Marine reanalysis wind field data correction method based on buoy observation data

Abstract

The invention provides a marine reanalysis wind field data correction method based on buoy observation data, and belongs to the technical field of meteorologies.The marine reanalysis wind field data correction method comprises the steps that firstly, abnormal value elimination and quality control are conducted on the buoy observation data, and space-time matching is conducted through a bilinear interpolation method; constructing a dynamic time window, and adaptively adjusting the time precision according to the meteorological event type; calculating wind speed and wind direction errors, and extracting spatio-temporal characteristic parameters; the method comprises the following core steps of: optimizing Gaussian process regression by using an ocean physical constraint neural network model, wherein the model comprises a surface layer physical process coding layer, a boundary layer dynamic embedding layer and a multi-head sparse attention layer; constructing a space-time covariance model to generate an error field; correcting the reanalyzed wind field data and outputting uncertainty evaluation; the technical problem that a significant space-time error exists between reanalysis wind field data and an actual observation value in a complex marine environment is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of meteorology. Specifically, it relates to a method for correcting reanalysis wind field data at sea based on buoy observation data. Background Art

[0002] Wind field data at sea is important basic information for ocean engineering construction, navigation safety assurance, ocean energy development, and ocean weather forecasting. Its accuracy and reliability directly affect the effects and safety of related applications. Currently, there are mainly two ways to obtain wind field data at sea: in-situ buoy observations and meteorological reanalysis data. Buoy observations can directly obtain sea surface wind field information, with high temporal resolution and measurement accuracy. However, due to factors such as high deployment costs and limited coverage, it is impossible to achieve comprehensive monitoring of the vast sea area; while reanalysis wind field data is generated by assimilating multi-source observation data through numerical models, with the advantage of global continuous coverage, and has become an important data source in ocean applications.

[0003] However, due to the complexity of the ocean environment and the limitations of the numerical model itself, there are often systematic biases in reanalysis wind field data in practical applications. Especially in cases such as nearshore complex terrain, strong air-sea interactions, and extreme meteorological events, there are significant spatio-temporal error distributions between reanalysis wind field data and in-situ buoy measurements. Research shows that these errors exhibit obvious non-uniformity and non-linearity characteristics in different sea areas, different seasons, and even different time periods within a day, seriously affecting the reliability of reanalysis wind field data in high-precision application scenarios.

[0004] Currently, the correction of reanalysis wind field data mainly uses statistical interpolation or simple regression methods, such as techniques like linear regression, polynomial fitting, and optimal interpolation. These methods usually establish an empirical relationship between in-situ buoy observations and reanalysis wind field data based on historical statistical characteristics within a fixed time window. However, this simple statistical correlation is difficult to effectively capture the spatio-temporal variability characteristics of wind field errors in complex ocean environments. Especially under rapidly changing meteorological conditions such as typhoons and strong convections, the error distribution between reanalysis wind field data and actual observations shows significant non-stationarity and physical process dependence, and the effect of traditional correction methods is very limited. That is to say, there is a technical problem in the prior art that there are significant spatio-temporal 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 reanalysis wind field data at sea based on buoy observation data, which can solve the technical problem in the prior art that there are significant spatio-temporal errors between reanalysis wind field data and actual observations in complex ocean environments.

[0006] The present invention is implemented as follows: The present invention provides a method for correcting reanalysis wind field data based on buoy observation data, which includes: receiving buoy observation data and reanalysis wind field data, and performing outlier removal and quality control processing on the buoy observation data; performing spatio-temporal 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 spatio-temporal characteristic parameters; optimizing Gaussian process regression using an ocean physical constraint neural network model, where the ocean physical constraint neural network model processes multi-scale spatio-temporal characteristics using a sparse attention mechanism, and the coupling index of the ocean physical constraint neural network model is determined by a boundary layer stability determination function; establishing a spatio-temporal covariance model; and correcting the reanalysis wind field data using the error field.

[0007] Among them, the outlier removal and quality control processing specifically identifies outliers of parameters such as wind speed, wind direction, and air pressure in the buoy observation data by setting physical thresholds in combination with the box plot method, and data points outside the reasonable range are marked as outliers and do not participate in subsequent calculations.

[0008] Among them, the spatio-temporal matching specifically uses the bilinear interpolation method to extract the wind field data of the grid points corresponding to the buoy position in the reanalysis wind field data to obtain spatio-temporal matching data.

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

[0010] Among them, the dynamic time window specifically automatically adjusts the width of the dynamic time window to 15 minutes to 1 hour according to the type of meteorological event to achieve precise alignment in 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; for the wind direction error, specifically, the angular difference between the wind direction of the buoy observation data and the wind direction of the reanalysis wind field data is calculated, and considering the angular cycle characteristic, the minimum angular difference with an absolute value not exceeding 180 degrees is taken.

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

[0013] Among them, the spatio-temporal 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: the surface layer physical process encoding layer, the boundary layer dynamics embedding layer, and the multi-head sparse attention layer.

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

[0016] Among them, the spatio-temporal feature parameters are specifically the key physical quantities and time features that affect the wind field error distribution. The sea surface temperature is obtained by satellite remote sensing. The pressure gradient is calculated by the air pressure difference between adjacent grid points. The wind field curvature reflects the spatial change rate of the wind field. The seasonal cycle coefficient extracts the annual change features through Fourier transform. The diurnal change index fits the intraday change features through a sine function.

[0017] Among them, the spatio-temporal covariance model uses the Matern 3 / 2 kernel function to construct the error field. The spatial scale of the spatio-temporal covariance model is set to 50 kilometers, and the time scale of the spatio-temporal covariance model is set to 6 hours.

[0018] Among them, the wind speed of the corrected wind field data is equal to the wind speed of the reanalysis wind field data plus the wind speed error field value at the corresponding position and time.

[0019] Among them, it also includes generating an uncertainty assessment result of the corrected wind field data based on the output of the Gaussian process regression variance, and calculating and outputting the 95% confidence interval.

[0020] The present invention realizes the precise correction of reanalysis wind field data by constructing a dynamic time window and an ocean physics constrained 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, ensuring the accuracy of the corrected basic data from the time dimension. The present invention extracts spatio-temporal characteristic parameters such as sea surface temperature, pressure gradient, wind field curvature, seasonal cycle coefficient, and diurnal variation index, and establishes an association mechanism between the wind field error distribution and the physical characteristics of the ocean environment. In particular, the introduced ocean physics constrained neural network model effectively captures the non-linear error distribution characteristics in complex ocean environments through a multi-level structure of the surface layer physical process encoding layer, boundary layer dynamics embedding layer, and multi-head sparse attention layer, realizing the accurate prediction of wind field errors. By combining Gaussian process regression with the physics constrained neural network, the present invention not only improves the accuracy of the corrected wind field data but also provides reliable uncertainty assessment, providing a basis for risk control in the application of wind field data. Experimental verification shows that this method can significantly reduce the systematic bias of reanalysis wind field data under various ocean environmental conditions, especially during extreme meteorological events, effectively solving the technical problem of significant spatio-temporal errors between reanalysis wind field data and actual observations in complex ocean environments. Brief Description of the Drawings

[0021] Figure 1 It is a flowchart of the method of the present invention. Detailed Embodiments

[0022] To make the objectives, 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] As Figure 1 shown, it is a flowchart of a method for correcting reanalysis wind field data at sea based on buoy observation data provided by the present invention. This method includes the following steps: S01. Receive buoy observation data and reanalysis wind field data, perform outlier rejection and quality control processing on the buoy observation data, and mark the buoy observation data with a wind speed greater than 50 meters per second or not conforming to physical laws as invalid data; S02. Perform spatio-temporal matching on the buoy observation data and the reanalysis wind field data, and use the bilinear interpolation method to extract the wind field data of the grid points corresponding to the buoy position in the reanalysis wind field data to obtain spatio-temporal matching data; S03. Construct a dynamic time window based on the spatio-temporal matching data, and automatically adjust the width of the dynamic time window to 15 minutes to 1 hour according to the meteorological event type to achieve precise alignment in the time dimension; S04. Calculate the wind speed error and wind direction error for each matching point based on the spatio-temporal matching data. 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; S05. Extract spatio-temporal characteristic parameters affecting the error distribution based on the wind speed error and the wind direction error. The spatio-temporal characteristic parameters include sea surface temperature, pressure gradient, wind field curvature, seasonal cycle coefficient, and diurnal variation index; S06. Optimize the Gaussian process regression using a pre-trained ocean physical constraint neural network model. The ocean physical constraint neural network model processes multi-scale spatio-temporal features using a sparse attention mechanism, and the coupling index of the ocean physical constraint neural network model is determined by a boundary layer stability determination function; S07. Establish a spatio-temporal covariance model based on the Gaussian process regression, and construct an error field using the Matern3 / 2 kernel function. The spatial scale of the spatio-temporal covariance model is set to 50 kilometers, and the time scale of the spatio-temporal covariance model is set to 6 hours; S08. Correct the reanalysis wind field data using the error field to obtain corrected wind field data. The wind speed of the corrected wind field data is equal to the wind speed of the reanalysis wind field data plus the wind speed error field value at the corresponding position and time; S09. Generate an uncertainty assessment result for the corrected wind field data based on the Gaussian process regression variance output, and calculate and output the 95% confidence interval; 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, bias, and correlation coefficient between the corrected wind field data and the buoy observation data, and verify the correction effect.

[0024] Among them, outlier removal specifically identifies outliers of parameters such as wind speed, wind direction, and pressure in the buoy observation data by setting physical thresholds and combining box plot methods. Data points outside the reasonable range are marked as outliers and do not participate in subsequent calculations.

[0025] Among them, the bilinear interpolation method specifically realizes the reconciliation of spatial resolution by calculating 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.

[0026] Among them, the dynamic time window is specifically a time period adaptively adjusted based on the type of meteorological event. A plus-minus 15-minute matching window is adopted under normal weather conditions, shortened to plus-minus 5 minutes in the rapidly changing meteorological environment of typhoons, and extended to plus-minus 1 hour under stable meteorological conditions.

[0027] Among them, 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 reanalysis wind field data. Considering the angular cycle characteristics, the minimum angular difference with an absolute value not exceeding 180 degrees is taken.

[0028] Among them, the spatio-temporal characteristic parameters are specifically the key physical quantities and time characteristics affecting the wind field error distribution. The sea surface temperature is obtained through satellite remote sensing. The pressure gradient is calculated by the air pressure difference between adjacent grid points. The wind field curvature reflects the spatial change rate of the wind field. The seasonal cycle coefficient extracts the annual change characteristics through Fourier transform. The diurnal change index fits the intra-day change characteristics through a sine function.

[0029] Among them, the specific structure of the ocean physics-constrained neural network model is a multi-layer encoding and decoding architecture, which includes three main parts: the surface layer physical process encoding layer, the boundary layer dynamics embedding layer, and the multi-head sparse attention layer. The surface layer physical process encoding layer fuses the Richardson number of the air-sea interface and the roughness parameter to convert the physical field characteristics into a latent representation. The boundary layer dynamics embedding layer contains five groups 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, and each head focuses on the feature correlations at different spatial scales. The number of heads in the multi-head sparse attention layer is determined by the gradient of the sea surface temperature. The attention sparsity of the multi-head sparse attention layer is determined by the wind field curvature and intensity. The output layer of the ocean physics-constrained neural network model is fused with the Gaussian process regression result through a residual connection method to achieve the dual constraints of physical laws and data-driven.

[0030] Among them, the steps for establishing the training dataset of the ocean physics-constrained neural network model specifically include screening out ten-year high-quality data from the global buoy observation network as the basic dataset, stratifying and sampling the basic dataset according to seasons and sea areas to ensure uniform spatial distribution of training samples, constructing a feature vector including the wind speed and direction differences and related physical field parameters for each training sample. The dimension of the feature vector is 128-dimensional, including the near-surface layer structure parameters and the turbulence intensity index. The error correction value of the labeled sample is calculated by comparing the historical buoy observation records with the reanalysis data at the corresponding location. The time span of the training dataset covers multiple typical meteorological events, including typhoon seasons and alternating extreme weather processes, to ensure the generalization ability of the ocean physics-constrained neural network model. For sparse regions, physical model simulation results are used for data augmentation to generate synthetic training samples to make up for the lack of observation data.

[0031] Among them, the steps of training the ocean physics-constrained neural network model specifically include first using unsupervised pre-training to learn the intrinsic representation of the sea-air interface physical structure. The loss function in the pre-training stage includes a physical consistency constraint term to ensure that the output of the ocean physics-constrained neural network model does not violate the basic ocean dynamics laws. Subsequently, supervised learning is used to fine-tune the parameters of the ocean physics-constrained neural network model, where the loss function is designed as a weighted combination of mean squared error and physical constraints to balance data fitting and physical rationality. The dynamic learning rate adjustment strategy is applied in the training process, and different weight update rules are adopted under different stability conditions. The batch training method is used, with each batch containing 64 samples, and the proportion of extreme weather samples is fixed at 25%. After the ocean physics-constrained neural network model converges after 500 training iterations, post-training fine-tuning is performed. Finally, the ocean physics-constrained neural network model is selected based on dual indicators of validation set performance and physical consistency.

[0032] Among them, Gaussian process regression is specifically a probabilistic machine learning model that describes the similarity between data points by defining a kernel function, establishes a function distribution in a continuous space, and can give both predicted values and uncertainty estimates.

[0033] Among them, the boundary layer stability determination function is specifically a Richardson number function calculated based on sea-air temperature difference, wind speed gradient, and pressure field. By analyzing the atmospheric stability state, the coupling strength of physical equations is determined. The boundary layer stability determination function first calculates the Richardson number at the sea-air interface, and then constructs a determination vector in combination with wave height and turbulent kinetic energy parameters. The determination vector consists of five components corresponding to different stability intervals. According to the value of the determination vector, the coupling index of the boundary layer physical equation in the ocean physics-constrained neural network model is dynamically adjusted. Under strongly unstable conditions, the weight of the turbulent closure equation is increased, and the influence of the thermodynamic equation is enhanced under stable stratification.

[0034] Among them, the Matern3 / 2 kernel function is specifically 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.

[0035] Among them, 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 spatio-temporal covariance model, which is used to correct the reanalysis wind field data.

[0036] Among them, the wind speed error field value specifically refers to the wind speed error predicted value given by the error field at a certain time and position point, which is added as a correction amount to the wind speed of the reanalysis wind field data.

[0037] Among them, the leave-one-buoy cross-validation method specifically excludes all the data of one buoy from the training dataset as the test set each time, uses the data of the remaining buoys to train the ocean physics-constrained neural network model and predicts the wind field error at the position of the excluded buoy, and repeats the process until each buoy has been used as the test set once, and finally comprehensively evaluates the prediction ability of the ocean physics-constrained neural network model at unknown positions.

[0038] The following describes the specific implementation manners of the above steps in detail.

[0039] The specific implementation manner of step S01 is to process the buoy observation data through a multi-level quality inspection system to ensure data reliability. First, receive the raw data uploaded in real time by buoy observation stations distributed globally, including parameters such as wind speed, wind direction, air pressure, and sea surface temperature. Subsequently, perform a preliminary screening using the physical threshold method, mark the data with a wind speed greater than 50 meters per second as outliers because wind speeds exceeding this threshold rarely occur in the natural environment; at the same time, check whether the wind direction data is within the range of 0 to 360 degrees, and mark those outside the range as invalid. Then apply the box plot method to identify outliers in the statistical sense, calculate the quartiles Q1 and Q3 of each parameter, and mark the data points outside the range of Q1 - 1.5×IQR or Q3 + 1.5×IQR as potential outliers, where IQR is the interquartile range. For the air pressure parameter, set a reasonable change rate threshold of 3 hPa per hour, and data exceeding this change rate requires further review. Finally, check the physical consistency, for example, verify whether the relationships between wind speed and significant wave height, wind direction and wave direction conform to the basic principles of marine meteorology. Through the above processing, high-quality data is retained for subsequent correction model construction, the reliability of the model input data is improved, and thus the accuracy of the final correction result is enhanced.

[0040] The specific implementation manner of step S02 is to achieve an accurate spatial matching between the buoy observation data and the reanalysis wind field data through the bilinear interpolation method. First, determine the longitude and latitude coordinates of the buoy, and then identify the four nearest reanalysis wind field data grid points around this coordinate. These grid points form a rectangular area containing the buoy position. For each buoy position, calculate its distances to the four grid points, and determine the weight coefficients based on the inverse of the distances. The specific weight calculation formula is: , where represents the distance from the buoy to the i-th grid point, and the sum of the weights is 1. Through these weights, the wind field data of the four grid points are weighted and averaged to obtain the interpolated wind field data at the buoy position. This step solves the problem of inconsistent positions between the buoy observation points and the reanalysis wind field grid points, realizes the coordination and unity in the spatial dimension, lays a foundation for subsequent error analysis, and ensures that the wind field information at the same position is compared. Bilinear interpolation can better retain the spatial continuity of the wind field compared to the simple nearest neighbor interpolation method and is suitable for expressing the spatial characteristics of the wind field under complex sea conditions.

[0041] The specific implementation of step S03 is to dynamically adjust the width of the time matching window according to meteorological conditions to achieve precise alignment in the time dimension. First, by analyzing the current type of meteorological conditions, they are classified into three categories: conventional weather conditions, extreme change meteorological conditions, and stable meteorological environments. For conventional weather conditions, the system sets a time window of ±15 minutes, that is, the reanalysis wind field data within 15 minutes before and after the buoy observation time are considered valid matches. For extreme change meteorological environments such as typhoons, due to the rapid change of the wind field, the system automatically reduces the time window to ±5 minutes to reduce the error caused by time differences. For stable meteorological conditions, such as the area controlled by high pressure, the system extends the time window to ±1 hour because the wind field changes slowly at this time. The system calculates the meteorological stability index by analyzing three indicators: wind speed standard deviation, pressure change rate, and wind direction stability, and dynamically adjusts the time window according to the index value: when the index value is lower than 0.2, it is determined as a stable condition; between 0.2 and 0.8 is a conventional condition; higher than 0.8 is a rapidly changing condition. This step effectively solves the problem of the difference in time resolution between buoy observation data and reanalysis wind field data, improves the time matching accuracy, and provides more reliable basic data for the subsequent construction of the error field.

[0042] The specific implementation of step S04 is to calculate the wind speed error and wind direction error of the spatio-temporal matching points. The wind speed error is calculated using a simple algebraic difference method, that is, the buoy observed wind speed value minus 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 consider the cyclic characteristics of the angle. First, calculate the difference between the buoy observed wind direction and the reanalysis wind field wind direction, and then apply angle normalization processing: when the difference is greater than 180 degrees, subtract the difference from 360 degrees and take the negative value; when the difference is less than -180 degrees, add the difference to 360 degrees and take the negative value. This processing ensures that the wind direction error always falls within the range of -180 degrees to 180 degrees, representing 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, the 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 deviation degree of the reanalysis wind field relative to the actual observation, provides basic data for the subsequent construction of the error field and wind field correction, and at the same time separates the two independent component errors of wind speed and wind direction, facilitating separate modeling and processing.

[0043] The specific implementation of step S05 is to extract the key spatio-temporal characteristic parameters affecting the wind field error distribution. First, the sea surface temperature field of the study area is extracted from satellite remote sensing data sources, with a spatial resolution of 0.05 degrees and a time resolution of daily average values. Then calculate the pressure gradient, and process the pressure values of adjacent grid points through the central difference method. The specific formula is: , where and respectively represent the pressure change rates in the east - west and north - south directions. The wind field curvature is calculated using the Laplace operator, approximated by the second - order derivative of the wind speed field: , where \(V\) is the wind speed scalar field. The seasonal cycle coefficient is obtained by performing a Fourier transform on historical wind field data, extracting the amplitudes and phases corresponding to the main frequency components, and constructing a seasonal variation term in the form of , where \(A\) is the annual amplitude, \(t\) is the current date, and \(\varphi\) is the initial phase. The diurnal variation index uses the cosine function to fit the intra - day variation characteristics, where \(B\) is the daily amplitude, \(t\) is the current hour, and \(\theta\) is the phase parameter. These spatio - temporal characteristic parameters comprehensively characterize the key physical factors affecting the wind field error, providing multi - dimensional input features for the Gaussian process regression model and improving the physical rationality and prediction accuracy of the error field construction.

[0044] The specific implementation of step S06 is to optimize the Gaussian process regression using an ocean physical - constraint neural network model. This model adopts a multi - layer encoding - decoding architecture, including 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 receives the air - sea interface Richardson number and sea surface roughness parameters as inputs, and converts the physical - field features into a 128 - dimensional latent representation vector through non - linear transformation. The boundary - layer dynamics embedding layer contains five sets of parameterized equations, which respectively describe the wind profile characteristics and turbulence effects under different stability conditions, and each set of equations corresponds to an atmospheric stability state. The multi - head sparse attention layer consists of eight attention heads, each head is responsible for capturing the feature correlations at different spatial scales. The number of attention heads is determined by the magnitude of the sea surface temperature gradient. When the gradient exceeds 0.05 °C / km, it increases to 12 heads. The attention sparsity is determined by the wind field curvature and intensity. The greater the wind field curvature, the lower the sparsity. The sparse attention mechanism filters out low - correlation connections by setting a threshold, significantly reducing the computational amount while retaining key information. The coupling index of the model is determined by a 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 a strongly unstable condition, the weight of the turbulence closure equation increases to 0.7; under stable stratification conditions, the influence of the thermodynamic equation increases to 0.6. This model effectively improves the prediction accuracy of the Gaussian process regression by integrating physical laws and data - driven methods, especially performing excellently in complex air - sea interaction regions and extreme weather conditions.

[0045] The specific implementation of step S07 is to establish a spatio - temporal covariance model based on the Gaussian process regression. First, the Matern3 / 2 kernel function is defined as the basis of the covariance function, and its functional 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 spatio-temporal 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 through maximum likelihood estimation. After the covariance matrix is constructed, sparse matrix technology is used to process large-scale data sets, and variational inference methods are adopted to accelerate model solution. To handle the characteristic differences in different regions, a local adaptive strategy is implemented, and the kernel function parameters are dynamically adjusted according to the regional wind field characteristics. Finally, the spatio-temporal covariance model outputs the wind speed error field and the wind direction error field. Both error fields are functions continuously defined on the ocean space and time dimensions, and error predictions can be made at any position and time point. This model makes full use of the spatial correlation and temporal continuity of the observation errors, can provide reasonable error interpolation in sparse observation point areas, and realizes the conversion from discrete observation points to continuous error fields.

[0046] The specific implementation of step S08 is to correct the reanalysis wind field data using the error field constructed in step S07. In the correction process, the spatio-temporal coordinates of the point to be corrected are first determined, and then the wind speed error field value and the wind direction error field value at the corresponding position and time are queried from the error field. The wind speed correction uses direct addition: corrected wind speed = reanalysis wind field wind speed + wind speed error field value. The wind direction correction needs to consider the cyclic property of the angle: corrected wind direction = (reanalysis wind field wind direction + wind direction error field value) % 360, where % represents the modulo operation to ensure the result is within the range of 0 to 360 degrees. To ensure the physical reasonableness of the correction result, the upper limit of the wind speed correction is set to 30% of the original wind speed, and correction values exceeding this threshold will be truncated to avoid unreasonable results caused by overcorrection. For wind direction correction, the maximum correction angle is limited to 45 degrees to maintain the spatial continuity of the wind field. In strong wind speed gradient regions such as typhoons, a special processing strategy is applied, and the correction intensity decreases as the wind speed gradient increases to ensure that the corrected wind field maintains the typhoon structure characteristics. Through this correction process, the systematic bias in the reanalysis wind field data is systematically eliminated, and the accuracy of the wind field data is improved, especially in areas affected by complex nearshore terrain and special weather systems, where the correction effect is more significant.

[0047] 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 predicting the mean value of the error field, Gaussian process regression also outputs the prediction variance, which directly quantifies the uncertainty of the prediction. The system uses this feature to calculate the 95% confidence interval for each correction point, and 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, while a wide interval indicates high uncertainty. The system simultaneously generates a spatial distribution map of uncertainty, visually showing the spatial variation characteristics of uncertainty using color gradients. To quantify the overall uncertainty level, the regional average uncertainty index is calculated, defined as the ratio of the confidence interval width to the wind speed correction amount. In regions with sparse observations, the uncertainty is usually high. The system automatically marks the regions where the uncertainty exceeds the threshold, and it is recommended that users use the correction results of these regions with caution. This step provides a reliability assessment of the correction results, which is helpful for risk control and decision-making support in the practical application of wind field data, especially in fields such as offshore engineering safety assessment and navigation decision-making.

[0048] 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. In this method, all the data of one buoy are excluded from the training dataset as the test set each time, and the data of the remaining buoys are used to train the model and predict the wind field error at the position of the excluded buoy. Specifically, during implementation, first, a buoy index list is constructed, which contains the identifiers of all the buoys participating in the validation. Then, the loop validation is performed. In each iteration, one buoy is selected as the test buoy, and all its data are excluded. The remaining data are used to train the model. The model prediction error is applied to the reanalysis wind field data at the position of the test buoy and corrected. The corrected result is compared with the actual observed data of the buoy, and three indicators, namely the root mean square error, bias, and correlation coefficient, are calculated. The formula for the root mean square error is: , where is the observed value, is the predicted value after correction. The bias is calculated as the average difference between the predicted value and the observed value: . The correlation coefficient reflects the linear correlation between the predicted value and the observed value, and the ideal value is 1. The loop is performed until each buoy has been used as the test set once. Finally, the validation results of all the buoys are comprehensively combined to evaluate the overall generalization performance of the model. This validation method simulates the application scenario of the model at unknown positions, can objectively evaluate the actual effect and reliability of the correction method, and provides a scientific basis for model optimization and practical application.

[0049] The detailed structure of the ocean physics-constrained neural network model is a multi-layer encoding and decoding architecture, which consists of three main parts: the surface layer physical process encoding layer, the boundary layer dynamics embedding layer, and the multi-head sparse attention layer. The surface layer physical process encoding layer is composed of three fully connected layers with the number of nodes being 256, 512, and 256 respectively. Using the GELU activation function, it receives the Richardson number at the air-sea interface and the roughness length parameter as inputs, and converts the physical field features into a 128-dimensional latent representation vector. Batch normalization and residual connections are used between layers to improve the training stability. The boundary layer dynamics embedding layer contains five sets of parameterized equations, each corresponding to an atmospheric stability state, graded from strongly unstable to strongly stable. Each set of equations is implemented by a three-layer fully connected network. The inputs are parameters such as height, wind speed, and temperature gradient, and the output is the wind profile feature under the corresponding stability condition. A soft switching function is used to achieve a smooth transition between different stability conditions. The multi-head sparse attention layer consists of 8 attention heads. Each head focuses on the feature correlations at different spatial scales. The number of heads is determined by the sea surface temperature gradient, and the attention sparsity is determined by the wind field curvature and intensity. The dimensions of the query, key, and value matrices of each attention head are 64, and the attention weights are calculated through scaled dot product. The output layer of the model is fused with the Gaussian process regression result through a residual connection method to achieve the dual constraints of physical laws and data-driven.

[0050] The detailed steps for establishing the training dataset of the ocean physics-constrained neural network model are as follows: First, select ten years of high-quality data from the global buoy observation network as the basic dataset, and select the observation records of a total of 387 buoys from 2010 to 2020, with the total sample size exceeding 78 million. The dataset is stratified sampled according to seasons and sea areas to ensure a uniform distribution of training samples in space. Each 5-degree × 5-degree grid contains at least 1000 training samples. For each training sample, a feature vector containing the differences in wind speed and direction and related physical field parameters is constructed. The dimension of the feature vector is 128, including the near-surface layer structure parameters and the turbulence intensity index. The error correction value of the labeled sample is calculated by comparing the historical buoy observation record with the reanalysis data at the corresponding location. The time span of the training dataset covers multiple typical meteorological events, including typhoons, seasonal alternations, and extreme weather processes, to ensure the generalization ability of the model. For sparse regions, the simulation results of the physical model are used for data augmentation to generate synthetic training samples to make up for the lack of observational data. A total of about 500,000 synthetic samples are generated, mainly distributed in the high-latitude regions of the southern hemisphere and the open ocean. The training data is divided into a training set, a validation set, and a test set according to the ratio of 80%, 10%, and 10% respectively, which are used for model parameter optimization, hyperparameter adjustment, and final performance evaluation. In particular, when dividing the dataset, ensure that the data of the same buoy only appears in one dataset to avoid overestimation of performance caused by data leakage.

[0051] Specifically, the principle of the present invention is as follows: The core principle of the present invention to solve the problem of significant spatio-temporal errors in reanalysis wind field data in complex marine 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: First, the present invention introduces a dynamic time window mechanism to dynamically adjust the matching accuracy between buoy observation data and reanalysis wind field data according to the characteristics of the meteorological system. This design is based on the significantly different time evolution characteristics of different meteorological systems: Under stable meteorological conditions, the wind field changes relatively slowly, and a relatively wide time window (such as plus or minus 1 hour) can be used for matching; while in a rapidly changing environment such as a typhoon, the wind field characteristics can change drastically in a short period of time, and at this time, a narrower time window (such as plus or minus 5 minutes) is required to ensure the matching accuracy. This adaptive adjustment strategy ensures the time alignment accuracy of the corrected basic data from the source.

[0052] Second, the ocean physical constraint neural network model constructed by the present invention embeds the knowledge of ocean boundary layer physical processes into the data-driven learning framework. The surface layer physical process encoding layer fuses the Richardson number and roughness parameters at the air-sea interface, converting the physical field characteristics into a latent 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 structure characteristics of the wind field within the boundary layer; the multi-head sparse attention layer can focus on the feature correlations at different spatial scales, adapting to the interactive characteristics of multi-scale physical processes in the ocean environment. This multi-level structure design enables the model to have the ability to describe the wind field error distribution in complex marine environments.

[0053] Third, the present invention dynamically adjusts the physical constraint intensity through the boundary layer stability determination 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, combines wave height and turbulent kinetic energy parameters to construct a stability determination vector, and then dynamically adjusts the coupling index of the physical equations in the neural network. Increase the weight of the turbulent closure equation under strong unstable conditions, and enhance the influence of the thermodynamic equation in a stable stratification environment. This adaptive adjustment mechanism ensures that the model can select the optimal physical constraint method for different air-sea interaction states.

[0054] Fourth, the present invention uses the Matern3 / 2 kernel function to construct a spatio-temporal covariance model of the error field. Compared with the traditional Gaussian kernel function, this kernel function has better adaptability to the non-smoothing characteristics of the data and can more accurately describe the local change characteristics of the wind field in regions with significant gradients such as circulation boundaries and frontal zones. This kernel function design highly matches the physical characteristics of the ocean wind field, improving the modeling ability of Gaussian process regression in complex wind field environments.

[0055] Finally, the present invention fuses the neural network output and the Gaussian process regression result through a residual connection method, achieving a dual optimization of physical constraints and statistical characteristics. This fusion strategy not only retains the prior knowledge of the physical model about the ocean wind field structure but also maintains the fitting ability of the statistical model to the observed data, effectively overcoming the limitations of a single method and improving the comprehensive performance of wind field correction.

[0056] The following provides a specific Embodiment 1 of the present invention, and the specific implementation manners of each step in this Embodiment 1 are described in detail as follows.

[0057] The specific implementation manner of step S01 is to process the buoy observation data through a multi-level quality inspection system to ensure data reliability. First, the original data uploaded in real time by buoy observation stations distributed globally is received, including parameters such as wind speed, wind direction, air pressure, and sea surface temperature. Subsequently, the physical threshold method is used for preliminary screening. Data with a wind speed greater than 50 meters per second is marked as an outlier because wind speeds exceeding this threshold rarely occur in the natural environment. At the same time, it is checked whether the wind direction data is within the range of 0 to 360 degrees, and those outside the range are marked as invalid. Then, the box plot method is applied to identify outliers in a statistical sense, and the quartiles 、 and the interquartile range of each parameter are calculated. The formula for the interquartile range is: ; In the formula, is the lower quartile of the data set, representing the value at the 25% position after sorting; is the upper quartile of the data set, representing the value at the 75% position after sorting.

[0058] Based on the interquartile range, an outlier determination criterion is constructed: ; ; In the formula, is the lower threshold of the data; is the upper threshold of the data; data points outside the range are marked as potential outliers.

[0059] For the air pressure parameter, a reasonable change rate threshold of 3 hPa per hour is set, and data exceeding this change rate requires further review. Finally, physical consistency is checked, for example, verifying whether the relationship between wind speed and significant wave height conforms to the basic principles of marine meteorology. Through the above processing, high-quality data is retained for subsequent correction model construction, improving the reliability of the model input data, and thus enhancing the accuracy of the final correction result.

[0060] The specific implementation of step S02 is to achieve the precise spatial matching of buoy observation data and reanalysis wind field data through the bilinear interpolation method. First, determine the longitude and latitude coordinates of the buoy, and then identify the four nearest reanalysis wind field data grid points around this coordinate. These grid points form a rectangular area containing the buoy position. For each buoy position , calculate its distances to the four grid points, and determine the weight coefficients based on the inverse of the distances. The specific weight calculation formula is: ; In the formula, represents the weight coefficient of the th grid point; represents the Euclidean distance from the buoy to the th grid point, and the calculation formula is ; is the coordinate of the th grid point; the sum of the weights is 1, that is .

[0061] Based on the calculated weight coefficients, perform weighted averaging on the wind field data of the four grid points to obtain the interpolated wind field data at the buoy position: ; ; In the formula, and are the wind speed and wind direction at the interpolated buoy position respectively; and are the wind speed and wind direction of the th grid point respectively.

[0062] This step solves the problem of inconsistent positions between buoy observation points and reanalysis wind field grid points, realizes the coordination and unity in the spatial dimension, and lays a foundation for subsequent error analysis. Compared with the simple nearest neighbor interpolation method, bilinear interpolation can better retain the spatial continuity of the wind field and is suitable for expressing the spatial characteristics of the wind field under complex sea conditions.

[0063] The specific implementation of step S03 is to dynamically adjust the width of the time matching window according to meteorological conditions to achieve precise alignment in the time dimension. First, by analyzing the current type of meteorological conditions, construct a meteorological stability index , and the calculation formula is: ; In the formula, is the meteorological stability index, and its value range is , and the larger the value, the faster the meteorological conditions change; is the standard deviation of the wind speed within a certain time period; is the average wind speed within the same time period; is the absolute change in air pressure per unit time; is the reference air pressure change, set to 5 hPa; is the standard deviation of the wind direction; is the reference wind direction change, set to 45 degrees; 、 、 are the weight coefficients, set to 0.5, 0.3, and 0.2 respectively, satisfying 。

[0064] Based on the meteorological stability index dynamically determine the time window width : ; In the formula, is the half-width of the time window; when is determined as the rapid change condition, the time window is set to minutes; when it is the normal condition, and the time window is minutes; when it is the stable condition, and the time window is extended to minutes.

[0065] This step effectively solves the problem of the difference in time resolution between buoy observation data and reanalysis wind field data, improves the time matching accuracy, and provides more reliable basic data for the subsequent construction of the error field.

[0066] The specific implementation of step S04 is to calculate the wind speed error and wind direction error of the spatio-temporal matching points. The wind speed error is calculated using the algebraic difference method: ; In the formula, is the wind speed error; is the buoy observed wind speed; is the corresponding reanalysis wind field wind speed. A positive value indicates that the reanalysis wind field underestimates the actual wind speed, and a negative value indicates an overestimation.

[0067] The calculation of the wind direction error needs to consider the cyclic characteristics of the angle: ; ; In the formula, is the original wind direction difference; is the buoy observed wind direction, with a range of ; is the reanalysis wind field wind direction, with a range of ; is the normalized wind direction error, with a range of .

[0068] To eliminate the influence of random errors, median filtering is applied to the error values at multiple consecutive time points: ; ; In the formula, and are the filtered wind speed and wind direction error at time point respectively; represents the median operation.

[0069] This step quantifies the deviation degree of the reanalysis wind field relative to the actual observation, provides basic data for subsequent error field construction and wind field correction, and at the same time separates the two independent component errors of wind speed and wind direction, facilitating separate modeling and processing.

[0070] The specific implementation of step S05 is to extract the key spatio-temporal characteristic parameters affecting the wind field error distribution. First, the sea surface temperature field of the study area is extracted from satellite remote sensing data sources , with a spatial resolution of 0.05 degrees and a temporal resolution of daily average. Then, the pressure gradient is calculated by processing the pressure values of adjacent grid points through the central difference method: ; ; ; In the formula, represents the pressure value at position ; and represent the grid spacings in the east-west and north-south directions respectively; and represent the pressure change rates in the east-west and north-south directions respectively; is the magnitude of the pressure gradient.

[0071] The wind field curvature is calculated using the Laplace operator and approximated by the second-order derivative of the wind speed field: ; ; ; In the formula, represents the wind speed value at position ; 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, reflecting the spatial change rate of the wind field.

[0072] Seasonal cycle coefficient Obtained by performing Fourier transform on historical wind field data: ; In the formula, is the seasonal cycle coefficient; is the annual amplitude, obtained through Fourier analysis, and the typical value is 15% - 25% of the wind speed; is the current date (the day of the year); is the initial phase, obtained by fitting historical data.

[0073] Day-night variation index Adopt cosine function to fit the intra-day variation characteristics: ; In the formula, is the day-night variation index; is the daily amplitude, generally 5% - 15% of the average wind speed; is the current hour (the hour of the day); is the phase parameter, obtained by data fitting.

[0074] These spatio-temporal characteristic parameters comprehensively characterize the key physical factors affecting the wind field error, provide multi-dimensional input features for the Gaussian process regression model, and improve the physical rationality and prediction accuracy of the error field construction.

[0075] The specific implementation of step S06 is to optimize the Gaussian process regression using the ocean physical constraint neural network model. This 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 determination function is based on the Richardson number Calculate: ; In the formula, is the Richardson number; is the acceleration due to gravity, with a value of 9.8 meters per square second; is the potential temperature difference, with the unit of Kelvin; is the height difference, with the unit of meter; is the average potential temperature, with the unit of Kelvin; is the wind speed difference, with the unit of meter per second.

[0076] Based on the Richardson number , wave height and turbulent kinetic energy Construct the determination vector : ; In the formula, is the judgment vector, which consists of five components corresponding to different stability intervals; the calculation of each component is as follows: ; ; ; ; ; In the formula, is the sigmoid function; is the significant wave height, with a value of 15 meters; is the maximum turbulent kinetic energy, with a value of 5 square meters per square second. to respectively represent the judgment scores of strong instability, weak instability, neutral, weak stability and strong stability conditions, and their value ranges are all .

[0077] According to the judgment vector , determine the coupling index of the physical equation of the boundary layer in the model: ; In the formula, is the coupling index, and its value range is , and the larger the value, the higher the physical constraint intensity; represents the maximum value in the vector .

[0078] In the multi-head sparse attention layer, the number of attention heads is determined by the sea surface temperature gradient : ; In the formula, is the number of attention heads; is the absolute value of the sea surface temperature gradient.

[0079] The attention sparsity is jointly determined by the wind field curvature and the wind speed : ; In the formula, is the attention sparsity, and its value range is , and the smaller the value, the denser the attention connection; is the maximum wind field curvature, with a value of per square meter; $V_{max}$ is the maximum wind speed, with a value of 50 m / s; and represent the operations of taking the maximum and minimum values respectively.

[0080] This step effectively improves the prediction accuracy of Gaussian process regression by integrating physical laws and data-driven methods, especially performing excellently in complex air-sea interaction regions and extreme weather conditions.

[0081] The specific implementation of step S07 is to establish a spatio-temporal covariance model based on Gaussian process regression. First, define the Matern 3 / 2 kernel function as the basis of the covariance function: ; In the formula, is the covariance between point and point ; is the signal variance, obtained by data estimation; is and is the distance between them; is the length scale parameter, controlling the rate of decay of correlation with distance.

[0082] To construct the spatio-temporal covariance structure, different characteristic length parameters are set for the spatial dimension and the temporal dimension respectively: ; In the formula, is the spatial distance; is the temporal distance; is the spatial scale parameter, set to 50 km; is the temporal scale parameter, set to 6 hours.

[0083] Given the observed data set , where is the input feature vector, is the observed error value, and the prediction distribution of Gaussian process regression is: ; In the formula, is the error value at the prediction point ; represents a normal distribution with a mean of and a variance of ; and are the mean and variance of the prediction distribution respectively, and the calculation formulas are: ; ; In the formula, For the prediction point and all the observation points the covariance vector between them; is the covariance matrix between the observation points, with elements ; is the observation noise variance; is the identity matrix; is the covariance of the prediction point itself; is the observation error value vector .

[0084] For large-scale data sets, a sparse approximation method is adopted to reduce the computational complexity: ; ; In the formula, is the selected inducing points; is the covariance vector between the prediction point and the inducing points; is the covariance matrix between the inducing points; is the correction matrix related to the observed data.

[0085] Finally, the spatio-temporal covariance model outputs the wind speed error field and the wind direction error field , and both error fields are functions continuously defined on the ocean space and time dimensions: ; ; In the formula, and are the predicted means of the wind speed and wind direction errors at the positions , time respectively.

[0086] This step makes full use of the spatial correlation and temporal continuity of the observation errors, can provide reasonable error interpolation in the sparse area of the observation points, and realizes the conversion from discrete observation points to a continuous error field.

[0087] The specific implementation of step S08 is to correct the reanalysis wind field data using the error field constructed in step S07. The correction process first determines the spatio-temporal coordinates of the point to be corrected, and then queries the wind speed error field value and the wind direction error field value at the corresponding position and time from the error field. The wind speed correction uses direct addition: ; In the formula, is the corrected wind speed; For re - analyzing the wind speed of the wind field; is the wind speed error field value.

[0088] To ensure the physical rationality of the correction result, set the upper limit of wind speed correction to 30% of the original wind speed: ; In the formula, is the sign function of, taking values of +1 or -1.

[0089] The wind direction correction considers the cyclic property of the angle: ; In the formula, is the corrected wind direction; is the wind direction of the re - analyzed wind field; is the wind direction error field value; represents the modulo operation, ensuring the result is within the range.

[0090] Limit the maximum wind direction correction angle to 45 degrees: ; ; In the formula, is the limited wind direction error field value.

[0091] In strong wind speed gradient regions such as typhoons, apply a special processing strategy, and the correction intensity decreases with the increase of the wind speed gradient : ; ; ; In the formula, is the correction coefficient, and its value range is , the greater the wind speed gradient, the smaller the correction coefficient; is the adjustment parameter, with a value of 0.1; is the norm of the wind speed gradient.

[0092] Through this correction process, the systematic deviation in the re - analyzed wind field data is systematically eliminated, improving the accuracy of the wind field data. Especially in areas affected by complex near - shore terrain and special weather systems, the correction effect is more significant.

[0093] The specific implementation of step S09 is to generate an uncertainty evaluation result of the corrected wind field data based on the variance output of Gaussian process regression. In addition to predicting the mean of the error field, Gaussian process regression also outputs the prediction variance, which directly quantifies the uncertainty of the prediction. The system utilizes this feature to calculate the 95% confidence interval for each correction point: ; where, is the 95% confidence interval at position , time ; is the predicted error value, i.e., or ; is the predicted standard deviation, which is directly output by the Gaussian process regression model; the coefficient 1.96 comes from the 95% confidence interval of the standard normal distribution.

[0094] To quantify the overall uncertainty level, calculate the regional average uncertainty index ; where, is the uncertainty index, and the smaller the value, the more reliable the prediction result; is the number of grid points in the evaluation area; is the predicted standard deviation at position , time ; is the absolute value of the prediction error.

[0095] The system automatically marks the areas where the uncertainty exceeds the threshold, and recommends that users use the correction results of these areas with caution: ; where, is the masking function, and the value of 1 indicates that the uncertainty of the correction result at this position is relatively high; is the uncertainty threshold, which is set to 0.5.

[0096] This step provides a reliability evaluation of the correction result for the user, which is helpful for risk control and decision-making support in the practical application of wind field data, especially in the fields of offshore engineering safety assessment and navigation decision-making.

[0097] The specific implementation of step S10 is to evaluate the generalization ability of the ocean physical constraint neural network model using the leave-one-buoy cross-validation method. In this method, all the data of one buoy are excluded from the training dataset as the test set each time, and the data of the remaining buoys are used to train the model and predict the wind field error at the position of the excluded buoy. Specifically, first construct a buoy index list , which contains all the buoy identifiers participating in the validation. For each buoy , perform the following verification steps: Exclude all data of the buoy from the training dataset to obtain a training subset and a test subset : ; ; where represents the buoy identifier associated with the data point .

[0098] Use the training subset to train the ocean physical constraint neural network model to obtain model parameters . For each data point in the test subset , use the trained model to predict the error value and compare it with the actual observed error value .

[0099] Calculate three evaluation metrics: root mean square error , bias and correlation coefficient : ; ; ; where represents the number of samples in the test subset ; and are the average values of the observed and predicted values in the test subset, respectively.

[0100] Repeat until each buoy has been used as the test set once, and then calculate the global evaluation metrics: ; ; ; where , and are the global root mean square error, bias, and correlation coefficient, respectively, used to comprehensively evaluate the generalization performance of the model.

[0101] To determine whether the correction result is significantly better than the original reanalysis wind field data, introduce the improvement rate index : ; In the formula, is the root mean square error between the reanalysis wind field before correction and the buoy observation; is the root mean square error between the wind field after correction and the buoy observation; is the improvement rate, with the unit of percentage. The larger the value, the more significant the correction effect.

[0102] This verification method simulates the application scenario of the model at unknown locations, can objectively evaluate the actual effect and reliability of the correction method, and provides a scientific basis for model optimization and practical application. In particular, for different sea conditions and meteorological conditions, the conditional evaluation indicators are further calculated: ; In the formula, is the root mean square error under specific conditions ; is the subset of test data that meets the condition . The condition can be the wind speed range, sea condition type, stability category, etc.

[0103] This step comprehensively evaluates the prediction ability of the ocean physics-constrained neural network model at unknown locations through the leave-one-buoy cross-validation method, verifies the effectiveness and reliability of the correction method, and provides a scientific basis for subsequent model optimization and practical application.

[0104] The detailed structure of the ocean physics-constrained neural network model is a multi-layer encoding-decoding architecture, which includes three main parts: the surface layer physical process encoding layer, the boundary layer dynamics embedding layer, and the multi-head sparse attention layer. The surface layer physical process encoding layer consists of three fully connected layers, and the hierarchical structure is as follows: ; ; ; In the formula, is the input feature vector, which includes the air-sea interface Richardson number and the roughness parameter; , , are the weight matrices; , , are the bias vectors; is the Gaussian error linear unit activation function; is the batch normalization operation; , , are the intermediate hidden layer representations; the residual connection is added to the last layer to improve the training stability.

[0105] The boundary layer dynamics embedding layer contains five sets of parameterized equations, each corresponding to an atmospheric stability state, graded from strongly unstable to strongly stable. Each set of parameterized equations is implemented by a three-layer fully connected network: ; where represents the parameterization function under the th stability condition, ; is the height; is the wind speed; is the temperature gradient; represents the th fully connected layer under the th stability condition; is the rectified linear unit activation function; represents the concatenation of input features.

[0106] The multi-head sparse attention layer consists of multiple attention heads, each responsible for capturing feature correlations at different spatial scales: ; ; ; ; where represents the output of the th attention head; is the attention calculation function; , , are the query, key, and value matrices respectively; is the input feature; , , are the corresponding weight matrices; is the attention sparsity.

[0107] The attention calculation function is defined as: ; where is the softmax normalization function; is the dimension of the key vector; is the sparsification function, which retains the largest part of the attention weights according to the sparsity and sets the rest to zero.

[0108] Finally, the output of the multi-head attention is the concatenation of the outputs of each head, followed by a linear transformation: ; where is the output of the multi-head attention; represents the concatenation operation; is the number of attention heads; is the weight matrix of the output linear transformation.

[0109] The model output layer is fused with the Gaussian process regression result through a residual connection: ; In the formula, is the error value of the final prediction; is the predicted output of the neural network model; is the predicted output of the Gaussian process regression; is the coupling index, which is determined by the boundary layer stability determination function and is used to balance the proportion of physical laws and data-driven.

[0110] The detailed steps for establishing the training dataset of the ocean physical constraint neural network model are as follows: First, select ten years of high-quality data from the global buoy observation network as the basic dataset, and select the observation records of a total of 387 buoys from 2010 to 2020, with a total sample size of more than 78 million. The dataset is stratified and sampled according to seasons and sea areas to ensure a uniform distribution of the training sample space, and each 5-degree × 5-degree grid contains at least 1000 training samples. A feature vector containing the wind speed and direction differences and related physical field parameters is constructed for each training sample. The dimension of the feature vector is 128, including the near-surface layer structure parameters and the turbulence intensity index. The error correction value of the labeled sample is calculated by comparing the historical buoy observation record with the reanalysis data at the corresponding location. The time span of the training dataset covers multiple typical meteorological events, including typhoons, seasonal alternations, and extreme weather processes, to ensure the generalization ability of the model. For sparse regions, the simulation results of the physical model are used for data augmentation to generate synthetic training samples to make up for the lack of observational data. A total of about 500,000 synthetic samples are generated, mainly distributed in the high-latitude regions of the southern hemisphere and the open ocean. The training data is divided into a training set, a validation set, and a test set according to the ratio of 80%, 10%, and 10% respectively, which are used for model parameter optimization, hyperparameter adjustment, and final performance evaluation. In particular, when dividing the dataset, ensure that the data of the same buoy only appears in one dataset to avoid overestimation of performance caused by data leakage.

[0111] To better understand and implement the present invention, the following provides Example 2 of a specific application scenario of the present invention: In a study on the accurate prediction of an ocean wind field carried out 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 wind field prediction accuracy in the offshore area. The study area covers the range from 18 degrees north latitude to 22 degrees north latitude and from 112 degrees east longitude to 116 degrees east longitude, and includes 10 meteorological and oceanographic buoy observation stations. The study used the buoy observation data and ERA5 reanalysis wind field data from January 2021 to December 2021 as the basic data set. First, outliers in the buoy observation data were screened. A total of 283 wind speed outliers and 169 wind direction outliers were identified through the box plot method, accounting for 2.6% of the total sample size. The statistical characteristics of the observation data at each buoy station after outlier screening are shown in Table 1.

[0112] Table 1 Statistical characteristics of wind field data at buoy observation stations

[0113] Subsequently, the spatio-temporal matching of the buoy observation data and the ERA5 reanalysis wind field data was carried out. The spatial resolution of the ERA5 data is 0.25 degrees, and the time resolution is 1 hour. The wind field data corresponding to the buoy positions were extracted by the bilinear interpolation method. Considering that there are many typhoon activities in the study area in summer and the influence of the northeast monsoon is relatively stable in winter, a dynamic time window was used for time matching. During the typhoon period (July - September), the window width was set to ±5 minutes, in winter (December - February), the window width was set to ±60 minutes, and for the rest of the time it was ±15 minutes. After spatio-temporal matching, the wind speed error and wind direction error of each buoy station were calculated. The results showed that the ERA5 reanalysis wind field generally underestimated the actual wind speed, with an average underestimation amplitude of 12.8%. The wind direction error showed an obvious sea-land distribution characteristic, and the wind direction error in the nearshore area was significantly greater than that in the open sea area.

[0114] When extracting the spatio-temporal characteristic parameters affecting the wind field error distribution, the sea surface temperature data obtained by satellite remote sensing showed that there was an obvious temperature front in the study area, and the maximum sea surface temperature gradient reached 0.08 °C / km. The pressure gradient was calculated from the ERA5 pressure field, and the maximum pressure gradient during the typhoon period 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 relatively large, and the maximum value reached per square meter. The seasonal cycle coefficient extracted by Fourier analysis showed that the annual amplitude of the wind field in the study area was 21.3% of the average wind speed, and the maximum wind speed occurred in winter. The diurnal variation index showed that the wind speed during the day was generally higher than that at night, and the daily amplitude was 8.7% of the average wind speed.

[0115] Determine the Richardson number distribution under different sea-air conditions through the boundary layer stability determination function, and construct a determination vector based on this. The results are shown in Table 2.

[0116] Table 2 Distribution table of boundary layer stability determination vectors under different conditions

[0117] When constructing a spatio-temporal covariance model based on Gaussian process regression, determine the optimal Matern 3 / 2 kernel function parameters through maximum likelihood estimation. The signal variance is 2.64, the spatial scale parameter is set to 52.3 km, and the temporal scale parameter is 6.2 h. Use the continuous wind speed error field and wind direction error field generated by this covariance model to correct the ERA5 reanalysis wind field data, and calculate the 95% confidence interval. The verification of the correction results shows that the improvement effects under various sea conditions evaluated by the leave-one-buoy cross-validation method are shown in Table 3.

[0118] Table 3 Improvement effects of wind field correction under different sea conditions

[0119] It can be seen from Table 3 that significant improvements have been achieved in the corrected wind field data under various sea conditions. The overall root mean square error has decreased from 1.92 m / s to 0.86 m / s, and the improvement rate has reached 55.2%. Especially in the nearshore area, the improvement rate is as high as 63.7%, which is mainly due to the good ability of the ocean physics constrained neural network model to capture the wind field deformation caused by complex terrain. At the same time, the wind direction error has also been significantly improved, and the average wind direction error has decreased from 18.3 degrees to 7.6 degrees.

[0120] Traditional offshore wind farm data correction methods mainly rely on statistical regression or simple physical models, and it is difficult to simultaneously consider complex spatio-temporal variation characteristics and air-sea interface physical processes. For example, the linear regression method and the nearest neighbor correction method can only capture simple linear or local relationships and perform poorly under complex sea conditions and extreme weather conditions. The improvement rate in the nearshore area is usually less than 20%. Although the numerical model nesting method considers physical processes, it has high computational costs and poor real-time performance. The present invention combines a hybrid method of an ocean physics-constrained neural network model and Gaussian process regression, which not only maintains the flexibility and efficiency of data-driven methods but also incorporates the constraints of ocean boundary layer physical laws. In particular, the spatio-temporal matching problem under different meteorological conditions is solved through a dynamic time window and a boundary layer stability determination function, significantly improving the wind field correction accuracy in extreme weather conditions such as typhoons and in the nearshore complex terrain area. At the same time, the uncertainty assessment provided by Gaussian process regression provides a quantitative basis for the reliability of wind field data, providing more reliable wind field data support for the siting of offshore wind farms, the safety assessment of ocean engineering, and navigation decision-making.

[0121] It should be noted that the detailed explanations of the variables involved in the present invention are shown in Tables 4, 5, and 6 below.

[0122] Table 4 Variable Explanation Table (Part 1)

[0123] Table 5 Variable Explanation Table (Part 2)

[0124] Table 6 Variable Explanation Table (Part 3)

[0125] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered within the protection scope of the present invention.

Claims

1. A method for correcting offshore reanalysis wind field data based on buoy observation data, characterized in that Including: Receiving buoy observation data and reanalysis wind field data, and performing outlier removal and quality control processing on the buoy observation data; Performing spatio-temporal 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 spatio-temporal characteristic parameters; optimizing Gaussian process regression using an ocean physical constraint neural network model, where the ocean physical constraint neural network model processes multi-scale spatio-temporal features using a sparse attention mechanism, and the coupling index of the ocean physical constraint neural network model is determined by a boundary layer stability determination function; establishing a spatio-temporal covariance model; and correcting the reanalysis wind field data using the error field.

2. The method for correcting offshore reanalysis wind field data based on buoy observation data according to claim 1, wherein The outlier removal and quality control processing specifically identifies outliers in the wind speed, wind direction, and pressure parameters of the buoy observation data by setting physical thresholds in combination with the box plot method, and data points outside the reasonable range are marked as outliers and do not participate in subsequent calculations.

3. The method for correcting the offshore reanalysis wind field data based on buoy observation data according to claim 2, wherein The spatio-temporal matching specifically uses the bilinear interpolation method to extract the wind field data of the grid points corresponding to the buoy position in the reanalysis wind field data to obtain spatio-temporally matched data.

4. The method for correcting the offshore reanalysis wind field data based on buoy observation data according to claim 3, wherein The bilinear interpolation method specifically realizes the reconciliation of spatial resolution by calculating the weighted average of the four nearest grid points of the reanalysis wind field data around the buoy position, and the weight is inversely proportional to the distance from the buoy to the grid point of the reanalysis wind field data.

5. The method for correcting offshore reanalysis wind field data based on buoy observation data according to claim 4, wherein The dynamic time window specifically automatically adjusts the width of the dynamic time window from 15 minutes to 1 hour according to the type of meteorological event to achieve precise alignment in the time dimension.

6. The method for correcting offshore reanalysis wind field data based on buoy observation data according to claim 5, wherein 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; for the wind direction error, specifically, the angular difference between the wind direction of the buoy observation data and the wind direction of the reanalysis wind field data is calculated, and considering the angular cycle characteristic, the minimum angular difference with an absolute value not exceeding 180 degrees is taken.

7. The method for correcting offshore reanalysis wind field data based on buoy observation data according to claim 6, wherein The dynamic time window is specifically a time period adaptively adjusted based on the type of meteorological event. A plus-minus 15-minute matching window is used under normal weather conditions, shortened to plus-minus 5 minutes under the rapidly changing meteorological environment of a typhoon, and extended to plus-minus 1 hour under stable meteorological conditions.

8. The method for correcting offshore reanalysis wind field data based on buoy observation data according to claim 7, wherein The spatio-temporal characteristic parameters include sea surface temperature, pressure gradient, wind field curvature, seasonal cycle coefficient, and diurnal variation index.

9. The method for correcting offshore reanalysis wind field data based on buoy observation data according to claim 8, wherein The structure of the ocean physical constraint neural network model is a multi-layer encoding and decoding architecture, including three main parts: a surface layer physical process encoding layer, a boundary layer dynamics embedding layer, and a multi-head sparse attention layer.

10. The method for correcting offshore reanalysis wind field data based on buoy observation data according to claim 9, wherein, The surface layer physical process encoding layer fuses the Richardson number of the air-sea interface and roughness parameters to convert the physical field characteristics into a latent representation. The boundary layer dynamics embedding layer contains five sets of parameterized equations to describe the wind profile characteristics and turbulent effects under different stability conditions. The multi-head sparse attention layer consists of eight attention heads, and each head focuses on the feature correlations at different spatial scales; Furthermore, the spatio-temporal characteristic parameters are specifically the key physical quantities and time characteristics that affect the distribution of wind field errors. The sea surface temperature is obtained through satellite remote sensing. The pressure gradient is calculated by the air pressure difference between adjacent grid points. The wind field curvature reflects the spatial change rate of the wind field. The seasonal cycle coefficient extracts the annual change characteristics through Fourier transform. The diurnal change index fits the intraday change characteristics through a sine function; Among them, the spatio-temporal covariance model constructs the error field using the Matern 3 / 2 kernel function. The spatial scale of the spatio-temporal covariance model is set to 50 kilometers, and the time scale of the spatio-temporal covariance model is set to 6 hours; Among them, the wind speed of the corrected wind field data is equal to the wind speed of the reanalysis wind field data plus the wind speed error field value at the corresponding position and time; Furthermore, it also includes generating an uncertainty assessment result of the corrected wind field data based on the output of the Gaussian process regression variance, and calculating and outputting the 95% confidence interval.

Images