GNSS buoy wind speed inversion method and system based on physical information neural network

Abstract

The invention belongs to the technical field of ocean observation, and particularly relates to a GNSS buoy wind speed inversion method and system based on a physical information neural network, by introducing the physical information neural network, a physical equation and a neural network are effectively combined, the precision and efficiency of wind speed inversion are optimized, the generalization ability of a model is improved, and the wind speed inversion accuracy is improved. Therefore, a more accurate, rapid and widely applicable wind speed inversion solution is realized.

Description

Technical Field

[0001] This application belongs to the field of marine observation technology, specifically relating to a GNSS buoy wind speed inversion method and system based on physical information neural networks. Background Art

[0002] Wind speed is a crucial parameter in marine meteorological monitoring. Traditional wind speed observation methods include weather buoys, satellite remote sensing, and ship-based measurements. These methods typically suffer from high costs, limitations, and poor real-time performance. GNSS buoys, as a novel marine observation device, can retrieve wind speed from wave spectrum data, providing a low-cost and efficient new approach to wind speed monitoring. However, existing GNSS buoy-based wind speed retrieval methods still suffer from insufficient accuracy, poor computational efficiency, and weak generalization capabilities.

[0003] Traditional wind speed retrieval methods mostly rely on classical physical models or simplified assumptions. These methods have low accuracy under complex sea conditions and require long training and inference times. Especially under extreme weather and strong wind conditions, the retrieval accuracy of existing methods cannot meet the needs of real-time monitoring. Summary of the Invention

[0004] This invention provides a GNSS buoy wind speed inversion method and system based on physical information neural networks (PINNs), aiming to solve the problems of insufficient wind speed inversion accuracy, computational efficiency, and model generalization ability in existing technologies. By introducing the constraints of physical equations into the neural network training process, this invention combines the advantages of physical models and data-driven methods. Specifically, this invention designs a deep neural network and uses physical information to guide the network's learning, ensuring that the network can accurately extract wind speed information from complex wave spectrum data. Simultaneously, PINNs optimize network training by directly integrating physical equations, avoiding the error accumulation problem that may occur in traditional methods, and significantly improving training efficiency and inversion accuracy. The technical solution is as follows: A method for inverting wind speed from a GNSS buoy based on a physical information neural network includes the following steps: S1. Data Acquisition and Preprocessing; S2. Construct a wind speed and direction inversion model based on the Physical Information Neural Network (PINNs); introduce physical equations during the training process of the PINNs to constrain the output of the PINNs model through inverse modeling. S3. Wind speed and direction prediction.

[0005] Preferably, the data preprocessing step S1 is as follows:

[0006] Acquire GNSS data to obtain the buoy's three-dimensional motion information in the NEU direction. The frequency spectrum and Fourier coefficients of the waves and It can be calculated using the following formula: ; ; ; ; in, Represents energy components in different directions; It is the imaginary part of the cross spectrum, reflecting the phase difference between two signals and describing their phase asymmetry; It is the real part of the cross spectrum, reflecting the synchronicity of the two signals and describing the common changes in their phase and amplitude.

[0007] Preferably, the physical information neural network model PINNs in step S2 includes an input layer, a hidden layer, and an output layer; The input layer receives wave spectrum data and Fourier coefficients as input features, including wave spectrum components. Representing waves at Frequency domain energy density distribution along the direction; Fourier coefficients , Characterizing wave signals at frequency The periodic decomposition characteristics at the location; The hidden layer processes the input data layer by layer through forward propagation, extracts the spatial distribution features and frequency domain decomposition characteristics of the wave spectrum, and maps them to the wind speed-related potential representation space. The output layer integrates the high-dimensional features extracted from the hidden layer to generate a joint prediction result for wind speed and wind direction. For the predicted wind speed value... A linear activation function is used to output continuous positive values; for wind direction prediction values... After selecting a linear activation function, the angle range is constrained through modulo operation.

[0008] Preferably, the constraint and inverse modeling steps for the physical equations are as follows: The inverse modeling method based on wave energy balance infers wind speed and direction from wave spectrum data. It assumes that the wave field is in a quasi-uniform and quasi-stationary state, and the source term balance equations are valid at all frequencies. and direction The components are approximately closed, that is: ; Among them: wind speed at a height of 10 meters and wind direction ; Input source for wind; For wave breakup and dissipation source terms; For four-wave interaction source terms; directional wave spectrum ; Directional wave spectrum Decomposed into frequency spectrum and directional distribution The product of: ; in: It is the vertical frequency spectrum ; It is a directional distribution function that satisfies the normalization condition. From Fourier coefficients reconstruction; The Fourier expansion of the directional distribution is as follows: ; By integrating the frequency and direction, we obtain: ; Wind direction through overall motion stress The wind direction alignment constraint is defined as: ; ; ; in: This represents the total kinetic stress vector exerted by the wind on the sea surface. It is the portion of the momentum that the wave gains from the wind; Indicates viscous stress; Indicates background stress, It is gravitational acceleration; It is the wavenumber vector, and its direction is the same as the direction of wave propagation. For wavenumber scalar, It is the wave angular frequency; ρ represents the density ratio of air to water; c represents the wave phase velocity. Characterizing the dynamic changes in sea surface roughness driven by wind, among which It is the friction speed. It is the Charnock constant; Surface roughness length, is the von Kármán constant.

[0009] Preferably, a source term approximation is introduced: a1. Wind input source item It can be expressed as a quasi-linear approximation as: ; ; ; when hour, ; ; ; and These are adjustable parameters; a2. Dissipation source term , representing the wave energy dissipation rate, is approximated using the direction-dependent saturation, and is expressed as: ; in: Indicates the wave in a given direction and wave number The saturation level on the wave surface measures whether the wave energy has reached the breaking point. The larger the wave, the more unstable it is and the more easily it breaks. ; ; Indicates the speed of the wave group. Represents the integral variable. The directional range of the dissipation effect; waves beyond this angle do not contribute to saturation. The parameter adjustment coefficient, representing the dissipation intensity, controls the dissipation intensity of the white wave breakage and is determined based on model calibration; The critical value of wave saturation, when When the waves break apart, dissipation occurs; otherwise, there is no dissipation.

[0010] Preferably, the physical equations are introduced into the training process through a loss function, namely, physical loss. Including overall source term equilibrium constraints and wind direction stress constraints, the calculation is as follows: Overall source term balancing constraints: ; Wind direction stress constraint: ; in This refers to the stress amplitude. The total physical loss is expressed as: .

[0011] Preferably, a physical equation is introduced: after each forward propagation, based on the predicted wind speed... Wind direction forecast Calculate physical loss and data loss Joint optimization, the total loss function is: ; ; For adjustable weights, N is used to calculate the number of samples for data loss.

[0012] Preferably, the physical equations are introduced during the model training phase, and the total loss is calculated using an optimization algorithm. gradient of network parameters Update network parameters until convergence; Calculate wind speed deviation and wind direction deviation: ； ; Actual wind speed value Actual wind direction.

[0013] Preferably, the data acquisition uses a sampling frequency of 2 Hz, and a complete data sequence is acquired every t time. The acquired data format is RTCM format.

[0014] A GNSS buoy wind speed inversion system based on physical information neural network includes a data acquisition module, a data processing module, and an output module; Data acquisition module: The buoy's built-in GNSS receiver can receive signals from multiple satellites in real time for positioning and provide accurate timestamps for data synchronization; the buoy's data acquisition uses a sampling frequency of 2 Hz, acquiring a complete data sequence every t time, and the acquired data is in RTCM format; the acquired data is processed and calculated internally by the buoy. Data processing module: Wind speed and direction inversion model based on Physical Information Neural Networks (PINNs); Physical equations are introduced during the training process of the PINNs model to constrain the model output through inverse modeling. Output module: Visualized output of wind speed and direction predictions.

[0015] Compared with the prior art, the beneficial effects of this application are as follows: 1. Improved Prediction Accuracy and Enhanced Physical Consistency: This invention embeds the physical laws of wind and waves into a neural network, enabling more accurate prediction of wind speed and direction from wave spectrum data compared to existing traditional physical methods and purely data-driven methods. Tests show that the wind speed prediction error is less than 1.5 m / s, and the wind direction prediction error is less than 25 degrees, with an accuracy improvement of approximately 15% to 20% compared to existing technologies. Furthermore, the prediction results conform to the scientific principles of wind-wave interaction.

[0016] 2. Enhanced robustness under sparse data conditions: This invention utilizes physical constraints to optimize the prediction process. Compared with existing technologies that are easily limited by assumptions or overfitting when data is sparse or noisy, this invention can still generate reliable wind speed and direction prediction results even with limited observation data, significantly improving the robustness of the method.

[0017] 3. Achieving efficient computation and real-time prediction: Compared with traditional numerical models with high computational complexity and long processing time, the neural network model trained in this invention can complete wind speed and direction prediction in milliseconds, greatly improving computational efficiency and meeting the needs of real-time marine applications.

[0018] 4. Reduced dependence on external data: This invention can predict wind speed and direction using only wave spectrum data. Compared with the limitations of existing technologies that often require additional wind field observations, it reduces dependence on external data, lowers observation costs, and improves ease of application. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the physical information neural network model architecture; Figure 2 A comparison of the error distributions for wind speed prediction using physical information neural network methods and classical physical methods; Figure 3 A comparison of the error distributions for wind direction prediction using physical information neural network methods and classical physical methods; Figure 4 This is a flowchart of the application process. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] A method for GNSS buoy wind speed inversion based on physical information neural networks includes the following steps: First, the data acquisition and preprocessing module: (1) Hardware structure and function of GNSS buoy: The GNSS buoy device used in this invention consists of a high-precision GNSS receiver, sensors, and a wireless data transmission module. The buoy's built-in GNSS receiver can receive signals from multiple satellites in real time, achieving high-precision positioning and providing accurate timestamps for data synchronization. The buoy's data acquisition uses a sampling frequency of 2 Hz, collecting a complete data sequence every 30 minutes in RTCM format. The acquired data undergoes subsequent processing and calculations within the buoy, providing reliable raw data for subsequent wind speed inversion.

[0022] (2) Data preprocessing process: The buoy's internal microprocessor system was used to process GNSS data, obtaining the buoy's three-dimensional motion information in the NEU (northeast and vertical) direction. Furthermore, the frequency spectrum and Fourier coefficients of the waves can be obtained. and It can be calculated using the following formula: ; ; ; ; in, Represents energy components in different directions; It is the imaginary part of the cross spectrum, reflecting the phase difference between two signals and describing their phase asymmetry; The cross spectrum is the real part, reflecting the synchronicity of the two signals and describing the common changes in their phase and amplitude. The cross spectrum above can be obtained through Fourier transform.

[0023] Second, a wind speed and direction inversion model based on PINNs (Physics-Informed Neural Networks): (1) Model network input: The input data for this model includes wave spectrum data. , Fourier coefficients , , , Together, they constitute a multidimensional input parameter reflecting the nonlinear relationship between waves and wind speed.

[0024] (2) Neural network hierarchical structure.

[0025] Input layer: The number of neurons in the input layer corresponds to the dimension of the input features, ensuring that the multidimensional information of the wave spectrum is fully transmitted to subsequent layers;

[0026] The input layer receives wave spectrum data and Fourier coefficients as input features. It mainly includes: wave spectrum components. Representing waves at Frequency domain energy density distribution along the direction; Fourier coefficients , Characterizing wave signals at frequency The periodic decomposition characteristics at the location.

[0027] Hidden layer: The hidden layer processes the input data layer by layer through forward propagation, extracts the spatial distribution features and frequency domain decomposition characteristics of the wave spectrum, and maps them to a wind speed-related latent representation space.

[0028] The hidden layer consists of multiple fully connected neural network layers, used to capture the nonlinear relationship between wave spectrum data and wind speed. The specific structure is as follows: Number of layers: Contains 3 to 10 hidden layers, preferably 5 hidden layers, to balance the computational complexity and expressive power of the model.

[0029] Number of neurons: 50 neurons per layer to fully represent the nonlinear combination of input features.

[0030] Activation function: Each hidden layer employs a nonlinear activation function, such as the hyperbolic tangent (tanh) or the modified linear unit (ReLU), to enhance the model's ability to model the complex relationship between wave spectrum and wind speed.

[0031] Output layer: The output layer integrates the high-dimensional features extracted from the hidden layer to generate a joint prediction result for wind speed and wind direction.

[0032] The output layer is a single fully connected layer used to generate wind speed forecasts. Wind direction forecast The specific design is as follows: Output dimension: The output layer contains 2 neurons, each corresponding to a predicted wind direction value. (unit: ) and wind direction forecast (Unit: degrees, range from 0° to 360°).

[0033] Activation function: for wind speed prediction values A linear activation function is used to output continuous positive values; for wind direction prediction values... After selecting a linear activation function, the angle range is constrained through modulo operation.

[0034] Multiple neurons in the hidden layer capture the complex patterns and nonlinear relationships in the input data, while the output layer transforms these features into final wind speed and direction predictions through fully connected operations. This design allows the model to fully utilize all information in the input data, thereby providing accurate wind speed and direction predictions.

[0035] (3) Introduction of physical equations: This invention introduces physical equations into the training process of Physical Information Neural Networks (PINNs) to constrain the model output through inverse modeling, thereby ensuring the accuracy of wind speed predictions. Wind direction forecast This conforms to the physical laws governing the interaction between wind and waves.

[0036] The physical equations are introduced specifically through a loss function, and the specific methods and steps are as follows: a) Physical equation constraints and inverse modeling: This invention employs a reverse modeling method based on wave energy balance to infer wind speed and direction from wave spectrum data. It assumes the wave field is in a quasi-uniform and quasi-stationary state, and the source term balance equations are valid at all frequencies. and direction The components are approximately closed, that is: ; in: The wind input source term explicitly depends on the unknown wind speed. and wind direction The subscript 10 indicates a height of 10 meters; For wave breakup and dissipation source terms; It is a four-wave interaction source term, which is conservative and approximately zero in the overall energy balance; The directional wave spectrum is approximately reconstructed from observational data, including the input spectral components. and Fourier coefficients .

[0037] Directional wave spectrum It can usually be decomposed into a frequency spectrum. and directional distribution The product of: ; in: It is the vertical frequency spectrum ; It is a directional distribution function that satisfies the normalization condition. From Fourier coefficients reconstruction.

[0038] The Fourier expansion of the directional distribution is as follows: ; To simplify calculations and improve robustness, this invention ignores minor terms and further balances the overall source terms. By integrating over frequency and direction, we obtain: ; In addition, wind direction affects the overall motion stress. The wind direction alignment constraint is defined as: ; in: Let be the total kinetic stress vector exerted by the wind on the sea surface, with direction relative to . Consistent, including wave stress, viscous stress, and background stress; It is the momentum contribution that waves gain from the wind, which is related to frequency. and direction The double integral is used to sum the contributions of all wave components to the corresponding forces. This represents viscous stress, the stress that air exerts directly on the sea surface through viscous friction; The background stress represents the contribution of unresolved high-frequency shortwave to the total stress. It is gravitational acceleration; It is the wavenumber vector, and its direction is the same as the direction of wave propagation. For wave number scalars, under the deep-water assumption ; It is the wave angular frequency; The density ratio of air to water is denoted by c, typically around 0.0012; c represents the wave phase velocity, as shown in the deep-water dispersion relation. . Characterizing the dynamic changes in sea surface roughness driven by wind, among which It is the friction speed. It is the Charnock constant (typical value 0.01-0.03); Surface roughness length, and , The relationship can be represented by a logarithmic wind corridor, such as ,in is the von Kármán constant, with a typical value of 0.4.

[0039] To achieve a numerical solution to the above equations, this invention introduces the following source term approximation: a1. Wind input source item It can be expressed as a quasi-linear approximation as: ; ; ; when hour, ; , where is the Miles parameter, representing the growth rate parameter in the wind-wave interaction; ; and These are adjustable parameters used to optimize the model's description of wind and wave energy transfer.

[0040] a2. Dissipation source term , representing the wave energy dissipation rate, is mainly caused by wave breaking. Using the direction-dependent saturation approximation, it is expressed as: ; in: Indicates the wave in a given direction and wave number The saturation level on the wave surface measures whether the wave energy has reached the breaking point. The larger the wave, the more unstable it is and the more easily it breaks.

[0041] ; ; Represents the integral variable. The directional range of the dissipation effect; waves beyond this angle do not contribute to saturation. The parameter adjustment coefficient, representing the dissipation intensity, controls the dissipation intensity of the white wave breakage and is determined based on model calibration; The critical value of wave saturation, when When the waves break apart, dissipation occurs; otherwise, there is no dissipation.

[0042] b) Loss function design for PINNs: The physical equations are introduced into the training process through a loss function, namely, physical loss. It consists of two parts: Overall source term balancing constraints: ; Wind direction stress constraint: ; in The stress amplitude is given; the total physical loss can be expressed as: ; in: These are the weighting coefficients.

[0043] c) Introducing physical equations and training: The physical equations are introduced during the model training phase. Specifically, after each forward propagation, based on the predicted values... and Calculate physical loss and data loss Joint optimization, the total loss function is: ; ; The weights are adjustable.

[0044] Training employs optimization algorithms (such as the Adam optimizer) to iteratively update network parameters until the loss converges. The training process minimizes the total loss function. Optimize neural network parameters, including weights and biases, to ensure the model simultaneously satisfies data consistency and physical laws. The training steps are as follows: c1. Data Preparation and Preprocessing: Input data: Wave spectral components obtained from GNSS buoy observation data. , , and Fourier coefficients .

[0045] Target data: the corresponding actual wind speed (unit: ) and wind direction (Unit: °). This data represents the actual wind field observed by nearby buoys.

[0046] Preprocessing: The input data is normalized to ensure numerical stability; wind direction data is converted to... and To handle periodicity.

[0047] c2. Forward Propagation and Output Calculation: Input data is used to generate predicted values ​​through the forward propagation of a neural network. For each training sample, a predicted value is calculated. and .

[0048] c3. Backpropagation and parameter optimization: Calculate the total loss using an optimization algorithm (such as the Adam optimizer). Gradients of network parameters (weights and biases) .

[0049] Update network parameters: learning rate, e.g., 0.001; number of iterations, e.g., 10,000, or until... convergence.

[0050] It's not a simple mean squared error, but involves nonlinear physical equations, making analytical differentiation impossible. Therefore, the gradient of the physical loss is calculated through automatic differentiation, specifically a computational graph-based differentiation method that automatically calculates the gradient using the chain rule to ensure... The complex dependencies were propagated correctly.

[0051] c4. Model Validation and Adjustment: Evaluate prediction accuracy using a validation set independent of the training set; if accuracy is insufficient, adjust accordingly. Alternatively, re-train the network based on its structure (number of layers, number of neurons). Accuracy validation metrics include root mean square error, mean absolute error, mean error, and standard deviation.

[0052] Mean absolute error: ; Average error: ; Root mean square error: ; Standard deviation: ; in: This represents the difference between the predicted and actual wind speed / direction values, where M represents the sample size.

[0053] Third, wind speed and direction forecast: After constructing and training the physical information neural network model, this invention applies it to predict wind speed and direction from wave spectrum data. The specific application steps are as follows: (1) Input data: Acquire wave spectrum data from GNSS buoys of the scenario to be predicted, including spectral components. and Fourier coefficients .

[0054] (2) Data preprocessing: The collected wave spectrum data is normalized to maintain consistency with the training phase, ensuring that the numerical scale of the input features matches the model's requirements. For example, ... Zoom to interval.

[0055] (3) Forward prediction of the model: The preprocessed input data is fed into the trained PINNs model, and the predicted value is calculated through forward propagation: the input layer receives a 7-dimensional feature vector. Processed through hidden layers (preferably 3 layers, 50 neurons per layer, tanh activation function).

[0056] (4) Results output and application: Output predicted value and .

[0057] A GNSS buoy wind speed inversion system based on physical information neural network includes a data acquisition module, a data processing module, and an output module; Data acquisition module: The buoy's built-in GNSS receiver can receive signals from multiple satellites in real time for positioning and provide accurate timestamps for data synchronization; the buoy's data acquisition uses a sampling frequency of 2 Hz, acquiring a complete data sequence every t time, and the acquired data is in RTCM format; the acquired data is processed and calculated internally by the buoy. Data processing module: Wind speed and direction inversion model based on Physical Information Neural Networks (PINNs); Physical equations are introduced during the training process of PINNs to constrain the model output through inverse modeling. Output module: Visualized output of wind speed and direction predictions.

[0058] Figure 1 This is the structure of the physical information neural network of this invention. Green marks the input layer, purple marks the hidden layer, and red marks the output layer. The left side is the input layer, containing n circular nodes, the content of which includes... Each parameter is a function of frequency, and the dimension n of the input layer is 128×7; there are 5 hidden layers in the middle, labeled as... Each layer has 50 nodes. The right side is the output layer, containing 2 nodes, labeled as... and , respectively representing wind speed and wind direction The layers are connected using a fully connected approach. The right side also includes a loss function representing physical constraints and data errors.

[0059] Figure 2 The horizontal axis of the graph represents prediction error (unit: m / s), and the vertical axis represents probability density. The blue bars represent the error distribution of the Physical Information Network (PIN) method, while the red bars and curves represent the error distribution of the classical physical method. The graph shows that the error distribution of the PIN method is more concentrated near zero, exhibiting smaller bias and dispersion. Specifically, the AI ​​method has a mean error (Mean) of 0.11 m / s and a standard deviation (Std) of 1.24 m / s, as shown in the blue text box in the upper right corner of the graph. In contrast, the classical physical method has a more dispersed error distribution, with a mean error of -0.24 m / s and a standard deviation of 1.48 m / s. Both methods exhibit an approximately normal distribution, but the PIN method has a higher peak and a narrower tail, indicating that it has higher accuracy and stability in wind speed prediction.

[0060] Figure 3 The horizontal axis represents prediction error (unit: °), and the vertical axis represents probability density. The blue bars represent the error distribution of the Physical Information Network (PIN) method, while the red bars and curves represent the error distribution of the classical physical method. As can be seen from the figures, the error distribution of the PSN method is more concentrated near zero, with smaller bias and dispersion. Specifically, the mean error (Mean) of the PSN method is approximately 0.07°, and the standard deviation (Std) is approximately 21.08°. In contrast, the error distribution of the classical physical method is more dispersed, with a mean error of approximately 0.3° and a standard deviation of approximately 35.84°. Although the error distributions of both methods generally exhibit approximate normal distribution characteristics, the PSN method has a higher peak value and a narrower tail, indicating a significant advantage in wind direction prediction accuracy and stability.

[0061] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A GNSS buoy wind speed inversion method based on physical information neural network, characterized in that: The following steps are involved: S1. Data collection and preprocessing; S2. Construct a wind speed and direction inversion model based on the physical information neural network model PINNs; introduce physical equations into the training process of the physical information neural network PINNs to constrain the output of the model PINNs through reverse modeling; S3. Wind speed and direction forecast.

2. The GNSS buoy wind speed inversion method based on physical information neural network according to claim 1 is characterized in that: Step S1: Data preprocessing steps are as follows: Obtain GNSS data and obtain the three-dimensional motion information of the buoy in the NEU direction ( ), the frequency spectrum and Fourier coefficients of the wave and , calculated by the following formula: ； ； ； ； in, Represents the energy components in different directions; is the imaginary part of the cross spectrum; is the real part of the cross spectrum.

3. The GNSS buoy wind speed inversion method based on physical information neural network according to claim 1 is characterized in that: Step S2: the physical information neural network model PINNs includes an input layer, a hidden layer and an output layer; The input layer receives wave spectrum data and Fourier coefficients as input features, including wave spectrum components Respectively represent the waves in Frequency domain energy density distribution in direction; Fourier coefficients , Characterize the frequency of the wave signal The periodic decomposition characteristics at The hidden layer processes the input data layer by layer through forward propagation, extracts the spatial distribution characteristics and frequency domain decomposition characteristics of the wave spectrum, and maps them to the potential representation space related to wind speed; The output layer generates a joint prediction result of wind speed and wind direction by integrating the high-dimensional features extracted by the hidden layer. , using a linear activation function to output continuous positive values; for wind direction prediction values , after selecting the linear activation function, the angle range is constrained by modulo operation.

4. The GNSS buoy wind speed inversion method based on physical information neural network according to claim 1 is characterized in that: The steps for constraining and inverse modeling the physical equations are as follows: Based on the inverse modeling method of wave energy balance, the wind speed and direction are inferred from the wave spectrum data. Assuming that the wave field is in a quasi-uniform and quasi-stationary state, the source term balance equation is and direction The components are approximately closed, that is: ； Where: Wind speed at 10m height and wind direction ; Enter the source term for wind; is the wave breaking dissipation source term; is the four-wave interaction source term; Directional wave spectrum ; Directional Wave Spectrum Decomposition into frequency spectrum and direction distribution The product of: ； in: is the vertical frequency spectrum ; is the directional distribution function, satisfying the normalization condition , from the Fourier coefficients reconstruction; The Fourier expansion form of the directional distribution is: ； By integrating over frequency and direction, we obtain: ； Wind direction through the overall motion stress Align with wind direction constraint, defined as: ； ； ； in: is the total motion stress vector exerted by wind on the sea surface; is the momentum contribution that the wave gains from the wind; represents the viscous stress; represents the background stress, is the acceleration due to gravity; is the wave number vector, and its direction is consistent with the direction of wave propagation; is the wave number scalar, is the wave angular frequency; is the density ratio of air to water; c is the wave phase velocity; , characterizing the dynamic changes of sea surface roughness driven by wind, where is the friction speed, is the Charnock constant; Surface roughness length, is the von Karman constant.

5. The GNSS buoy wind speed inversion method based on physical information neural network according to claim 4 is characterized in that: Introduce source term approximation: a1. Wind input source term , using the quasi-linear approximation: ； ； ； when hour, ; ； ； and is an adjustable parameter; a2. Dissipative source term , represents the wave energy dissipation rate, using the direction-dependent saturation approximation, expressed as: ； in: Indicates that the wave is in a given direction and wave number Saturation on ； ； is the wave group velocity, represents the integral variable, The directional range of the dissipation effect, waves beyond this angle do not contribute to saturation; The parameter adjustment coefficient representing the dissipative intensity controls the dissipative intensity of white wave breaking and is determined based on the model calibration; The critical value of wave saturation is When , the waves begin to break and dissipation occurs, otherwise there is no dissipation.

6. The GNSS buoy wind speed inversion method based on physical information neural network according to claim 1 is characterized in that: Physical equations are introduced into the training process through the loss function. Including the overall source balance constraint and wind direction stress constraint, the calculation is as follows: Overall source term balance constraint: ； Wind stress constraint: ； in is the stress amplitude; The total physical loss is expressed as: 。 7. The GNSS buoy wind speed inversion method based on physical information neural network according to claim 5 is characterized in that: Introducing Physics Equation: After each forward propagation, based on the wind speed prediction value and wind direction forecast , calculate the physical loss and with data loss Joint optimization, the total loss function is: ； ； is an adjustable weight, and N is the number of samples used to calculate data loss.

8. The GNSS buoy wind speed inversion method based on physical information neural network according to claim 7 is characterized in that: The physical equations are introduced during the model training phase, and the total loss is calculated using an optimization algorithm. Gradients of network parameters , update the network parameters until convergence; Calculate wind speed deviation and wind direction deviation: u= - ; ; Actual wind speed value, Actual value of wind direction.

9. The GNSS buoy wind speed inversion method based on physical information neural network according to claim 1 is characterized in that: The data acquisition adopts a sampling frequency of 2 Hz, and a complete data sequence is collected every t time. The collected data format is RTCM format.

10. A GNSS buoy wind speed inversion system based on physical information neural network, using the GNSS buoy wind speed inversion method based on physical information neural network as claimed in any one of claims 1 to 9, characterized in that: It includes a data acquisition module, a data processing module and an output module; Data acquisition module: The buoy's built-in GNSS receiver can receive multiple satellite signals in real time, perform positioning, and provide accurate timestamps for data synchronization; the buoy's data acquisition uses a sampling frequency of 2 Hz, and a complete data sequence is collected every t time. The data collection format is RTCM format; the collected data is subsequently processed and calculated inside the buoy; Data processing module: wind speed and direction inversion model based on physical information neural network PINNs; physical equations are introduced in the training process of physical information neural network model PINNs to constrain model output through reverse modeling; Output module: Visual output of wind speed and direction forecasts.

Images