A correction method for offshore reanalysis wind field data based on buoy observation data

Abstract

The present invention provides a method for correcting marine reanalysis wind field data based on buoy observation data, which belongs to the technical field of meteorology. The present invention proposes firstly to remove outliers and perform quality control on the buoy observation data, and adopt a bilinear interpolation method to perform spatiotemporal matching; then to construct a dynamic time window, and adaptively adjust the time accuracy according to the type of meteorological event; then to calculate the wind speed and wind direction errors, and extract spatiotemporal characteristic parameters; the core step is to optimize Gaussian process regression using an ocean physics constrained neural network model, which includes a surface layer physical process encoding layer, a boundary layer dynamics embedding layer and a multi-head sparse attention layer; to construct a spatiotemporal covariance model to generate an error field; to correct the reanalysis wind field data and output an uncertainty assessment; and to solve the technical problem that the reanalysis wind field data has significant spatiotemporal errors compared with the actual observation values ​​in a complex ocean environment.

Description

Technical Field

[0001] The present invention belongs to the technical field of meteorology, and in particular relates to a method for correcting offshore reanalysis wind field data based on buoy observation data. Background Art

[0002] Offshore wind farm data is crucial foundational information for marine engineering construction, navigation safety, marine energy development, and marine meteorological forecasting. Its accuracy and reliability directly impact the effectiveness and safety of related applications. Currently, offshore wind farm data is primarily acquired through two methods: field observations from buoys and meteorological reanalysis data. Buoy observations can directly capture sea surface wind farm information with high temporal resolution and measurement accuracy, but are limited by high deployment costs and limited coverage, making comprehensive monitoring of vast ocean areas impossible. Reanalysis wind farm data, on the other hand, is generated by assimilating multi-source observational data using numerical models. It offers the advantage of continuous global coverage and has become an important data source for marine applications.

[0003] However, due to the complexity of the ocean environment and the limitations of numerical models, reanalysis wind data often exhibit systematic biases in practical applications. In particular, under conditions of complex nearshore terrain, strong air-sea interactions, and extreme weather events, significant spatial and temporal errors can be observed between reanalysis wind data and buoy measurements. Studies have shown that these errors exhibit significant non-uniformity and nonlinear characteristics across different sea areas, seasons, and even time of day, seriously impacting the reliability of reanalysis wind data in high-precision applications.

[0004] At present, the correction of reanalysis wind field data mainly adopts statistical interpolation or simple regression methods, such as linear regression, polynomial fitting, optimal interpolation and other technologies. These methods are usually based on historical statistical characteristics within a fixed time window to establish an empirical relationship between buoy observations and reanalysis wind field data. However, this simple statistical correlation is difficult to effectively capture the spatiotemporal variation characteristics of wind field errors in complex ocean environments. Especially under rapidly changing meteorological conditions such as typhoons and severe convection, the error distribution between reanalysis wind field data and actual observations shows significant non-stationarity and dependence on physical processes, and the effect of traditional correction methods is very limited. In other words, there is a technical problem in the existing technology that there are significant spatiotemporal errors between reanalysis wind field data and actual observations in complex ocean environments. Summary of the Invention

[0005] In view of this, the present invention provides a method for correcting offshore reanalysis wind field data based on buoy observation data, which can solve the technical problem in the prior art that reanalysis wind field data has significant temporal and spatial errors compared with actual observation values ​​in complex ocean environments.

[0006] The present invention is implemented as follows: The present invention provides a method for correcting offshore reanalysis wind field data based on buoy observation data, comprising: receiving buoy observation data and reanalysis wind field data, eliminating outliers and performing quality control processing on the buoy observation data; performing spatiotemporal matching on the buoy observation data and the reanalysis wind field data; constructing a dynamic time window; calculating wind speed error and wind direction error; extracting spatiotemporal characteristic parameters; optimizing Gaussian process regression using an ocean physics constraint neural network model, wherein the ocean physics constraint neural network model adopts a sparse attention mechanism to process multi-scale spatiotemporal characteristics, and the coupling index of the ocean physics constraint neural network model is determined by a boundary layer stability judgment function; establishing a spatiotemporal covariance model; and correcting the reanalysis wind field data using an error field.

[0007] Among them, the outlier elimination and quality control processing are specifically to identify abnormal points of wind speed, wind direction, air pressure and other parameters in the buoy observation data by setting physical thresholds and combining the box plot method. Data points that exceed the reasonable range are marked as outliers and do not participate in subsequent calculations.

[0008] The time-space matching is specifically to extract the wind field data of the reanalysis wind field data grid points corresponding to the buoy position using a bilinear interpolation method to obtain time-space matching data.

[0009] Among them, the bilinear interpolation method specifically achieves spatial resolution coordination by calculating the weighted average of the four nearest reanalysis wind field data grid points around the buoy position, where the weight is inversely proportional to the distance from the buoy to the reanalysis wind field data grid point.

[0010] Among them, the dynamic time window automatically adjusts the width of the dynamic time window to 15 minutes to 1 hour according to the type of meteorological event, so as to achieve precise alignment of the time dimension.

[0011] Among them, the wind speed error is the wind speed of the buoy observation data minus the wind speed of the reanalysis wind field data, and the wind direction error is the wind direction of the buoy observation data minus the wind direction of the reanalysis wind field data; the wind direction error is specifically calculated by calculating the angular difference between the wind direction of the buoy observation data and the wind direction of the reanalysis wind field data, taking into account the angle cycle characteristics, and taking the minimum angle difference whose absolute value does not exceed 180 degrees.

[0012] Among them, the dynamic time window is specifically a time period that is adaptively adjusted based on the type of meteorological event. Under normal weather conditions, a matching window of plus or minus 15 minutes is adopted, which is shortened to plus or minus 5 minutes in a rapidly changing typhoon meteorological environment, and extended to plus or minus 1 hour under stable meteorological conditions.

[0013] The spatiotemporal characteristic parameters include sea surface temperature, pressure gradient, wind field curvature, seasonal cycle coefficient and diurnal variation index.

[0014] Among them, the structure of the ocean physical constraint neural network model is a multi-layer encoding and decoding architecture, which includes three main parts: surface layer physical process encoding layer, boundary layer dynamics embedding layer and multi-head sparse attention layer.

[0015] Among them, the surface layer physical process encoding layer integrates the Richardson number and roughness parameter of the sea-air interface to convert the physical field characteristics into potential representations. The boundary layer dynamics embedding layer contains five sets of parameterized equations to describe the wind profile characteristics and turbulence effects under different stability conditions. The multi-head sparse attention layer consists of eight attention heads, each of which focuses on the feature correlation of different spatial scales.

[0016] Among them, the spatiotemporal characteristic parameters are specifically the key physical quantities and time characteristics that affect the wind field error distribution. The sea surface temperature is obtained through satellite remote sensing, the pressure gradient is calculated by the pressure difference between adjacent grid points, the wind field curvature reflects the spatial change rate of the wind field, the seasonal cycle coefficient is extracted through Fourier transform to extract the annual change characteristics, and the diurnal change index is fitted with the intraday change characteristics by a sine function.

[0017] The space-time covariance model uses the Matern3 / 2 kernel function to construct the error field, the spatial scale of the space-time covariance model is set to 50 kilometers, and the time scale of the space-time covariance model is set to 6 hours.

[0018] The wind speed of the corrected wind field data is equal to the wind speed of the reanalyzed wind field data plus the wind speed error field value at the corresponding position and time.

[0019] This also includes generating uncertainty assessment results of corrected wind field data based on Gaussian process regression variance output, and calculating and outputting 95% confidence intervals.

[0020] The present invention achieves accurate correction of reanalysis wind field data by constructing a dynamic time window and an ocean physics constraint neural network model. This method adaptively adjusts the time matching accuracy according to different meteorological event types, using a wider time window under normal weather conditions and a narrower time window in rapidly changing environments such as typhoons, thereby ensuring the accuracy of the corrected basic data from the time dimension. The present invention establishes a correlation mechanism between the wind field error distribution and the physical characteristics of the ocean environment by extracting spatiotemporal characteristic parameters such as sea surface temperature, pressure gradient, wind field curvature, seasonal cycle coefficient, and diurnal variation index. In particular, the introduced ocean physics constraint neural network model, through a multi-level structure of surface layer physical process encoding layer, boundary layer dynamics embedding layer and multi-head sparse attention layer, effectively captures the nonlinear error distribution characteristics in complex ocean environments and achieves accurate prediction of wind field errors. By combining Gaussian process regression with a physical constraint neural network, the present invention not only improves the accuracy of the corrected wind field data, but also provides a reliable uncertainty assessment, providing a risk control basis for the application of wind field data. Experimental verification shows that this method can significantly reduce the systematic deviation of reanalysis wind field data under various marine environmental conditions, especially during extreme meteorological events, and effectively solve the technical problem that reanalysis wind field data have significant temporal and spatial errors compared with actual observations in complex marine environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0023] like Figure 1 FIG. 1 is a flow chart of a method for correcting offshore reanalysis wind field data based on buoy observation data provided by the present invention. The method comprises the following steps:

[0024] S01, receiving buoy observation data and reanalyzed wind field data, performing outlier removal and quality control on the buoy observation data, and marking the buoy observation data with wind speeds greater than 50 meters per second or that do not conform to physical laws as invalid data;

[0025] S02, performing spatiotemporal matching on the buoy observation data and the reanalysis wind field data, and extracting wind field data of the grid points of the reanalysis wind field data corresponding to the buoy position using a bilinear interpolation method to obtain spatiotemporal matching data;

[0026] S03. Construct a dynamic time window based on the spatiotemporal matching data, and automatically adjust the width of the dynamic time window to 15 minutes to 1 hour according to the type of meteorological event to achieve precise alignment of the time dimension;

[0027] S04. Calculating a wind speed error and a wind direction error at each matching point based on the spatiotemporal matching data, wherein the wind speed error is the wind speed of the buoy observation data minus the wind speed of the reanalyzed wind field data, and the wind direction error is the wind direction of the buoy observation data minus the wind direction of the reanalyzed wind field data;

[0028] S05. Extracting spatiotemporal characteristic parameters that affect error distribution based on the wind speed error and the wind direction error, where the spatiotemporal characteristic parameters include sea surface temperature, pressure gradient, wind field curvature, seasonal cycle coefficient, and diurnal variation index;

[0029] S06. Optimizing Gaussian process regression using a pre-trained ocean physics constraint neural network model, wherein the ocean physics constraint neural network model uses a sparse attention mechanism to process multi-scale spatiotemporal features, and wherein a coupling index of the ocean physics constraint neural network model is determined by a boundary layer stability determination function;

[0030] S07, establishing a spatiotemporal covariance model based on the Gaussian process regression, constructing an error field using the Matern3 / 2 kernel function, wherein the spatial scale of the spatiotemporal covariance model is set to 50 kilometers, and the time scale of the spatiotemporal covariance model is set to 6 hours;

[0031] S08. Correcting the reanalyzed wind field data using the error field to obtain corrected wind field data, wherein the wind speed of the corrected wind field data is equal to the wind speed of the reanalyzed wind field data plus the wind speed error field value at the corresponding position and time;

[0032] S09. Generating an uncertainty assessment result of the corrected wind field data based on the Gaussian process regression variance output, and calculating and outputting a 95% confidence interval;

[0033] S10. Use the leave-one-buoy cross-validation method to evaluate the generalization ability of the ocean physical constraint neural network model, calculate the root mean square error, deviation and correlation coefficient of the corrected wind field data and the buoy observation data, and verify the correction effect.

[0034] Among them, outlier elimination is specifically to identify abnormal points of wind speed, wind direction, air pressure and other parameters in the buoy observation data by setting physical thresholds and combining the box plot method. Data points that exceed the reasonable range are marked as outliers and do not participate in subsequent calculations.

[0035] Among them, the bilinear interpolation method specifically calculates the weighted average of the four nearest reanalysis wind field data grid points around the buoy position, and the weight is inversely proportional to the distance from the buoy to the reanalysis wind field data grid point, so as to achieve spatial resolution coordination.

[0036] Among them, the dynamic time window is a time period that is adaptively adjusted based on the type of meteorological event. Under normal weather conditions, a matching window of plus or minus 15 minutes is used, which is shortened to plus or minus 5 minutes in the rapidly changing meteorological environment of a typhoon, and extended to plus or minus 1 hour under stable meteorological conditions.

[0037] The wind direction error is specifically calculated as the angular difference between the wind direction of the buoy observation data and the wind direction of the reanalyzed wind field data, and the minimum angular difference whose absolute value does not exceed 180 degrees is taken into account in considering the angle cycle characteristics.

[0038] Among them, the spatiotemporal characteristic parameters are specifically the key physical quantities and time characteristics that affect the wind field error distribution. The sea surface temperature is obtained through satellite remote sensing, the pressure gradient is calculated by the pressure difference between adjacent grid points, the wind field curvature reflects the spatial change rate of the wind field, the seasonal cycle coefficient is extracted through Fourier transform to extract the annual change characteristics, and the diurnal change index is fitted with the intraday change characteristics by sine function.

[0039] Among them, the specific structure of the ocean physical constraint neural network model is a multi-layer encoding and decoding architecture, which includes three main parts: a surface layer physical process encoding layer, a boundary layer dynamics embedding layer, and a multi-head sparse attention layer. The surface layer physical process encoding layer integrates the Richardson number and roughness parameter of the sea-air interface to convert the physical field characteristics into potential representations. The boundary layer dynamics embedding layer contains five sets of parameterized equations to describe the wind profile characteristics and turbulence effects under different stability conditions. The multi-head sparse attention layer consists of eight attention heads, each head focuses on the feature correlation of different spatial scales. The number of heads in the multi-head sparse attention layer is determined by the gradient of the sea surface temperature, and the attention sparsity of the multi-head sparse attention layer is determined by the curvature and intensity of the wind field. The output layer of the ocean physical constraint neural network model is integrated with the Gaussian process regression results through residual connection to realize the dual constraints of physical laws and data-driven.

[0040] Among them, the steps of establishing the training data set of the ocean physical constraint neural network model specifically include screening out ten years of high-quality data from the global buoy observation network as the basic data set, stratifying the basic data set according to season and sea area to ensure that the training sample space is evenly distributed, and constructing a feature vector for each training sample including wind speed and direction differences and related physical field parameters. The feature vector dimension is 128 dimensions and includes near-shore surface layer structure parameters and turbulence intensity indicators. The error correction value of the labeled sample is calculated by comparing the historical buoy observation records with the reanalysis data of the corresponding position. The time span of the training data set covers multiple typical meteorological events, including typhoon season alternation and extreme weather processes to ensure the generalization ability of the ocean physical constraint neural network model. For sparse areas, the physical model simulation results are used for data enhancement to generate synthetic training samples to make up for the lack of observation data.

[0041] Among them, the steps of training the ocean physical constraint neural network model specifically include first using an unsupervised pre-training method to learn the intrinsic representation of the physical structure of the sea-air interface. The loss function in the pre-training stage includes a physical consistency constraint term to ensure that the output of the ocean physical constraint neural network model does not violate the basic ocean dynamics laws. Subsequently, supervised learning is used to fine-tune the parameters of the ocean physical constraint neural network model, where the loss function is designed as a weighted combination of mean square error and physical constraints to balance data fitting and physical rationality. The training process applies a dynamic learning rate adjustment strategy and adopts different weight update rules under different stability conditions. Batch training is adopted, and each batch contains 64 samples, of which the proportion of extreme weather samples is fixed at 25%. The ocean physical constraint neural network model is trained for 500 rounds of iterations and then fine-tuned. Finally, the selection of the ocean physical constraint neural network model is determined based on the dual indicators of verification set performance and physical consistency.

[0042] Among them, Gaussian process regression is specifically a probabilistic machine learning model. By defining a kernel function to describe the similarity between data points and establishing a function distribution in a continuous space, it can simultaneously give a predicted value and its uncertainty estimate.

[0043] Among them, the boundary layer stability judgment function is specifically a Richardson number function calculated based on the sea-air temperature difference, wind speed gradient and pressure field. The coupling strength of the physical equation is determined by analyzing the atmospheric stability state. The boundary layer stability judgment function first calculates the Richardson number of the sea-air interface, and then constructs a judgment vector in combination with the wave height and turbulent kinetic energy parameters. The judgment vector consists of five components corresponding to different stability intervals. According to the value of the judgment vector, the coupling index of the boundary layer physical equation in the ocean physical constraint neural network model is dynamically adjusted, the weight of the turbulence closure equation is increased under strong instability conditions, and the influence of the thermodynamic equation is enhanced under stable stratification.

[0044] Among them, the Matern3 / 2 kernel function is a covariance function widely used in geophysical data. Compared with the Gaussian kernel function, it has better adaptability to the non-smooth characteristics of data and can more accurately describe the local change characteristics of the wind field.

[0045] The error field specifically refers to the distribution field of wind speed error and wind direction error at any position and time point in the ocean space based on the spatiotemporal covariance model, which is used to correct the reanalysis wind field data.

[0046] The wind speed error field value specifically refers to the wind speed error prediction value given by the error field at a certain time and position point, which is added to the wind speed of the reanalyzed wind field data as a correction amount.

[0047] Among them, the leave-one-buoy cross-validation method specifically excludes all the data of one buoy from the training data set each time as a test set, uses the remaining buoy data to train the ocean physical constraint neural network model and predicts the wind field error at the excluded buoy position, and repeats the cycle until each buoy is used as a test set once, and finally comprehensively evaluates the prediction ability of the ocean physical constraint neural network model at unknown positions.

[0048] The specific implementation of the above steps is described in detail below.

[0049] The specific implementation of step S01 involves processing buoy observation data through a multi-level quality inspection system to ensure data reliability. First, raw data uploaded in real time from globally distributed buoy observation sites is received, including parameters such as wind speed, wind direction, air pressure, and sea surface temperature. Subsequently, a physical threshold method is used for preliminary screening, marking data with wind speeds greater than 50 meters per second as outliers, as wind speeds exceeding this threshold rarely occur in the natural environment. Wind direction data is also checked to ensure it is within the range of 0 to 360 degrees; those outside this range are marked as invalid. A boxplot method is then used to identify statistically significant outliers. The quartiles Q1 and Q3 of each parameter are calculated, and data points outside the range of Q1 - 1.5 × IQR or Q3 + 1.5 × IQR are marked as potential outliers, where IQR is the interquartile range. For air pressure parameters, a reasonable rate of change threshold of 3 hPa / h is set; data exceeding this rate of change requires further review. Finally, physical consistency is checked, for example, to verify that the relationships between wind speed and significant wave height, and wind direction and wave direction, conform to basic principles of marine meteorology. Through the above processing, high-quality data is retained for subsequent correction model construction, which improves the reliability of the model input data and thus enhances the accuracy of the final correction results.

[0050] The specific implementation of step S02 is to achieve accurate spatial matching of the buoy observation data and the reanalysis wind field data through bilinear interpolation. First, the latitude and longitude coordinates of the buoy are determined, and then the four nearest reanalysis wind field data grid points around these coordinates are identified. These grid points form a rectangular area containing the buoy position. For each buoy position, its distance to the four grid points is calculated, and a weight coefficient is determined based on the inverse ratio of the distance. The specific weight calculation formula is: ,in represents the distance from the buoy to the i-th grid point, and the sum of the weights is 1. Using these weights, the wind field data of the four grid points are weighted averaged to obtain the interpolated wind field data at the buoy position. This step solves the problem of inconsistent positions between the buoy observation point and the reanalysis wind field grid point, achieves the coordination and unification of spatial dimensions, lays the foundation for subsequent error analysis, and ensures that the wind field information at the same location is compared. Compared with the simple nearest neighbor interpolation method, bilinear interpolation can better preserve the spatial continuity of the wind field and is suitable for expressing the spatial characteristics of the wind field under complex sea conditions.

[0051] The specific implementation of step S03 involves dynamically adjusting the time matching window width based on meteorological conditions to achieve precise alignment in the temporal dimension. First, by analyzing the current meteorological conditions, they are divided into three categories: normal weather conditions, extreme weather conditions, and stable weather environments. For normal weather conditions, the system sets a ±15-minute time window, meaning that reanalyzed wind field data within 15 minutes before and after the buoy observation time is considered a valid match. For extreme weather conditions such as typhoons, due to the rapid changes in the wind field, the system automatically narrows the time window to ±5 minutes to reduce errors caused by time differences. For stable weather conditions, such as high-pressure areas, the system expands the time window to ±1 hour, as the wind field changes slowly in these conditions. The system calculates a meteorological stability index by analyzing three indicators: wind speed standard deviation, pressure change rate, and wind direction stability. The time window is dynamically adjusted based on this index value: an index value below 0.2 is considered stable; between 0.2 and 0.8 is considered normal; and above 0.8 is considered rapidly changing. This step effectively solves the problem of difference in temporal resolution between buoy observation data and reanalysis wind field data, improves the time matching accuracy, and provides more reliable basic data for subsequent error field construction.

[0052] The specific implementation of step S04 is to calculate the wind speed error and wind direction error at the spatiotemporal matching point. The wind speed error is calculated using a simple algebraic difference method, that is, the buoy observed wind speed value is subtracted from the corresponding reanalysis wind field wind speed value. A positive value indicates that the reanalysis wind field underestimates the actual wind speed, and a negative value indicates an overestimation. The calculation of the wind direction error needs to take into account the cyclic characteristics of the angle. First, the difference between the buoy observed wind direction and the reanalysis wind field wind direction is calculated, and then the angle normalization process is applied: when the difference is greater than 180 degrees, the difference is subtracted from 360 degrees and the negative value is taken; when the difference is less than -180 degrees, the difference is added from 360 degrees and the negative value is taken. This process ensures that the wind direction error always falls within the range of -180 degrees to 180 degrees, indicating the minimum angle difference. For each matching point, the wind speed error and wind direction error are recorded simultaneously to form an error sample set. To eliminate the influence of random errors, a 5-point median filter is applied to the error values ​​at multiple consecutive time points to enhance the stability of the error signal. This step quantifies the degree of deviation of the reanalyzed wind field relative to the actual observation, providing basic data for subsequent error field construction and wind field correction, while separating the errors of the two independent components of wind speed and wind direction, making it easier to model and process them separately.

[0053] The specific implementation of step S05 is to extract the key spatiotemporal characteristic parameters that affect the wind field error distribution. First, the sea surface temperature field of the study area is extracted from the satellite remote sensing data source, with a spatial resolution of 0.05 degrees and a temporal resolution of the daily average. Then, the pressure gradient is calculated, and the pressure values ​​of adjacent grid points are processed using the central difference method. The specific formula is: ,in and Represent the rate of change of air pressure in the east-west and north-south directions respectively. The wind field curvature is calculated using the Laplace operator and is approximately expressed using the second-order derivative of the wind speed field: , where V is the wind speed scalar field. The seasonal period coefficient is obtained by Fourier transforming the historical wind field data, extracting the amplitude and phase corresponding to the main frequency components, and constructing a The seasonal variation term is , where A is the annual amplitude, t is the current date, and φ is the initial phase. The diurnal variation index uses the cosine function Fitting intraday variation characteristics, where B is the daily amplitude, t is the current hour, and θ is the phase parameter. These spatiotemporal characteristic parameters comprehensively characterize the key physical factors affecting wind field errors, providing multi-dimensional input features for the Gaussian process regression model, improving the physical rationality of error field construction and prediction accuracy.

[0054] The specific implementation of step S06 involves optimizing Gaussian process regression using an ocean physics-constrained neural network model. This model employs a multi-layer encoding-decoding architecture, comprising a surface layer physics encoding layer, a boundary layer dynamics embedding layer, and a multi-head sparse attention layer. The surface layer physics encoding layer receives the Richardson number of the air-sea interface and the sea surface roughness parameter as input and converts the physical field features into a 128-dimensional latent representation vector through a nonlinear transformation. The boundary layer dynamics embedding layer contains five sets of parameterized equations, describing wind profile characteristics and turbulence effects under different stability conditions, with each set of equations corresponding to a different atmospheric stability state. The multi-head sparse attention layer consists of eight attention heads, each responsible for capturing feature correlations at different spatial scales. The number of attention heads is determined by the magnitude of the sea surface temperature gradient, increasing to 12 heads when the gradient exceeds 0.05°C / km. The sparsity of attention is determined by the curvature and intensity of the wind field; greater wind field curvature results in lower sparsity. The sparse attention mechanism uses a threshold to filter out low-correlation connections, significantly reducing computational effort while preserving critical information. The model's coupling index is determined by the boundary layer stability determination function, which calculates a five-component determination vector based on the Richardson number, wave height, and turbulent kinetic energy parameters. When the determination vector indicates strong instability, the weight of the turbulent closure equation increases to 0.7; under stable stratification conditions, the influence of the thermodynamic equation increases to 0.6. By integrating physical laws with a data-driven approach, the model effectively improves the prediction accuracy of Gaussian process regression, particularly in regions with complex air-sea interactions and under extreme weather conditions.

[0055] The specific implementation of step S07 is to establish a spatiotemporal covariance model based on Gaussian process regression. First, define the Matern3 / 2 kernel function as the basis of the covariance function, and the function form is: , where d is the distance between two points, σ ​​is the signal variance, and l is the length scale parameter. Based on this, a spatiotemporal covariance structure is constructed, with different characteristic length parameters set for the spatial and temporal dimensions, respectively. The spatial scale is set to 50 kilometers, and the temporal scale is set to 6 hours. These two parameters are optimized from historical data using maximum likelihood estimation. After the covariance matrix is ​​constructed, sparse matrix techniques are used to process large-scale datasets, and variational inference methods are employed to accelerate the model solution. To address the characteristic differences between different regions, a local adaptive strategy is implemented, dynamically adjusting the kernel function parameters based on the regional wind field characteristics. Ultimately, the spatiotemporal covariance model outputs wind speed and wind direction error fields. Both error fields are continuously defined functions in the ocean space and time dimensions, enabling error prediction at any location and time point. This model fully utilizes the spatial correlation and temporal continuity of observation errors, providing reasonable error interpolation in areas with sparse observation points, and achieving the transformation from discrete observation points to a continuous error field.

[0056] The specific implementation of step S08 involves correcting the reanalysis wind field data using the error field constructed in step S07. The correction process first determines the spatiotemporal coordinates of the point to be corrected. The wind speed and wind direction error field values ​​corresponding to the location and time are then retrieved from the error field. Wind speed correction utilizes direct addition: Corrected wind speed = Reanalysis wind field wind speed + Wind speed error field value. Wind direction correction considers the cyclic nature of angles: Corrected wind direction = (Reanalysis wind field wind direction + Wind direction error field value) % 360, where % represents a modulo operation, ensuring the result is within the range of 0 to 360 degrees. To ensure the physical plausibility of the correction results, the wind speed correction is capped at 30% of the original wind speed. Corrected values ​​exceeding this threshold are truncated to avoid overcorrection and resulting in unreasonable results. For wind direction correction, the maximum correction angle is limited to 45 degrees to maintain the spatial continuity of the wind field. In areas with strong wind speed gradients, such as typhoons, a special processing strategy is applied, where the correction intensity decreases as the wind speed gradient increases, ensuring that the corrected wind field retains the structural characteristics of the typhoon. Through this correction process, the systematic bias in the reanalysis wind field data was systematically eliminated, and the accuracy of the wind field data was improved. The correction effect was particularly significant in areas with complex nearshore terrain and special weather systems.

[0057] The specific implementation of step S09 is to generate an uncertainty assessment result for the corrected wind field data based on the variance output of Gaussian process regression. In addition to the prediction error field mean, Gaussian process regression also outputs the prediction variance, which directly quantifies the uncertainty of the prediction. The system uses this characteristic to calculate a 95% confidence interval for each correction point. The specific formula is: , where μ is the predicted error value and σ is the predicted standard deviation. The width of the confidence interval reflects the reliability of the prediction. A narrow interval indicates high confidence, and a wide interval indicates high uncertainty. The system also generates a spatial distribution map of uncertainty, using a color gradient to visually display the spatial variation characteristics of uncertainty. In order to quantify the overall uncertainty level, the regional average uncertainty index is calculated, which is defined as the ratio of the confidence interval width to the wind speed correction amount. In areas with sparse observations, uncertainty is usually higher. The system automatically marks areas where uncertainty exceeds the threshold, and users are advised to use the correction results of these areas with caution. This step provides users with a reliability assessment of the correction results, which helps in risk control and decision support of wind field data in practical applications, especially in the fields of marine engineering safety assessment and navigation decision-making.

[0058] The specific implementation method of step S10 is to use the leave-one-buoy cross-validation method to evaluate the generalization ability of the ocean physical constraint neural network model. This method excludes all the data of a buoy from the training data set as a test set each time, uses the remaining buoy data to train the model and predict the wind field error of the excluded buoy position. In the specific implementation, first construct a buoy index list containing the identifiers of all buoys involved in the verification. Then perform a cyclic verification, select a buoy as a test buoy in each iteration, exclude all its data, and use the remaining data to train the model. Apply the model prediction error to the reanalyzed wind field data at the test buoy position and perform correction, compare the correction result with the actual observation data of the buoy, and calculate the three indicators of root mean square error, deviation and correlation coefficient. The root mean square error calculation formula is: ,in is the observed value, is the corrected predicted value. The bias is calculated as the average difference between the predicted and observed values: The correlation coefficient reflects the linear correlation between the predicted and observed values, with an ideal value of 1. This process is repeated until each buoy has been used as a test set once. Finally, the validation results from all buoys are combined to evaluate the overall generalization performance of the model. This validation method simulates the application of the model in unknown locations and objectively evaluates the actual effectiveness and reliability of the correction method, providing a scientific basis for model optimization and practical application.

[0059] The detailed structure of the ocean physics constraint neural network model is a multi-layer encoder-decoder architecture consisting of three main components: a surface layer physics encoding layer, a boundary layer dynamics embedding layer, and a multi-head sparse attention layer. The surface layer physics encoding layer consists of three fully connected layers with 256, 512, and 256 nodes, respectively. It uses the GELU activation function, accepts the Richardson number and roughness length parameter of the air-sea interface as input, and converts the physical field features into a 128-dimensional latent representation vector. Batch normalization and residual connections are used between layers to improve training stability. The boundary layer dynamics embedding layer contains five sets of parameterized equations, each corresponding to a different atmospheric stability state, graded from strongly unstable to strongly stable. Each set of equations is implemented by a three-layer fully connected network. Input parameters include altitude, wind speed, and temperature gradient, and the output is the wind profile characteristics under the corresponding stability condition. A soft switching function is used to achieve smooth transitions between different stability conditions. The multi-head sparse attention layer consists of eight attention heads, each of which focuses on feature correlations at different spatial scales. The number of heads is determined by the sea surface temperature gradient, and the sparsity of attention is determined by the wind field curvature and intensity. The query, key, and value matrices for each attention head have a dimension of 64, and attention weights are calculated via scaled dot products. The model output layer is fused with the Gaussian process regression results via a residual connection, achieving dual constraints based on physical laws and data-driven approaches.

[0060] The detailed steps for establishing a training dataset for the ocean physics-constrained neural network model include: First, a decade of high-quality data was selected from the global buoy observation network as the base dataset. Observation records from 387 buoys from 2010 to 2020 were selected, totaling over 78 million samples. The dataset was stratified by season and ocean area to ensure uniform spatial distribution of training samples, with each 5-degree x 5-degree grid containing at least 1,000 training samples. For each training sample, a 128-dimensional feature vector was constructed, containing wind speed and direction differences and related physical field parameters, including nearshore surface layer structure parameters and turbulence intensity indicators. Error corrections for annotated samples were calculated by comparing historical buoy observations with reanalysis data at the corresponding locations. The training dataset spanned a number of typical meteorological events, including typhoons, seasonal transitions, and extreme weather events, to ensure model generalization. In sparsely populated areas, data augmentation was performed using physical model simulation results to generate synthetic training samples to compensate for insufficient observational data. A total of approximately 500,000 synthetic samples were generated, primarily in high-latitude regions of the Southern Hemisphere and in the open ocean. The training data is divided into a training set, a validation set, and a test set at a ratio of 80%, 10%, and 10%, respectively, for model parameter optimization, hyperparameter adjustment, and final performance evaluation. In particular, when dividing the dataset, we ensure that data from the same buoy appears in only one dataset to avoid overestimation of performance due to data leakage.

[0061] Specifically, the core principle of this invention to solve the problem of significant spatiotemporal errors in reanalyzed wind field data in complex ocean environments lies in establishing a deep fusion mechanism of physical process constraints and advanced statistical learning. This mechanism is achieved through the following key technical links:

[0062] First, the present invention introduces a dynamic time window mechanism to dynamically adjust the matching accuracy between buoy observations and reanalysis wind data based on the characteristics of the meteorological system. This design is based on the fact that different meteorological systems have significantly different temporal evolution characteristics: in stable meteorological conditions, wind field changes relatively slowly, so a wider time window (e.g., ±1 hour) can be used for matching. In rapidly changing environments such as typhoons, wind field characteristics can change dramatically in a short period of time, requiring a narrower time window (e.g., ±5 minutes) to ensure matching accuracy. This adaptive adjustment strategy fundamentally ensures the temporal alignment accuracy of the calibration base data.

[0063] Secondly, the ocean physics constraint neural network model constructed by the present invention embeds the knowledge of ocean boundary layer physical processes into a data-driven learning framework. The surface layer physical process encoding layer integrates the Richardson number and roughness parameter of the air-sea interface to convert the physical field characteristics into a potential representation that can reflect the state of air-sea interaction; the boundary layer dynamics embedding layer describes the wind profile characteristics and turbulence effects under different stability conditions through five sets of parameterized equations, effectively capturing the vertical structural characteristics of the wind field in the boundary layer; the multi-head sparse attention layer can focus on the feature correlation of different spatial scales and adapt to the interactive characteristics of multi-scale physical processes in the ocean environment. This multi-level structural design enables the model to have the ability to describe the error distribution of the wind field in a complex ocean environment.

[0064] Third, the present invention dynamically adjusts the strength of physical constraints through a boundary layer stability judgment function, achieving a balanced optimization of physical laws and data characteristics. This function calculates the Richardson number based on the sea-air temperature difference, wind speed gradient, and pressure field, and constructs a stability judgment vector based on wave height and turbulent kinetic energy parameters, thereby dynamically adjusting the coupling index of the physical equations in the neural network. This adaptive adjustment mechanism increases the weight of the turbulent closed equation under strong instability conditions, while enhancing the influence of the thermodynamic equation in a stable stratified environment. This ensures that the model can select the optimal physical constraint method for different sea-air interaction states.

[0065] Fourth, the present invention uses the Matern3 / 2 kernel function to construct a spatiotemporal covariance model of the error field. Compared to the traditional Gaussian kernel function, this kernel function is more adaptable to the non-smooth characteristics of data and can more accurately describe the local variations in wind fields in areas with significant gradients, such as circulation boundaries and frontal regions. This kernel function design is highly consistent with the physical characteristics of ocean wind fields and improves the modeling capabilities of Gaussian process regression in complex wind field environments.

[0066] Finally, the present invention fuses the neural network output with the Gaussian process regression results through a residual connection, achieving dual optimization of physical constraints and statistical properties. This fusion strategy preserves both the physical model's prior knowledge of the ocean wind field structure and the statistical model's ability to fit the observed data, effectively overcoming the limitations of a single method and improving the overall performance of wind field correction.

[0067] A specific embodiment 1 of the present invention is provided below. The specific implementation of each step in this embodiment 1 is described in detail as follows.

[0068] The specific implementation method of step S01 is to process the buoy observation data through a multi-level quality inspection system to ensure data reliability. First, the raw data uploaded in real time by buoy observation stations distributed around the world are received, including parameters such as wind speed, wind direction, air pressure, and sea surface temperature. Subsequently, a physical threshold method is used for preliminary screening, and data with wind speeds greater than 50 meters per second are marked as outliers, because wind speeds exceeding this threshold rarely occur in the natural environment; at the same time, the wind direction data is checked to see if it is within the range of 0 to 360 degrees, and those outside the range are marked as invalid. The box plot method is then used to identify statistically outliers and calculate the quartiles of each parameter. 、 and interquartile range , the interquartile range calculation formula is:

[0069] ;

[0070] Where, The lower quartile of the data set indicates the value at 25% after sorting; The upper quartile of the data set, indicating the value at 75% after sorting.

[0071] Based on the interquartile range, construct the outlier judgment criteria:

[0072] ;

[0073] ;

[0074] Where, is the lower threshold of the data; is the upper threshold of the data; if it exceeds Data points outside the range are marked as potential outliers.

[0075] For pressure parameters, a reasonable rate of change threshold of 3 hPa / h was set; data exceeding this rate of change required further review. Finally, physical consistency was checked, for example, verifying that the relationship between wind speed and significant wave height adhered to fundamental principles of marine meteorology. This process preserved high-quality data for subsequent calibration model construction, improving the reliability of the model input data and ultimately enhancing the accuracy of the final calibration results.

[0076] The specific implementation of step S02 is to achieve accurate spatial matching between the buoy observation data and the reanalysis wind field data through bilinear interpolation. First, the latitude and longitude coordinates of the buoy are determined, and then the four nearest reanalysis wind field data grid points around the coordinates are identified. These grid points form a rectangular area containing the buoy position. For each buoy position , calculate its distance from the four grid points, and determine the weight coefficient based on the inverse ratio of the distance. The specific weight calculation formula is:

[0077] ;

[0078] Where, Indicates the The weight coefficient of each grid point; Indicates that the buoy has reached The Euclidean distance of grid points is calculated as follows: ; For the The coordinates of the grid points; the weight sum is 1, that is .

[0079] Based on the calculated weight coefficient, the wind field data of the four grid points are weighted averaged to obtain the interpolated wind field data at the buoy position:

[0080] ;

[0081] ;

[0082] Where, and are the wind speed and wind direction at the buoy position obtained by interpolation, respectively; and Respectively The wind speed and direction at each grid point.

[0083] This step resolves the inconsistency between the buoy observation points and the reanalysis wind field grid points, achieving a coordinated and unified spatial dimension and laying the foundation for subsequent error analysis. Bilinear interpolation better preserves the spatial continuity of the wind field than simple nearest neighbor interpolation, making it suitable for expressing the spatial characteristics of wind fields in complex sea conditions.

[0084] The specific implementation of step S03 is to dynamically adjust the time matching window width according to the meteorological conditions to achieve accurate alignment of the time dimension. First, by analyzing the current meteorological condition type, a meteorological stability index is constructed. , the calculation formula is:

[0085] ;

[0086] Where, is the meteorological stability index, and its value range is , the larger the value, the faster the meteorological conditions change; is the standard deviation of wind speed within a certain period of time; is the average wind speed during the same time period; is the absolute change in air pressure per unit time; is the reference pressure change, set to 5 hPa; is the standard deviation of wind direction; To refer to the wind direction change, it is set to 45 degrees; 、 、 are weight coefficients, which are set to 0.5, 0.3 and 0.2 respectively, satisfying .

[0087] Based on the meteorological stability index , dynamically determine the time window width :

[0088] ;

[0089] Where, is the half-width of the time window; when When it is determined to be a rapid change condition, the time window is set to minute; when When is the normal condition, the time window is minute; when When the condition is stable, the time window is extended to minute.

[0090] This step effectively solves the problem of difference in temporal resolution between buoy observation data and reanalysis wind field data, improves the time matching accuracy, and provides more reliable basic data for subsequent error field construction.

[0091] The specific implementation of step S04 is to calculate the wind speed error and wind direction error of the time-space matching point. The wind speed error is calculated using the algebraic difference method:

[0092] ;

[0093] Where, is the wind speed error; Observe wind speed for buoys; is the wind speed of the corresponding reanalysis wind field. A positive value indicates that the reanalysis wind field underestimates the actual wind speed, while a negative value indicates an overestimation.

[0094] The calculation of wind direction error needs to take into account the cyclic characteristics of the angle:

[0095] ;

[0096] ;

[0097] Where, is the original wind direction difference; The buoy observes the wind direction, the range is ; To reanalyze the wind direction, the range is ; is the normalized wind direction error, ranging from .

[0098] In order to eliminate the influence of random errors, median filtering is applied to the error values ​​at multiple consecutive time points:

[0099] ;

[0100] ;

[0101] Where, and Time points The wind speed and direction errors after filtering; Represents the median operation.

[0102] This step quantifies the degree of deviation of the reanalyzed wind field relative to the actual observation, providing basic data for subsequent error field construction and wind field correction, while separating the errors of the two independent components of wind speed and wind direction, making it easier to model and process them separately.

[0103] The specific implementation of step S05 is to extract the key spatiotemporal characteristic parameters that affect the wind field error distribution. First, the sea surface temperature field of the study area is extracted from the satellite remote sensing data source. , the spatial resolution is 0.05 degrees, and the temporal resolution is the daily average. Then calculate the pressure gradient , the pressure values ​​of adjacent grid points are processed by the central difference method:

[0104] ;

[0105] ;

[0106] ;

[0107] Where, Indicates location The air pressure value at and Represents the grid spacing in the east-west and north-south directions respectively; and They represent the rate of change of air pressure in the east-west and north-south directions respectively; is the magnitude of the pressure gradient.

[0108] Wind field curvature The calculation uses the Laplace operator and the second-order derivative of the wind speed field is approximated:

[0109] ;

[0110] ;

[0111] ;

[0112] Where, Indicates location The wind speed value at and represent the second-order derivatives of the wind speed field in the east-west and north-south directions respectively; is the wind field curvature, which reflects the spatial variation rate of the wind field.

[0113] Seasonal cycle coefficient Obtained by Fourier transform of historical wind field data:

[0114] ;

[0115] Where, is the seasonal period coefficient; is the annual amplitude, obtained through Fourier analysis, with a typical value of 15% to 25% of the wind speed; is the current date (day of the year); is the initial phase, which is obtained by fitting historical data.

[0116] Diurnal variation index Use the cosine function to fit the intraday variation characteristics:

[0117] ;

[0118] Where, It is an indicator of diurnal variation; It is the daily amplitude, which is generally 5% to 15% of the average wind speed; is the current hour (hour of the day); is the phase parameter, obtained by data fitting.

[0119] These spatiotemporal characteristic parameters comprehensively characterize the key physical factors affecting wind field errors, provide multi-dimensional input features for the Gaussian process regression model, and improve the physical rationality of error field construction and prediction accuracy.

[0120] The specific implementation of step S06 is to optimize the Gaussian process regression using the ocean physics constraint neural network model. The model adopts a multi-layer encoding and decoding architecture, including a surface layer physical process encoding layer, a boundary layer dynamics embedding layer, and a multi-head sparse attention layer. The boundary layer stability judgment function is based on the Richardson number. calculate:

[0121] ;

[0122] Where, is the Richardson number; is the acceleration due to gravity, which is 9.8 meters per second squared; is the potential temperature difference in Kelvin; is the height difference in meters; is the mean potential temperature in Kelvin; is the wind speed difference in meters per second.

[0123] Based on Richardson number , wave height and turbulent kinetic energy Constructing a decision vector :

[0124] ;

[0125] Where, is the determination vector, which consists of five components, corresponding to different stability intervals; each component is calculated as follows:

[0126] ;

[0127] ;

[0128] ;

[0129] ;

[0130] ;

[0131] Where, is the sigmoid function; is the maximum significant wave height, which is 15 meters; is the maximum turbulent kinetic energy, which is 5 square meters per square second. to They represent the judgment scores of strong instability, weak instability, neutrality, weak stability and strong stability conditions, and the value range is .

[0132] According to the decision vector , determines the coupling index of the boundary layer physics equations in the model :

[0133] ;

[0134] Where, is the coupling index, and its value range is ,The larger the value, the higher the physical constraint strength; Represents a vector The maximum value in .

[0135] The number of attention heads in a multi-head sparse attention layer The sea surface temperature gradient Decide:

[0136] ;

[0137] Where, is the number of attention heads; is the absolute value of the sea surface temperature gradient.

[0138] Attention sparsity By wind field curvature and wind speed Jointly decided:

[0139] ;

[0140] Where, is the attention sparsity, and its value range is , the smaller the value, the denser the attention connection; is the maximum wind field curvature, which is per square meter; is the maximum wind speed, which is 50 meters per second; and Represents the maximum and minimum operations respectively.

[0141] This step effectively improves the prediction accuracy of Gaussian process regression by integrating physical laws and data-driven methods, especially in areas with complex sea-air interactions and extreme weather conditions.

[0142] The specific implementation of step S07 is to establish a spatiotemporal covariance model based on Gaussian process regression. First, define the Matern3 / 2 kernel function as the basis of the covariance function:

[0143] ;

[0144] Where, for point with dot The covariance between is the signal variance, estimated from the data; for and the distance between them; is the length scale parameter that controls the rate at which the correlation decays with distance.

[0145] To construct the spatiotemporal covariance structure, different characteristic length parameters are set for the spatial dimension and the temporal dimension respectively:

[0146] ;

[0147] Where, is the spatial distance; is the time distance; is the spatial scale parameter, which is set to 50 kilometers; is the time scale parameter, set to 6 hours.

[0148] Given an observation dataset ,in is the input feature vector, For the observed error value, the prediction distribution of Gaussian process regression is:

[0149] ;

[0150] Where, Forecast point The error value at ; Indicates the mean , the variance is Normal distribution; and are the mean and variance of the predicted distribution, respectively, and the calculation formula is:

[0151] ;

[0152] ;

[0153] Where, Forecast point With all observation points The covariance vector between ; is the covariance matrix between observation points, the elements ; is the observation noise variance; is the identity matrix; is the covariance of the prediction point itself; is the observation error vector .

[0154] For large-scale data sets, sparse approximation methods are used to reduce computational complexity:

[0155] ;

[0156] ;

[0157] Where, For the selected Induction points; is the covariance vector between the prediction point and the induced point; is the covariance matrix between the induced points; is the correction matrix associated with the observed data.

[0158] Finally, the spatiotemporal covariance model outputs the wind speed error field and wind direction error field , both error fields are defined continuously in the ocean space and time Functions on dimensions:

[0159] ;

[0160] ;

[0161] Where, and Position ,time The predicted mean of wind speed and direction errors at .

[0162] This step makes full use of the spatial correlation and temporal continuity of observation errors, can provide reasonable error interpolation in areas with sparse observation points, and realizes the transformation from discrete observation points to a continuous error field.

[0163] The specific implementation of step S08 is to use the error field constructed in step S07 to correct the reanalysis wind field data. The correction process first determines the time and space coordinates of the point to be corrected , and then query the wind speed error field value and wind direction error field value of the corresponding position and time from the error field. Wind speed correction uses direct addition:

[0164] ;

[0165] Where, is the corrected wind speed; Wind speed for reanalysis wind field; is the wind speed error field value.

[0166] To ensure the physical rationality of the correction results, the upper limit of wind speed correction is set to 30% of the original wind speed:

[0167] ;

[0168] Where, for The sign function takes the value of +1 or -1.

[0169] Wind direction correction takes into account the cyclic characteristics of the angle:

[0170] ;

[0171] Where, is the corrected wind direction; To reanalyze the wind field and direction; is the wind direction error field value; Indicates a modulo operation, ensuring that the result is within the range.

[0172] Limit the maximum wind direction correction angle to 45 degrees:

[0173] ;

[0174] ;

[0175] Where, is the wind direction error field value after limitation.

[0176] In areas with strong wind speed gradients such as typhoons, special processing strategies are applied to correct the intensity of the wind speed gradient. Increase and decrease:

[0177] ;

[0178] ;

[0179] ;

[0180] Where, is the correction coefficient, and its value range is , the greater the wind speed gradient, the smaller the correction coefficient; is an adjustment parameter, and its value is 0.1; is the norm of the wind speed gradient.

[0181] Through this correction process, the systematic bias in the reanalysis wind field data was systematically eliminated, and the accuracy of the wind field data was improved. The correction effect was particularly significant in areas with complex nearshore terrain and special weather systems.

[0182] The specific implementation of step S09 is to generate an uncertainty assessment result for the corrected wind field data based on the variance output of Gaussian process regression. In addition to the prediction error field mean, Gaussian process regression also outputs the prediction variance, which directly quantifies the uncertainty of the prediction. The system uses this property to calculate a 95% confidence interval for each correction point:

[0183] ;

[0184] Where, For location ,time 95% confidence interval at ; is the predicted error value, that is or ; To predict the standard deviation, it is directly output from the Gaussian process regression model; the coefficient of 1.96 comes from the 95% confidence interval of the standard normal distribution.

[0185] To quantify the overall uncertainty level, the regional average uncertainty index was calculated ;

[0186] Where, is the uncertainty index, the smaller the value, the more reliable the prediction result; is the number of grid points in the evaluation area; For location ,time The standard deviation of the prediction at ; is the absolute value of the prediction error.

[0187] The system automatically marks areas where the uncertainty exceeds the threshold, and users are advised to use the correction results in these areas with caution:

[0188] ;

[0189] Where, is a mask function, a value of 1 indicates that the correction result at that position has high uncertainty; is the uncertainty threshold, set to 0.5.

[0190] This step provides users with a reliability assessment of the correction results, which helps in risk control and decision support in practical applications of wind farm data, especially in areas such as marine engineering safety assessment and navigation decision-making.

[0191] The specific implementation of step S10 is to use the leave-one-buoy cross-validation method to evaluate the generalization ability of the ocean physical constraint neural network model. This method excludes all data of one buoy from the training data set as the test set each time, uses the remaining buoy data to train the model and predict the wind field error at the excluded buoy location. In the specific implementation, first construct a buoy index list , including all those involved in the verification buoy identification. For each buoy , perform the following verification steps:

[0192] From the training dataset Exclusion buoy All data of and the test subset :

[0193] ;

[0194] ;

[0195] Where, Representation and data points The associated buoy ID.

[0196] Using the training subset Train the ocean physics constraint neural network model to obtain model parameters . For the test subset Each data point in , use the trained model to predict the error value , and the actual observed error value Compare.

[0197] Calculate three evaluation indicators: root mean square error ,deviation and correlation coefficient :

[0198] ;

[0199] ;

[0200] ;

[0201] Where, Represents a test subset The number of samples; and are the means of the observed and predicted values ​​in the test subset, respectively.

[0202] The loop continues until each buoy has been used as a test set once, and then the global evaluation metrics are calculated:

[0203] ;

[0204] ;

[0205] ;

[0206] Where, 、 and They are the global root mean square error, bias and correlation coefficient, respectively, which are used to comprehensively evaluate the generalization performance of the model.

[0207] In order to judge whether the correction result is significantly better than the original reanalysis wind field data, the improvement rate index is introduced :

[0208] ;

[0209] Where, is the root mean square error between the reanalysis wind field and the buoy observations before correction; is the root mean square error between the corrected wind field and the buoy observation; is the improvement rate in percentage. The larger the value, the more significant the correction effect.

[0210] This verification method simulates the application scenario of the model in an unknown location, and can objectively evaluate the actual effect and reliability of the correction method, providing a scientific basis for model optimization and practical application. In particular, for different sea and meteorological conditions, the condition evaluation indicators are further calculated:

[0211] ;

[0212] Where, For specific conditions The root mean square error under To meet the conditions A subset of the test data. This could be wind speed range, sea state type, stability category, etc.

[0213] This step comprehensively evaluated the prediction ability of the ocean physical constraint neural network model at unknown locations through the leave-one-buoy cross-validation method, verified the effectiveness and reliability of the correction method, and provided a scientific basis for subsequent model optimization and practical application.

[0214] The detailed structure of the ocean physics constraint neural network model is a multi-layer encoder-decoder architecture, consisting of three main parts: a surface layer physics process encoding layer, a boundary layer dynamics embedding layer, and a multi-head sparse attention layer. The surface layer physics process encoding layer consists of three fully connected layers with the following hierarchical structure:

[0215] ;

[0216] ;

[0217] ;

[0218] Where, is the input feature vector, which contains the Richardson number and roughness parameter of the air-sea interface; 、 、 is the weight matrix; 、 、 is the bias vector; is the Gaussian error linear unit activation function; It is a batch normalization operation; 、 、 It represents the middle hidden layer; the last layer adds residual connection to improve training stability.

[0219] The boundary layer dynamics embedding layer contains five sets of parameterized equations, each corresponding to a state of atmospheric stability, graded from strongly unstable to strongly stable. Each set of parameterized equations is implemented by a three-layer fully connected network:

[0220] ;

[0221] Where, Indicates the A parameterized function under stability conditions, ; is the height; is the wind speed; is the temperature gradient; Indicates the Under the stability condition Layer fully connected layer; is the rectified linear unit activation function; Represents the concatenation of input features.

[0222] The multi-head sparse attention layer consists of multiple attention heads, each of which is responsible for capturing feature correlations at different spatial scales:

[0223] ;

[0224] ;

[0225] ;

[0226] ;

[0227] Where, Indicates the The output of an attention head; Calculate the attention function; 、 、 are query, key, and value matrices respectively; is the input feature; 、 、 is the corresponding weight matrix; is the attention sparsity.

[0228] The attention calculation function is defined as:

[0229] ;

[0230] Where, is the softmax normalization function; is the dimension of the key vector; is the sparse function, according to the sparsity Keep the largest part of the attention weights and set the rest to zero.

[0231] Finally, the output of the multi-head attention is the concatenation of the outputs of each head, and then undergoes a linear transformation:

[0232] ;

[0233] Where, is the output of multi-head attention; Represents a splicing operation; is the number of attention heads; is the weight matrix of the output linear transformation.

[0234] The model output layer is fused with the Gaussian process regression results through residual connection:

[0235] ;

[0236] Where, is the final predicted error value; is the predicted output of the neural network model; is the prediction output of Gaussian process regression; It is a coupling index determined by the boundary layer stability judgment function and is used to balance the weights of physical laws and data-driven factors.

[0237] The detailed steps for establishing a training dataset for the ocean physics-constrained neural network model include: First, a decade of high-quality data was selected from the global buoy observation network as the base dataset. Observation records from 387 buoys from 2010 to 2020 were selected, totaling over 78 million samples. The dataset was stratified by season and ocean area to ensure uniform spatial distribution of training samples, with each 5-degree x 5-degree grid containing at least 1,000 training samples. For each training sample, a 128-dimensional feature vector was constructed, containing wind speed and direction differences and related physical field parameters, including nearshore surface layer structure parameters and turbulence intensity indicators. Error corrections for annotated samples were calculated by comparing historical buoy observations with reanalysis data at the corresponding locations. The training dataset spanned a number of typical meteorological events, including typhoons, seasonal transitions, and extreme weather events, to ensure model generalization. In sparsely populated areas, data augmentation was performed using physical model simulation results to generate synthetic training samples to compensate for insufficient observational data. A total of approximately 500,000 synthetic samples were generated, primarily in high-latitude regions of the Southern Hemisphere and in the open ocean. The training data is divided into a training set, a validation set, and a test set at a ratio of 80%, 10%, and 10%, respectively, for model parameter optimization, hyperparameter adjustment, and final performance evaluation. In particular, when dividing the dataset, we ensure that data from the same buoy appears in only one dataset to avoid overestimation of performance due to data leakage.

[0238] In order to better understand and implement the present invention, Example 2 of a specific application scenario of the present invention is provided below: In a study on accurate forecasting of ocean wind fields conducted in a certain sea area, researchers used the method of the present invention to correct the ERA5 reanalysis wind field data of the European Centre for Medium-Range Weather Forecasts (ECMWF) in order to improve the accuracy of wind field forecasts in offshore areas. The study area covers the range of 18 to 22 degrees north latitude and 112 to 116 degrees east longitude, and includes 10 meteorological ocean buoy observation stations. The study used buoy observation data and ERA5 reanalysis wind field data from January 2021 to December 2021 as the basic data sets. First, the buoy observation data was screened for outliers, and a total of 283 wind speed outliers and 169 wind direction outliers were identified by the box plot method, accounting for 2.6% of the total sample size. The statistical characteristics of the observation data of each buoy station after outlier screening are shown in Table 1.

[0239] Table 1 Statistical characteristics of wind field data at buoy observation sites

[0240]

[0241] The buoy observation data were then temporally and spatially matched with the ERA5 reanalysis wind field data. The ERA5 data has a spatial resolution of 0.25 degrees and a temporal resolution of 1 hour. Wind field data corresponding to the buoy locations were extracted using bilinear interpolation. Considering the study area's frequent typhoon activity in the summer and the relatively stable influence of the northeast monsoon in the winter, a dynamic time window was used for temporal matching. The window width was set to ±5 minutes during typhoon season (July-September), ±60 minutes during winter (December-February), and ±15 minutes for the rest of the year. After temporal and spatial matching, the wind speed and direction errors at each buoy site were calculated. The results show that the ERA5 reanalysis wind field generally underestimates the actual wind speed, with an average underestimation of 12.8%. Wind direction errors exhibit distinct land-sea distribution patterns, with significantly greater wind direction errors nearshore than offshore.

[0242] When extracting the spatiotemporal characteristic parameters that affect the wind field error distribution, the sea surface temperature data obtained by satellite remote sensing showed that there was a clear temperature front in the study area, with the maximum sea surface temperature gradient reaching 0.08℃ / km. The pressure gradient was calculated using the ERA5 pressure field, and the maximum pressure gradient during the typhoon reached 5.7 hPa / km. The wind field curvature calculated using the Laplace operator showed that the wind field curvature in the coastal area was larger, with a maximum value of Per square meter. The seasonal coefficients extracted through Fourier analysis indicate that the annual amplitude of the wind field in the study area is 21.3% of the average wind speed, with the maximum wind speed occurring in winter. Diurnal variation indicators show that daytime wind speeds are generally higher than nighttime wind speeds, with a daily amplitude of 8.7% of the average wind speed.

[0243] The Richardson number distribution under different sea-air conditions is determined by the boundary layer stability judgment function, and the judgment vector is constructed accordingly. The results are shown in Table 2.

[0244] Table 2 Distribution of boundary layer stability determination vectors under different conditions

[0245]

[0246] When constructing a spatiotemporal covariance model based on Gaussian process regression, the optimal Matern3 / 2 kernel function parameters are determined by maximum likelihood estimation, and the signal variance is 2.64, the spatial scale parameter Set to 52.3 km, the time scale parameter The covariance model generates continuous wind speed and direction error fields, which are then used to correct the ERA5 reanalysis wind data and calculate 95% confidence intervals. The correction results demonstrate the improved performance under various sea conditions, as assessed using the leave-one-buoy cross-validation method, as shown in Table 3.

[0247] Table 3 Wind field correction improvement effect under different sea conditions

[0248]

[0249] As shown in Table 3, the corrected wind data achieved significant improvements across all sea conditions, with the overall root mean square error reduced from 1.92 m / s to 0.86 m / s, an improvement rate of 55.2%. In the nearshore region, the improvement reached a staggering 63.7%, primarily due to the ocean physics-constrained neural network model's ability to capture wind field deformations caused by complex terrain. Wind direction error also significantly improved, with the average wind direction error decreasing from 18.3 degrees to 7.6 degrees.

[0250] Traditional offshore wind farm data correction methods mainly rely on statistical regression or simple physical models, which makes it difficult to simultaneously consider complex spatiotemporal variation characteristics and physical processes of the sea-air interface. For example, linear regression and nearest neighbor correction methods can only capture simple linear or local relationships, and perform poorly under complex sea conditions and extreme weather conditions. The improvement rate in nearshore areas is usually less than 20%. Although the numerical model nesting method takes physical processes into account, it has high computational costs and poor real-time performance. The present invention combines a hybrid method of marine physical constraint neural network model and Gaussian process regression, which not only maintains the flexibility and efficiency of the data-driven method, but also incorporates the constraints of the physical laws of the marine boundary layer. In particular, the spatiotemporal matching problem under different meteorological conditions is solved through dynamic time windows and boundary layer stability judgment functions, significantly improving the wind field correction accuracy in extreme weather conditions such as typhoons and nearshore complex terrain areas. At the same time, the uncertainty assessment provided by Gaussian process regression provides a quantitative basis for the reliability of wind farm data, providing more reliable wind farm data support for offshore wind farm site selection, marine engineering safety assessment and navigation decision-making.

[0251] It should be noted that the variables involved in the present invention are explained in detail as shown in Tables 4, 5 and 6 below.

[0252] Table 4 Variable Explanation Table (Part I)

[0253]

[0254] Table 5 Variable Explanation Table (Part II)

[0255]

[0256] Table 6 Variable Explanation Table (Part 3)

[0257]

[0258] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be covered by the scope of protection of the present invention.

Claims

1. A method for correcting offshore reanalysis wind field data based on buoy observation data, characterized in that: include: Receive buoy observation data and reanalysis wind field data, remove outliers and perform quality control on the buoy observation data; Performing spatiotemporal matching on the buoy observation data and the reanalysis wind field data, extracting the wind field data of the reanalysis wind field data grid points corresponding to the buoy position using a bilinear interpolation method to obtain spatiotemporal matching data; constructing a dynamic time window based on the spatiotemporal matching data, and automatically adjusting the width of the dynamic time window to 15 minutes to 1 hour based on the type of meteorological event, thereby achieving precise alignment of the time dimension; The wind speed error and wind direction error of each matching point are calculated based on the spatiotemporal matching data, where the wind speed error is the wind speed of the buoy observation data minus the wind speed of the reanalysis wind field data, and the wind direction error is the wind direction of the buoy observation data minus the wind direction of the reanalysis wind field data; the spatiotemporal characteristic parameters affecting the error distribution are extracted based on the wind speed error and the wind direction error, where the spatiotemporal characteristic parameters include sea surface temperature, pressure gradient, wind field curvature, seasonal cycle coefficient, and diurnal variation index; the Gaussian process regression is optimized using an ocean physics constraint neural network model, which uses a sparse attention mechanism to process multi-scale spatiotemporal features, and the coupling index of the ocean physics constraint neural network model is determined by a boundary layer stability judgment function; a spatiotemporal covariance model is established based on the Gaussian process regression, and an error field is constructed using a Matern 3 / 2 kernel function; and the reanalysis wind field data is corrected using the error field.

2. The method for correcting marine reanalysis wind field data based on buoy observation data according to claim 1, characterized in that: The outlier elimination and quality control processing specifically involves identifying outliers in wind speed, wind direction, and air pressure parameters in buoy observation data by setting physical thresholds and combining them with a boxplot method. Data points that exceed a reasonable range are marked as outliers and do not participate in subsequent calculations.

3. The method for correcting marine reanalysis wind field data based on buoy observation data according to claim 2, characterized in that: The bilinear interpolation method specifically achieves spatial resolution harmony by calculating the weighted average of the four nearest reanalysis wind field data grid points around the buoy position, where the weight is inversely proportional to the distance from the buoy to the reanalysis wind field data grid point.

4. The method for correcting offshore reanalysis wind field data based on buoy observation data according to claim 3, characterized in that: The dynamic time window is specifically a time period that is adaptively adjusted based on the type of meteorological event. Under normal weather conditions, a matching window of plus or minus 15 minutes is adopted, which is shortened to plus or minus 5 minutes in a rapidly changing typhoon weather environment, and extended to plus or minus 1 hour under stable weather conditions.

5. The method for correcting offshore reanalysis wind field data based on buoy observation data according to claim 4, characterized in that: The structure of the ocean physical constraint neural network model is a multi-layer encoding and decoding architecture, which includes a surface layer physical process encoding layer, a boundary layer dynamics embedding layer and a multi-head sparse attention layer.

6. The method for correcting offshore reanalysis wind field data based on buoy observation data according to claim 5, characterized in that: The surface layer physical process encoding layer integrates the Richardson number and roughness parameter of the air-sea interface to convert the physical field characteristics into potential representations. The boundary layer dynamics embedding layer contains five sets of parameterized equations to describe the wind profile characteristics and turbulence effects under different stability conditions. The multi-head sparse attention layer consists of eight attention heads, each of which focuses on feature correlations at different spatial scales. The spatiotemporal characteristic parameters are specifically key physical quantities and time characteristics that affect the wind field error distribution. The sea surface temperature is obtained through satellite remote sensing, the pressure gradient is calculated by the pressure difference between adjacent grid points, the wind field curvature reflects the spatial change rate of the wind field, the seasonal cycle coefficient is extracted through Fourier transform to extract the annual variation characteristics, and the diurnal variation index is fitted with the intraday variation characteristics by sine function. The wind speed of the corrected wind field data is equal to the wind speed of the reanalyzed wind field data plus the wind speed error field value at the corresponding position and time; This also includes generating uncertainty assessment results of corrected wind field data based on Gaussian process regression variance output, and calculating and outputting 95% confidence intervals.

Images