of data. For stable tuning, the validation set is slightly larger. Based on this principle, the training, validation, and test sets are approximately divided into 70%, 20%, and 10% respectively.

[0058] Static model construction and optimization steps S3

[0059] The LightGBM framework was used for model training, and the core hyperparameters were finely tuned using the Bayesian optimization algorithm. Based on the optimal parameter combination, the model was retrained on an extended dataset consisting of the training and validation sets to obtain a static optimal prediction model.

[0060] Furthermore, based on the LightGBM model, model training is performed separately for u, v, and w, divided into preliminary forecast model training, hyperparameter tuning, and final model training. The preliminary forecast model training involves using the training set to initially set the model parameters and establish a baseline forecast model based on LightGBM.

[0061] like Figure 2 As shown, the static model construction and optimization step S3 further includes the preliminary prediction model training step S31, the hyperparameter tuning step S32, and the final model training step S33. Step S31 includes the following steps:

[0062] In step S311, initial parameters are set based on the dataset size and feature dimensions. A set of empirically validated, computationally efficient LightGBM model parameters is established. The number of iterations, `n_estimators`, is set to a moderate fixed value (e.g., 1000), and early stopping is enabled. Early stopping is a regularization technique that terminates training before overfitting by monitoring validation set performance, aiming to prevent overfitting, save computational resources, and improve model generalization ability. The learning rate, `learning_rate`, is set to a relatively high value (e.g., 0.1) to accelerate the convergence of initial training. The number of leaf nodes, `num_leaves`, and the maximum tree depth, `max_depth`, are set to conservative values ​​(e.g., `num_leaves=31`, `max_depth=7`) to control the initial complexity of the model.

[0063] In the initial model training step S312, the model is trained on the training set based on LightGBM using the initial parameters mentioned above.

[0064] In the baseline performance evaluation step S313, the trained preliminary model is evaluated on the validation set, and its mean absolute error (MAE) and root mean square error (RMSE) are calculated as performance metrics. These metrics will serve as the baseline 1 that subsequent hyperparameter tuning needs to surpass.

[0065] The Mean Absolute Error (MAE) is the average of the deviations of all observations from the mean, and it accurately reflects the actual error. The calculation formula is as follows:

[0066] .

[0067] The root mean square error (RMSE) is the arithmetic square root of the variance, reflecting the degree of dispersion of a dataset. The calculation formula is as follows:

[0068] .

[0069] In the formula, P is the predicted wind speed, R is the actual wind speed, and n is the number of all points in all test sets.

[0070] In step S32, hyperparameter tuning involves using Bayesian optimization to find the optimal combination of parameters so that the root mean square error (RMSE) of the model on the validation set is minimized as much as possible.

[0071] The parameter combination includes the number of iterations n_estimators, learning rate learning_rate, number of leaf nodes num_leaves, maximum tree depth max_depth, L1 regularization coefficient reg_alpha, L2 regularization coefficient reg_lambda, minimum number of leaf nodes min_child_samples, and minimum split gain min_split_gain, etc. Step S32 further includes the following steps:

[0072] The core performance parameter optimization step S321 includes the number of iterations n_estimators, the learning rate learning_rate, the number of leaf nodes num_leaves, and the maximum tree depth max_depth. Specifically...

[0073] (1) n_estimators and learning_rate: Under the premise of stable convergence, a larger learning_rate tends to iterate faster (smaller n_estimators), while a smaller learning_rate iterates slower (larger n_estimators). However, exceeding the convergence interval can lead to excessively large learning_rates, causing oscillations or even divergence. Larger n_estimators may even prevent convergence. Therefore, it is necessary to optimize n_estimators and learning_rate in a coordinated manner. Specifically, in Bayesian optimization, a large upper limit for n_estimators (e.g., 5000) is fixed, while searching for the optimal learning_rate (e.g., within the interval [0.01, 0.2]), and an early stopping mechanism is used to automatically determine the actual required n_estimators.

[0074] (2) num_leaves and max_depth: These two parameters together determine the complexity of the tree and the model's fitting ability. The search is performed within the intervals [20, 150] and [5, 20].

[0075] Step S322, regularization and anti-overfitting parameter optimization: After determining the core performance parameters, it is necessary to improve the model's generalization ability and prevent it from overfitting on the training set. Therefore, the regularization parameters (L1 regularization coefficient reg_alpha, L2 regularization coefficient reg_lambda) and anti-overfitting parameters (minimum number of leaf node samples min_child_samples, minimum split gain min_split_gain) need to be optimized. Specifically...

[0076] (1) reg_alpha and reg_lambda: Search in the logarithmic space (e.g., [1e-8, 1.0]) to fine-tune the regularization strength of the model.

[0077] (2) min_child_samples and min_split_gain: further control the growth of the tree.

[0078] The step of determining the optimal parameter combination (S323) involves selecting the hyperparameter combination that minimizes the RMSE on the validation set after completing the above optimization, as the parameters of the final model.

[0079] The final model training step S33 involves merging the training and validation sets from S2 to form a larger augmented training set. Using the optimal parameter combination found in hyperparameter tuning, a new prediction model is retrained.

[0080]

[0081] In the formula, x is a wind field vector containing three wind components: u, v, and w, and t is time. The new forecast model is evaluated using the test set in S2, and the new MAE and RMSE are calculated. They should be lower than the baseline 1 in S313. At the same time, this performance index is used as the baseline 2 that the subsequent Kalman filter dynamic correction needs to surpass.

[0082] Dynamic assimilation system integration step S4

[0083] The static optimal forecast model is used as the core prediction engine and deeply integrated with the ensemble Kalman filter (EnKF) framework to construct a dynamic assimilation forecast system. The performance of the dynamic assimilation forecast system is evaluated and verified on the test set to ensure that it meets the preset accuracy requirements.

[0084] The dynamic assimilation system integration step S4 includes an initialization step S41 and a cyclic assimilation step S42. In the initialization step S41, the observed wind field y(0) at t=0 is obtained, and N random perturbations are added to generate N set members.

[0085] ;

[0086] in, Let x be a vector representing the wind field, containing three wind components: u, v, and w. Let a represent the analysis. Let represent the wind field vector of the i-th member at time t=0, which has been subjected to random perturbation. The observed wind field at time t=0 R is the random noise drawn from a Gaussian distribution N(0, R) with zero mean. R is the observation error covariance matrix, which represents the degree of uncertainty of lidar observation. Usually, the diagonal elements of R are set as the standard deviation of the observation error.

[0087] In the iterative assimilation step S42, t=1,2,3,…, based on the static optimal forecast model in S3, the latest observation data is periodically incorporated, and the ensemble Kalman filter algorithm is used to continuously correct and optimize the output of the deep learning wind field forecast model. For example… Figure 3 As shown, the cyclic assimilation step S42 includes the following steps:

[0088] In step S421 of calculating the noise covariance matrix, the static optimal prediction model from S3 is used to push all set members from the previous time step (t-1) forward to the current time step (t). Simultaneously, to compensate for the shortcomings of a single model and prevent set collapse (over-concentration of set members), a random perturbation is added to each member, i.e.

[0089] ;

[0090] in, Let x be a wind field vector containing three wind components: u, v, and w; a represent analysis; f represent forecast; and M be the final forecast model output by S3. ; indicates that the analysis state of the i-th member at time t-1 is predicted. It is process noise, randomly sampled from a zero-mean Gaussian distribution N(0, Q).

[0091] The process noise covariance matrix Q represents the degree of distrust in the deep learning model M. It needs to be adjusted in advance through multiple wind field forecast evaluation results. The larger the error of the evaluation results, the larger Q needs to be, and vice versa.

[0092] The Kalman gain is calculated in step S422. The Kalman gain K is essentially a weighting coefficient. Its calculation can be viewed as: K ≈ (model uncertainty) / (model uncertainty + observation uncertainty). The formula is as follows:

[0093] ;

[0094] Where H is the observation operator, a matrix that maps the state vector from the "model space" to the "observation space," where the model space is... For the uncertainty of model prediction, R represents the uncertainty of the model prediction in the observation space, and R represents the observation error covariance.

[0095] In practice, the Kalman gain is usually approximated using a set of methods. The specific steps are as follows:

[0096] (1) The forecast set is obtained based on the perturbation applied at time t. Calculate the forecast mean:

[0097] ;

[0098] And calculate the perturbation matrix:

[0099] ;

[0100] (2) Map the disturbance to the observation space:

[0101] ;

[0102] If H is the identity matrix, then HA' is equal to A'.

[0103] (3) The Kalman gain is approximated by using set theory to calculate each part, and finally the approximate Kalman gain is obtained:

[0104] ;

[0105] in, Depend on To approximate, Depend on To approximate;

[0106] Update each member in step S423 to obtain the true wind field y(obs)(t) at time t. To avoid numerical problems and preserve the statistical properties of the set, the "perturbation observation" method is used for updating.

[0107] First, generate N observations with perturbations: ,in ;

[0108] Then, update each member i = 1, ..., N: .

[0109] S424, based on the latest analysis set for each member, obtains the set of analyses closest to the true value:

[0110] ;

[0111] As input to the next round of forecasting (time t+1) (S4201), the weighted average of each member can be calculated to obtain the final forecast result at time t. .

[0112] System performance verification step S5

[0113] The performance of the entire assimilation prediction system is evaluated using the test set in S2. The mean absolute error (MAE) and root mean square error (RMSE) are calculated, and they are acceptable if they are below the baseline 2 in S3.

[0114] Business-oriented operation steps S6

[0115] In operational use, the dynamic assimilation forecasting system fills in and identifies grid locations with missing observation data using NaN values ​​in the final forecast product, generating the final wind field forecast result.

[0116] Load the latest low-altitude 3D wind field, generate an initial set, perform single-step prediction using the optimal model, and then use EnKF for error correction. Iterate through single-step prediction and EnKF assimilation, gradually advancing the forecast time to 2 hours. At the same time, replace the missing measurement points in each layer with NaN values, and output the final wind field forecast results.

[0117] It should be noted that the examples and application scenarios implemented by the above modules and corresponding steps are the same, but are not limited to the content disclosed in the above implementation examples. It should also be noted that the above modules, as part of the device, can be implemented through software or hardware, where the hardware environment includes a network environment.

[0118] Example 2

[0119] Embodiment 2 of the present invention provides a near-term forecasting device 1 for low-altitude three-dimensional wind fields, characterized in that it includes: a data collection and preprocessing unit 10, a feature engineering and target construction unit 20, a static model construction and optimization unit 30, a dynamic assimilation system integration unit 40, a system performance verification unit 50, and an operational unit 60.

[0120] First, the data collection and preprocessing unit 10 is used to collect low-altitude three-dimensional wind field observation data in real time through the API interface, store it as a NetCDF(nc) file, and perform preprocessing such as cleaning, standardization, and interpolation on the data.

[0121] Secondly, the feature engineering and target construction unit 20 is used to construct spatiotemporal feature engineering, extract key factors that can reflect the dynamic evolution of the wind field, and generate training, validation, and test sets according to the time sequence principle. Specifically, the feature engineering and target construction unit is used to extract key factors including periodic time features, lagged variables, sliding window statistics, wind speed change trend features, static spatial features, neighborhood features, spatial gradient features, and physical features.

[0122] Furthermore, the static model construction and optimization unit 30 is used to train the model using the LightGBM framework and fine-tune the core hyperparameters using the Bayesian optimization algorithm. Based on the optimal parameter combination, it is retrained on the extended dataset (training set + validation set) to obtain a static optimal prediction model. The static model construction and optimization unit includes a hyperparameter search module, which uses the Bayesian optimization method to automatically search for the number of iterations, learning rate, number of leaf nodes, maximum tree depth, and regularization parameters of LightGBM to obtain the parameter combination that minimizes the root mean square error (RMSE) on the validation set.

[0123] Furthermore, the dynamic assimilation system integration unit 40 is used to deeply fuse the static optimal forecast model with the ensemble Kalman filter framework to construct a complete dynamic assimilation forecast system. The Kalman gain calculated by the dynamic assimilation system integration unit is based on the covariance between the perturbation matrix of the forecast ensemble and the observation perturbation matrix, and uses this to construct the ensemble approximation gain matrix.

[0124] Furthermore, the system performance verification unit 50 is used to evaluate and verify the performance of the entire system on the test set to ensure that it meets the preset accuracy requirements.

[0125] Furthermore, the operational unit 60 is used to operationalize the dynamic assimilation forecast system. For grid locations where observation data is missing, NaN values ​​are used to fill and mark them in the final forecast product to generate the final wind field forecast result.

[0126] In summary, this invention directly utilizes high-density observed wind field data, combined with advanced machine learning methods and data assimilation techniques, to realize an operational system for refined wind field forecasting at an ultra-small scale, effectively solving the current problem of insufficient refined forecasting capabilities for low-altitude three-dimensional wind fields.

[0127] While various embodiments of the invention have been described above, it should be understood that these are merely exemplary and not limiting. Therefore, the scope of the invention should not be limited to the above embodiments, but should be defined solely by the following claims and their equivalents.

Claims

1. A method for nowcasting low-level three-dimensional wind field, characterized in that, The method comprises the following steps: a data collection and preprocessing step, which collects low-altitude three-dimensional wind field observation data in real time through an API interface, stores the data as a NetCDF (nc) file, and performs cleaning, standardization and interpolation processing on the observation data to generate a spatially continuous and complete three-dimensional wind field data matrix; a feature engineering and target construction step, which constructs a spatio-temporal feature engineering, extracts key factors that can reflect the dynamic evolution of the wind field, and divides historical data into mutually independent training set, validation set and test set according to the time sequence principle; a static model construction and optimization step, which trains a model using the LightGBM framework, fine-tunes the core hyperparameters with the help of a Bayesian optimization algorithm, re-trains the model on an extended data set composed of the training set and the validation set based on the optimal parameter combination, and obtains a static optimal prediction model; a dynamic assimilation system integration step, which integrates the static optimal prediction model as the core prediction engine with the ensemble Kalman filter framework to construct a dynamic assimilation prediction system; a system performance verification step, which uses the test set in the feature engineering and target construction step to evaluate the performance of the entire assimilation prediction system, calculates the mean absolute error (MAE) and root mean square error (RMSE), and is lower than the baseline 2 in the static model construction and optimization step; and a business operation step, which business operation the dynamic assimilation prediction system, fills and identifies NaN values in the final prediction product for grid locations with missing observation data, and generates the final wind field prediction result.

2. The nowcast method according to claim 1, characterized in that, The interpolation processing in the data collection and preprocessing step adopts the inverse distance weighted interpolation method, which comprises: finding all missing points nan_indexs, determining a three-dimensional cubic space with a missing point as the center and a side length of 2R+1 as the maximum neighborhood range based on each missing point (x, y, z), where R is a preset maximum search radius; identifying all non-missing valid data points within the maximum neighborhood range to form a neighboring point set for weighted calculation; calculating the distance between the missing point and the valid data point, and calculating the interpolation weight according to the distance; and calculating the average interpolation weight and filling it into the corresponding missing point.

3. The nowcast method of claim 1, wherein, The spatio-temporal feature engineering in the static model construction and optimization step extracts features including a time feature and a space feature, the time feature includes periodic time feature, lag variable, sliding window statistics and wind speed change trend feature; the space feature includes static space feature, neighborhood feature, spatial gradient feature and divergence or vertical vorticity physical feature.

4. The nowcast method of claim 1, wherein, The static model construction and optimization step comprises: a preliminary prediction model training step, which uses empirical parameters to establish a baseline prediction model based on LightGBM on the training set, and evaluates the performance on the validation set to obtain baseline 1; The hyperparameter tuning step optimizes core performance parameters, including the number of iterations n_estimators, the learning rate learning_rate, the number of tree leaf nodes num_leaves, the maximum depth of the tree max_depth, and the regularization and anti-overfitting parameters, using Bayesian optimization to determine the optimal parameter combination that minimizes the RMSE on the validation set; The final model training step combines the training set and the validation set to form an enhanced training set, re-trains a new prediction model using the optimal parameter combination, and evaluates the performance on the test set to obtain a baseline 2.

5. The nowcast method of claim 1, wherein, The dynamic assimilation system integration step includes: An initialization step that obtains the initial observation wind field y(0) at the initial time and adds N random perturbations to generate N ensemble members; A cyclic assimilation step that uses the static optimal prediction model to continuously correct and optimize the output of the deep learning wind field prediction model by periodically incorporating the latest observation data using the ensemble Kalman filter algorithm; Where the static optimal prediction model is used to extrapolate all ensemble members at the previous time (t-1) to the current time (t), and to compensate for the shortcomings of a single model and prevent ensemble collapse, a random perturbation is added to each member.

6. The nowcasting method of claim 5, wherein, The Kalman gain K of the dynamic assimilation system integration step is calculated using ensemble approximation, which involves the perturbation matrix of the prediction ensemble and the matrix that maps the perturbation to the observation space.

7. A device for nowcasting low-level three-dimensional wind field, characterized by, It includes: A data collection and preprocessing unit for real-time collection of low-altitude three-dimensional wind field observation data through an API interface and storage as NetCDF (nc) files, and preprocessing of the data such as cleaning, standardization, and interpolation; A feature engineering and target construction unit for constructing spatio-temporal feature engineering, extracting key factors that can reflect the dynamic evolution of wind fields, and generating training, validation, and test sets based on time series principles; A static model construction and optimization unit for model training using the LightGBM framework and fine-tuning of core hyperparameters using Bayesian optimization algorithms, re-training based on the optimal parameter combination on an extended data set to obtain a static optimal prediction model; A dynamic assimilation system integration unit for deep integration of a static optimal prediction model with the ensemble Kalman filter framework to build a complete dynamic assimilation prediction system; A system performance verification unit for evaluating and verifying the performance of the entire system on the test set to ensure that it meets the predetermined accuracy requirements; And A business operation unit for business operation of the dynamic assimilation prediction system, filling and identifying NaN values in the final prediction product for grid locations with missing observation data to generate the final wind field prediction result.

8. The nowcasting apparatus according to claim 7, characterized by The feature engineering and target construction unit is specifically configured to extract key factors including periodic temporal features, lag variables, sliding window statistics, wind speed change trend features, static spatial features, neighborhood features, spatial gradient features, and physical features.

9. The nowcast apparatus according to claim 7, wherein The static model construction and optimization unit comprises a hyperparameter search module, which is configured to automatically search for the number of iterations, the learning rate, the number of leaf nodes, the maximum depth of the tree and the regularization parameter of LightGBM by using a Bayesian optimization method, so as to obtain a combination of parameters that minimizes the root mean square error (RMSE) on the validation set.

10. The nowcast apparatus according to claim 7, wherein The Kalman gain calculated by the dynamic assimilation system integration unit is based on the covariance of the disturbance matrix of the prediction ensemble and the observation disturbance matrix, and the gain matrix of the ensemble approximation is constructed based on the covariance.

Images