A sea typhoon profile generating method based on a downward cast typhoon sounding data and deep learning

Abstract

The application discloses a sea typhoon profile generation method based on a downward typhoon sounding data and deep learning, comprising: obtaining typhoon environmental parameters and a ground surface wind component of a sounding observation point at a preset reference height; preprocessing the typhoon environmental parameters and the ground surface wind component to generate a conditional input feature; determining a target generation height, and mapping the target generation height to a normalized interval to obtain a normalized height; inputting the conditional input feature and the normalized height into a trained deep learning model to output a radial wind component and a tangential wind component at the target generation height; and generating a sea typhoon wind profile corresponding to the target generation height based on the radial wind component and the tangential wind component. The method effectively solves the technical problems of the prior art, such as insufficient description of complex typhoon structure, insufficient use of discrete sounding data, weak continuity and generalization ability of generated profiles, and the like.

Description

Technical Field

[0001] This invention relates to the fields of marine meteorology, tropical cyclone boundary layer modeling, atmospheric sounding data processing, meteorological artificial intelligence, and marine engineering disaster prevention and mitigation technology. More specifically, it relates to a method for generating marine typhoon profiles based on drop-in typhoon sounding data and deep learning. Background Technology

[0002] Near-surface wind fields of typhoons at sea provide fundamental information for storm surge numerical forecasting, wind-resistant design of offshore platforms and offshore wind power structures, risk assessment of ship routes, determination of construction windows for nearshore engineering projects, and coastal disaster prevention. Unlike simple single-point wind speeds, wind profiles simultaneously reflect radial wind, tangential wind, total wind speed, and inflow angle at different altitudes, serving as a crucial bridge connecting typhoon dynamics with engineering applications.

[0003] Current methods for representing near-surface wind profiles of typhoons at sea typically rely on logarithmic law models, power law models, Ekman spiral-like analytical models, or empirical regression models. While these methods are simple in form and easy to interpret, they often struggle to accurately describe the increased curvature, variations in inflow angle with height, local bends, and non-monotonic structures under strong typhoon conditions within the 60m to 200m transition layer. On the other hand, purely statistical or data-driven methods, while offering greater flexibility, are prone to non-physical oscillations, unstable tail extrapolation, and inconsistencies with 10m observation points when physical constraints are lacking.

[0004] Furthermore, downcast typhoon sounding data is characterized by highly irregular sampling, large differences in storm scale, periodic wind direction, and different sign constraints in the Northern and Southern Hemispheres. This makes it difficult for traditional discrete interpolation or fixed function family methods to simultaneously achieve continuous representation capability, physical plausibility, and cross-typhoon generalization capability. Therefore, it is necessary to propose an algorithm for generating continuous wind profiles of offshore typhoons that can both absorb complex morphological information from observational data and suppress incomprehensible information through physical priors.

[0005] Currently, existing technologies for generating continuous wind profiles of typhoons at sea have the following problems:

[0006] (1) Acquiring marine typhoon observation data is costly and risky. While conventional surface stations, buoys, ship observations, and satellite remote sensing can provide some environmental information, it is difficult to directly and accurately obtain continuous vertical structure information of the typhoon core region. Downward-projected typhoon sounding data can provide detailed profiles of wind speed, air pressure, temperature, humidity, and other elements as a function of altitude, and is an important data source for studying the typhoon core structure. However, this type of data has a limited number of samplings, discrete deployment locations, and uneven coverage. The amount of data varies greatly between different typhoon cases, resulting in significant sparsity, non-uniformity, and incompleteness of the observation samples. Existing methods can usually only analyze a single measured profile, making it difficult to generate a continuous typhoon profile field covering different radii, quadrants, intensities, and altitudes under conditions of scarce observations.

[0007] (2) Many existing methods rely on empirical statistical models, parametric typhoon models, or simple interpolation extrapolation to estimate typhoon structure. These methods are usually based on strong prior assumptions, such as that typhoons are approximately axisymmetric in the horizontal or vertical direction, have smooth structural changes, and have fixed functional relationships between different elements. However, real typhoons, especially under conditions of rapid intensification, eyewall replacement, asymmetric development, and strong environmental wind shear, often exhibit significant spatial asymmetry and high dependence. Traditional parametric methods are unable to accurately characterize the complex coupling relationships between the typhoon boundary layer, eyewall region, outer rainband, and upper outflow region, leading to deviations in the generated results in key areas.

[0008] (3) Existing technologies have limited ways of utilizing dropsonde data, usually focusing on sample statistics, case analysis, or local structure fitting, and cannot fully explore the high-dimensional nonlinear mapping relationships hidden in historical typhoon sounding samples. In particular, in the problem of mapping "relative position of typhoon center, intensity parameters, and structural characteristic parameters" to "complete vertical profile", existing technologies lack a general method that can automatically learn patterns from a large number of historical dropsonde samples and generate high-precision profiles for unobserved locations. Therefore, when the number of observation points is limited or the observation location is not ideal, existing methods often cannot effectively recover the true profile characteristics of the target area.

[0009] (4) Existing technologies also have shortcomings in terms of the consistency and applicability of the generated results. On the one hand, there is often a lack of unified modeling between different altitude layers and different physical quantities, which can easily lead to problems such as vertical discontinuity, excessive local mutations, or poor matching between multiple elements. On the other hand, some methods are only applicable to specific intensity levels, specific sea areas, or a small number of individual typhoons, and their generalization ability is insufficient, making it difficult to directly apply them to scenarios such as operational analysis, numerical model initialization, marine engineering risk assessment, and flight mission auxiliary decision-making.

[0010] Therefore, how to solve the technical problems of existing methods, such as insufficient characterization of complex typhoon structures, inadequate utilization of discrete radiosonde data, and weak continuity and generalization ability of generated profiles, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0011] In view of the above problems, the present invention proposes a method for generating marine typhoon profiles based on downdrop typhoon sounding data and deep learning to overcome or at least partially solve the above problems.

[0012] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a method for generating marine typhoon profiles based on downdrop typhoon sounding data and deep learning, comprising: Acquire typhoon environmental parameters and surface wind components at preset reference heights from radiosonde observation points; The typhoon environmental parameters and the surface wind components are preprocessed to generate conditional input features; Determine the target generation height and map the target generation height to a normalized interval to obtain the normalized height; The conditional input features and the normalized height are input into the trained deep learning model, which outputs the radial wind component and tangential wind component at the target generation height. Based on the radial wind component and the tangential wind component, a typhoon wind profile corresponding to the target generation height is generated.

[0013] Furthermore, the specific input features include: dimensionless radius, azimuth sine, azimuth cosine, maximum wind speed, maximum wind speed radius, absolute value of Coriolis parameter, sign of Coriolis parameter, latitude of typhoon center, longitude of typhoon center, radial wind at a preset reference height, and tangential wind at a preset reference height.

[0014] Furthermore, the deep learning model is constructed using an implicit profile network architecture, which includes: Conditional branching is used to map the conditional input features into a conditional vector. The height branch is used to map the normalized height to a high-dimensional latent space to obtain a high-dimensional latent representation; A shared backbone is used to receive and fuse the conditional vector and the height latent representation to generate a shared semantic representation of the continuous profile; Dual output heads are used to output radial wind increment and tangential wind increment respectively according to the shared semantic representation; A differential anchoring layer is used to perform differential anchoring processing based on the radial wind increment and tangential wind increment to determine the radial wind component and tangential wind component.

[0015] Furthermore, the height branch employs a hybrid spectral basis function encoder, which fuses position encoding, Fourier features, and radial basis functions to map the normalized height to a high-dimensional latent space.

[0016] Furthermore, the width of the radial basis function is positively correlated with the dimensionless radius.

[0017] Furthermore, the shared backbone includes multiple cascaded residual blocks, each residual block employing a characteristic linear modulation layer. The modulation parameters of the characteristic linear modulation layer are generated by the condition vector and are used to dynamically modulate the shared semantic representation in the shared backbone layer by layer.

[0018] Furthermore, the residual blocks of odd-numbered layers in the shared backbone also introduce gated jump connections to fuse the initial output of the height branch with the output of the current residual block.

[0019] Furthermore, the dual output head includes a low-rank coupling term, which constructs a cross-component physical correlation between the radial wind increment and the tangential wind increment through a low-rank matrix.

[0020] Furthermore, the total loss function of the deep learning model Represented as:

[0021]

[0022]

[0023]

[0024]

[0025]

[0026]

[0027]

[0028]

[0029] in, For data fitting loss; For curvature regularization loss; For tangential wind sign consistency loss; For the inflow angle prior loss; For logarithmic soft prior loss; This is due to the loss of modulation parameter stability; This is due to the low-rank coupling stability loss; For slope regularization loss; , , , , , , , These are the weighting coefficients for the corresponding loss terms; This represents the total number of sampling points; For interval The total number of valid sampling points involved in the calculation of the prior loss of the inflow angle; For interval The total number of valid sampling points involved in the calculation of logarithmic-law soft prior loss; For the first The height of each sampling point; For the first The weight of each sampling point; For robust penalty functions; Indicates a threshold parameter Robust penalty function; Indicates a threshold parameter Robust penalty function; and They are respectively height Predicted and observed values ​​of radial wind at the location; and They are respectively height Predicted and observed values ​​of tangential wind; For height The predicted wind vector at that location, and ; For height Curvature weight at the location; Coriolis parameter symbol; It is a second-order difference operator along the height direction; This represents the change in the inflow angle between adjacent heights; For height The reference wind speed modulus value is obtained based on the logarithmic law of the near-surface layer. This represents the total number of FiLM modulation layers; For the first One FiLM modulation layer; For the first A vector of scaling modulation parameters for each FiLM modulation layer; It is a low-rank coupling matrix; This represents the Frobenius norm.

[0030] Furthermore, it also includes: obtaining a lightweight model by reducing the number of Fourier features in the deep learning model, reducing the number of residual blocks, or reducing the rank of low-rank coupling terms; and using the lightweight model for inference deployment on an edge computing platform or in a real-time wind field generation scenario.

[0031] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method for generating marine typhoon profiles based on downdrop typhoon radiosonde data and deep learning, which has the following beneficial effects: This invention makes full use of historical downdrop typhoon radiosonde data and combines deep learning to model the multi-element, multi-height layer, and nonlinear structural relationships of typhoons. It also achieves a method for generating high-precision vertical profiles of typhoons at sea under sparse observation conditions. This effectively solves the technical problems of existing methods, such as insufficient characterization of complex typhoon structures, inadequate utilization of discrete radiosonde data, and weak continuity and generalization ability of generated profiles. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0033] Figure 1 This is a schematic diagram of the method for generating marine typhoon profiles based on drop-in typhoon sounding data and deep learning provided in an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the variation of the sampling density of the lower profile of different typhoon levels with altitude, provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of the relative position distribution of HRD drop-in radiosonde samples provided in an embodiment of the present invention; Figure 4 This is a scatter plot illustrating the overall consistency of the model provided in this embodiment of the invention; Figure 5 This is a schematic diagram of the consistency and error distribution of the inflow angle prediction provided in an embodiment of the present invention; Figure 6 Radial wind provided in the embodiments of the present invention Cross-sectional comparison diagram; Figure 7 Tangential wind provided in the embodiments of the present invention Cross-sectional comparison diagram; Figure 8 This is a comparative schematic diagram of the total wind speed profile provided in an embodiment of the present invention; Figure 9This is a schematic diagram comparing the average profile and interquartile range (IQR) provided in an embodiment of the present invention; Figure 10 This is a schematic diagram illustrating the variation of radial wind RMSE with altitude under different typhoon levels provided in this embodiment of the invention; Figure 11 This is a schematic diagram illustrating the variation of tangential wind RMSE with altitude under different typhoon levels provided in this embodiment of the invention; Figure 12 This is a schematic diagram illustrating the variation of total wind speed (RMSE) with altitude under different typhoon levels provided in this embodiment of the invention. Detailed Implementation

[0034] 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.

[0035] This invention discloses a method for generating marine typhoon profiles based on drop-in typhoon radiosonde data and deep learning, such as... Figure 1 As shown, it includes the following steps: S1. Obtain typhoon environmental parameters and surface wind components at preset reference heights from radiosonde observation points; S2. Preprocess the typhoon environmental parameters and surface wind components to generate conditional input features; S3. Determine the target generation height and map the target generation height to the normalized interval to obtain the normalized height; S4. Input the conditional input features and normalized height into the trained deep learning model, and output the radial wind component and tangential wind component at the target generation height; S5. Based on the radial and tangential wind components, generate the offshore typhoon wind profile corresponding to the target generation height.

[0036] This invention provides a method for generating marine typhoon profiles based on downdrop typhoon radiosonde data and deep learning. It makes full use of historical downdrop typhoon radiosonde data and combines deep learning to model the multi-element, multi-height, and nonlinear structural relationships of typhoons. Under the condition of sparse observation, it achieves high-precision generation of marine typhoon vertical profiles, thereby solving the technical problems of existing methods such as insufficient characterization of complex typhoon structures, insufficient utilization of discrete radiosonde data, and weak continuity and generalization ability of generated profiles.

[0037] This method can be used in the following scenarios: 1) Reconstruction, interpolation, and completion of continuous wind profiles of offshore typhoons in the near-surface layer from 10 m to 300 m; 2) Construction of input fields for storm surge, ocean waves, air-sea flux, and boundary layer diagnostic models; 3) Load assessment of offshore wind turbines, offshore platforms, offshore bridges, floating equipment, and other structures under wind conditions at different heights; 4) Optimization of ship routes, port closure decisions, and safety early warning for offshore construction under typhoon conditions; 5) Construction of historical typhoon sample databases, generation of digital twin wind fields, and research on engineering risks in extreme weather; 6) Used as a core module for generating continuous vertical wind fields when jointly modeling with radar wind profilers, microwave scatterometers, buoys, and numerical weather prediction results.

[0038] It should be noted that the labels S1-S5 above are only for ease of explanation and do not limit the execution order of the steps. The following is a detailed explanation of steps S1 to S5.

[0039] In step S1 above, typhoon environmental parameters and surface wind components at a preset reference height (e.g., 10m) are obtained; specifically: This invention uses typhoon / hurricane dropsonde data acquired by the HRD (Hurricane Research Division) of NOAA / AOML between 1996 and 2022 as the training and validation data source. The original data package contains 14,358 records.

[0040] To ensure consistent near-surface anchoring, only east-west winds at locations that simultaneously include a preset reference height (e.g., 10m) are retained. North-south winds on the ground The sample was processed; records missing the 10 m wind component, exhibiting significant gross errors, or failing to meet consistency check requirements were discarded. After quality control, 8,909 valid samples were obtained, covering 107 different tropical cyclone cases.

[0041] The original variables in the sounding data used in this embodiment of the invention are shown in Table 1.

[0042] Table 1: Variable Table of Original Data

[0043] In step S2 above, based on the principles of scale equivariance, rotation equivariance, intensity modulation, and surface anchoring consistency, the typhoon environmental parameters and surface wind components are preprocessed to generate conditional input features, as shown in Table 2. These conditional input features specifically include: dimensionless radius, azimuth sine, azimuth cosine, maximum wind speed, maximum wind speed radius, absolute value of Coriolis parameter, sign of Coriolis parameter, latitude of typhoon center, longitude of typhoon center, radial wind at preset reference height, and tangential wind at preset reference height. Table 2: Conditional Input Features

[0044] Among them, dimensionless radius is used With the radius of maximum wind speed Combining these methods allows for feature sharing among typhoons of different scales; using azimuth sine waves With azimuth cosine Replace direct input of azimuth angle This avoids the discontinuity problem caused by angular periodicity; using the absolute value of the Coriolis parameter With Coriolis parameter symbol Decomposing the Coriolis parameters allows the model to simultaneously perceive the differences in absolute intensity and sign between the Northern and Southern Hemispheres.

[0045] (1) The above dimensionless radius azimuth sine azimuth cosine The steps to obtain it are as follows: Given the latitude and longitude of the typhoon center Latitude and longitude of the observation point Converted to radians, i.e., the latitude and longitude of the typhoon center. Latitude and longitude of the observation point Under the local plane approximation, calculate the relative displacement:

[0046]

[0047] in, This represents the meridional displacement of the observation point relative to the typhoon center. This represents the zonal displacement of the observation point relative to the typhoon center. The radius of the Earth; Further calculations were made of the radial distance of the observation point relative to the typhoon center. and azimuth :

[0048]

[0049] radial distance Divide by the radius of the typhoon's maximum wind speed The dimensionless radius is obtained. :

[0050] opposite angle Calculate the sine and cosine values ​​to obtain the azimuth sine. azimuth cosine .

[0051] (2) The steps for obtaining the radial wind and tangential wind at the preset reference height are as follows: To achieve rotational isomorphic processing, the wind component in the geographic coordinate system is represented by the following formula. Radial wind transformed to cylindrical coordinates With tangential wind :

[0052]

[0053] Since the preset reference height is set to 10m in this embodiment, it is similarly possible to convert the 10m surface wind into the corresponding 10m radial wind. and 10m tangential wind .

[0054] (3) Absolute values ​​of the Coriolis parameters mentioned above With Coriolis parameter symbol The Coriolis parameters are derived from their decomposition. The steps for obtaining the Coriolis parameters are as follows:

[0055] in This is the Earth's angular velocity. For the approximate gradient wind equilibrium near the top of the boundary layer, it can be expressed as:

[0056] in It is air density. It's pressure. It is the radial distance. It is a tangential wind. These are Coriolis parameters. Deviations from this approximate equilibrium in the frictional boundary layer lead to radial inflow and varying inflow angles at different heights, thus triggering the physical constraints used in the training objective. Therefore, it can be seen that the near-surface wind field, under the combined effects of turbulent friction, vertical shear, pressure gradient force, and Coriolis force, will produce highly correlated radial inflow and inflow angle variations. This invention incorporates this physical understanding into subsequent model design and loss function constraints.

[0057] In step S3 above, the target generation height is determined and mapped to a normalized interval to obtain the normalized height; Specifically, this embodiment of the invention focuses on a continuous wind profile within the range of 10m to 300m. To facilitate network learning, the target generation height z is mapped to a normalized interval to obtain the normalized height. , is represented as:

[0058] This normalization process enables the model to learn the continuous variation patterns of different height layers on a uniform scale, while providing a uniform independent variable for implicit function expression.

[0059] In step S4 above, the conditional input features and normalized height are input into the trained deep learning model, and the radial wind component and tangential wind component at the target generation height are output. In this embodiment of the invention, the continuous mapping function to be learned can be expressed as:

[0060] in, This represents the standardized conditional feature vector; This represents the normalized height; the output is the corresponding height. Predicted radial wind With tangential wind .

[0061] In this embodiment of the invention, the deep learning model is constructed using an implicit profile network architecture, which includes: (1) Conditional branch, used to map conditional input features into conditional vectors ; (2) Height branch, used to normalize the height Mapping to a high-dimensional latent space yields a high-level latent representation; Specifically, this height branch employs a hybrid spectral basis function encoder, fusing position encoding, Fourier features, and radial basis functions to map the normalized height to a high-dimensional latent space. Its form is:

[0062] in, For normalized height, This represents the total number of Fourier features; For Fourier feature index; This represents the total number of radial basis functions; Radial basis function index; The first uniformly distributed in the interval [0,1] Radial basis function centers The width of the radial basis function is related to the dimensionless radius.

[0063] The width of the radial basis function is written as:

[0064] in, Based on the width, To and The relevant adjustment coefficient, , These are the lower and upper limits of the width, respectively. That is, the width of the radial basis function is set to be positively correlated with the dimensionless radius: in the typhoon's core region, When smaller, Smaller size is advantageous for depicting abrupt changes in near-surface strata; in the outer region, When larger A relatively larger value helps avoid overfitting to weak perturbations.

[0065] (3) Shared backbone, used to receive and merge condition vectors and the initial output of the height branch Generate a shared semantic representation of continuous profiles. ; Specifically, the shared backbone comprises multiple cascaded residual blocks, each employing a Feature-wise Linear Modulation (FiLM) layer. The modulation parameters of this FiLM layer are determined by a conditional vector. Generation is used to dynamically modulate the feature representations in the shared backbone layer by layer to control the transmission and evolution of wind profile morphology under different typhoon environmental conditions. For the first... For layer residual blocks, their basic form is:

[0066] in, For the first Output of the layer residual block; For the first Output of the layer residual block; and For the first The linear mapping weights corresponding to the layers; For layer normalization; For activation functions; For the first A vector of scaling modulation parameters for each FiLM modulation layer; For the first A vector of translation modulation parameters for each FiLM modulation layer; This represents a regular expression combining random depth and dropout.

[0067] Furthermore, to enhance information flow across layers, gated jump connections are introduced into the residual blocks of odd-numbered layers in the shared backbone to fuse the initial output of the high-order branch with the output of the current residual block; represented as:

[0068] in, For the first The gating parameter vector of an odd number of residual blocks; For the Sigmoid function; It is a linear projection matrix; The initial output for the high-order branch; (4) Dual output heads, used for sharing semantic representation Unanchored radial wind increments were obtained through two lightweight MLP heads. and tangential wind increment And by introducing the necessary cross-component physical correlation through a low-rank coupling term, it is expressed as:

[0069] in, Radial wind increment The weights; For tangential wind increment The weights; For shared semantic representation; This represents the feature terms after applying a nonlinear mapping to the shared semantic representation; Representation matrix The rank of a low-rank matrix. By employing low-rank and small-amplitude regularization, the model can express cross-component physical relationships such as the inflow angle that vary with height, while retaining the advantages of component decoupling.

[0070] (5) Differential anchoring layer, used for radial wind increment and tangential wind increment Differential anchoring is performed to ensure that the model output is within the specified range. The radial and tangential wind components are strictly equal to the input preset reference height. For example, when the preset reference height is set to 10m, the radial wind component and the tangential wind component are expressed as follows:

[0071] in, This refers to either the radial wind component or the tangential wind component. 10m radial wind or 10m tangential wind; In order to achieve normalization Radial wind increment at the location or tangential wind increment ; In order to achieve normalization Radial wind increment at the location or tangential wind increment ; The above design makes , Strictly established. During training... Explicit sampling is performed on the anchor point and its neighboring points, and the FiLM scaling is slightly attenuated near the anchor point. At the same time, weight norm constraints and small amplitude Tanh constraints are applied to the output layer to reduce the impact of high-frequency oscillations on the stability of the anchor point.

[0072] Furthermore, in this embodiment of the invention, to balance data fitting ability and physical rationality, the total loss function of the deep learning model during training is designed as a weighted combination of multiple losses, specifically including: (1) Data fitting loss : The data fitting term employs a weighted Huber loss to reduce sensitivity to outliers and assigns higher weights to near-surface and high-curvature layers:

[0073] in, This represents the total number of sampling points; For the first The height of each sampling point; For the first The weight of each sampling point; For robust penalty functions; and They are respectively height Predicted and observed values ​​of radial wind at the location; and They are respectively height Predicted and observed values ​​of tangential wind; (2) Curvature regularization loss : To suppress non-physical high-frequency oscillations, a curvature regularity that varies with height is introduced:

[0074] in, For height Curvature weight at the location; It is a second-order difference operator along the height direction; Indicates a threshold parameter Robust penalty function; (3) Tangential wind sign consistency loss : Based on the hemispherical wind field structure characteristics, further constraints on tangential wind sign consistency are applied:

[0075] in, Coriolis parameter symbol; (4) Prior loss : Using the physical prior knowledge that the inflow angle varies with altitude as a constraint, a monotonic constraint is applied to the change of the inflow angle between adjacent altitudes:

[0076] Among them, among them, Representing an interval The total number of valid sampling points involved in the calculation of the prior loss of the inflow angle; This represents the change in the inflow angle between adjacent heights; (5) Soft prior loss : The near-surface wind speed modulus approximately follows a logarithmic law; therefore, a soft prior is introduced in the range of 10 m to 40 m:

[0077]

[0078] in, For height The predicted wind vector at that location, and ; This represents the total number of valid sampling points within the interval [10, 40]m that participate in the calculation of the logarithmic soft prior loss. Indicates a threshold parameter Robust penalty function; Indicates height The reference wind speed modulus value is obtained based on the logarithmic law of the near-surface layer. The friction speed; Kármán's constant; Length represents the surface roughness. (6) Modulation parameter stability loss :

[0079] in, Indicates the total number of FiLM modulation layers; Indicates the first One FiLM modulation layer; For the first A vector of scaling modulation parameters for each FiLM modulation layer; (7) Low-rank coupling stability loss :

[0080] in, It is a low-rank coupling matrix; Denotes the Frobenius norm; (8) Slope regularity loss :

[0081] in, For height Predicted wind vector at the location; Finally, the total loss function Represented as:

[0082] in, For data fitting loss; For curvature regularization loss; For tangential wind sign consistency loss; For the inflow angle prior loss; For logarithmic soft prior loss; This is due to the loss of modulation parameter stability; This is due to the low-rank coupling stability loss; For slope regularization loss; , , , , , , , These are the weighting coefficients for the corresponding loss terms; Based on a pre-trained deep learning model, during actual inference, for any... By sampling, continuous radial and tangential winds at any height from 10m to 300m can be output, thus forming a continuous wind profile of a typhoon at sea.

[0083] In step S5 above, the wind profile of the typhoon at sea corresponding to the target generation height is generated based on the radial wind component and the tangential wind component.

[0084] Figure 2 The results show that after quality control, the samples have a relatively stable sampling density at different typhoon levels and different altitudes. Figure 3 The sample data shows that it covers the core region, the vicinity of the eyewall, and the outer rainband within the relative spatial location, indicating that the input data of this invention can support wind profile modeling at different spatial locations and scales. Figure (a) shows the representation in an adaptive coordinate system, and Figure (b) shows the representation in a fixed coordinate system. Figure 4 and Figure 5 This is used to illustrate the overall consistency of the output results and the stability of the direction variable in this invention. Among them, Figure 4 This indicates that the predicted values ​​and observed values ​​are in... , The total wind speed and wind direction are generally distributed along the 1:1 reference line; Figure 5This indicates that the inflow angle error is mainly concentrated around 0°, demonstrating that the present invention can reconstruct the boundary layer orientation structure relatively stably.

[0085] Figure 6 , Figure 7 and Figure 8 The paper presents comparative results of radial wind, tangential wind, and total wind speed profiles with height in several representative cases. As can be seen from the figure, the continuous profile generated by this invention can follow the slope, amplitude, and inflection point position of the observed curve well, indicating that this invention can not only perform point value fitting but also has the ability to preserve the profile shape.

[0086] Figure 9 Furthermore, the present invention can better match the statistical distribution of the observed samples in both the average profile and the interquartile range, indicating that the method can not only generate individual profiles, but also has good overall statistical consistency.

[0087] Figure 10 , Figure 11 and Figure 12 The results show that the RMSE (root mean square error) varies with altitude and typhoon classification at different intensity levels. The overall error is lower for weak typhoons, while the error increases at higher altitudes for strong typhoons, but the overall trend is smooth without significant uncontrolled oscillations. This demonstrates that the present invention still possesses good generalization ability and interpretable error evolution characteristics under extreme strong wind conditions.

[0088] In summary, the method for generating marine typhoon profiles based on downdrop typhoon radiosonde data and deep learning provided by this invention addresses several issues: first, it solves the problem that traditional analytical functions are insufficient to describe the complex multi-scale wind profiles of marine typhoons in the near-surface layer; second, it addresses the problems of pure data-driven methods lacking physical constraints and being prone to non-physical oscillations and extrapolation instability; third, it addresses the problem that existing methods cannot strictly satisfy the consistency of 10m observation anchor points; and fourth, it resolves the contradiction between the need for relatively independent modeling capabilities and necessary coupled expression capabilities for radial and tangential winds.

[0089] Compared with the prior art, the present invention has at least the following advantages: (1) Strong continuous representation capability. This invention directly learns the continuous mapping relationship from height ζ to wind component, and can sample any height in the range of 10m to 300m without being limited by fixed layer discrete points.

[0090] (2) Strictly anchored at 10 m for observation. This invention ensures that the output is within the specified range by using structural anchoring rather than relying solely on loss penalty. The wind speed is strictly equal to 10 m, thus providing a stable benchmark for engineering applications.

[0091] (3) It can characterize complex transition layers. The present invention uses a hybrid spectral basis function composed of position encoding, Fourier features and adaptive RBF, which can better describe the curvature enhancement, inflow angle change and local inflection in the 60 m to 200 m transition layer.

[0092] (4) Better physical consistency. This invention significantly suppresses non-physical oscillations and directional jumps through mechanisms such as hemispherical consistency, monotonic prior of inflow angle, logarithmic soft constraint of near-surface layer, and curvature regularization.

[0093] (5) Balancing expression and explanation. This invention utilizes conditional vectors. Layered control of wind profile is achieved using FiLM modulation parameters, while utilizing a low-rank coupling matrix. It retains only the necessary cross-component couplings, thus combining high expressiveness with good interpretability.

[0094] (6) Generalization stability. According to the test results in the original paper, the model's coefficients of determination R² for radial wind, tangential wind and total wind speed reached 0.956, 0.967 and 0.987 respectively, indicating that the present invention has high fitting consistency on independent test samples.

[0095] (7) The model remains applicable under strong typhoon conditions. Although the sample error for high-intensity typhoons has increased, the statistical results of classification error and the changes in RMSE with altitude show that the model can still maintain a high correlation and controllable error level under strong typhoon conditions of C3 to C4+.

[0096] Further quantitative evidence can be obtained by using the statistical analysis of intensity errors, as shown in Tables 3 and 4.

[0097] Table 3: Statistics of Radial Wind Prediction Errors under Different Intensity Categories

[0098] Table 4: Statistics of Shear Wind Prediction Errors under Different Intensity Categories

[0099] Next, we will illustrate the above-mentioned method for generating marine typhoon profiles based on downdrop typhoon sounding data and deep learning provided by the present invention through several specific embodiments: Example 1: Standard Full-Function Example: This embodiment adopts a complete conditional branching-high-level branching-shared trunking-dual-headed output-low-rank coupling-anchored subtraction structure, and uses weighted Huber loss, curvature regularization, hemispherical consistency, inflow angle prior, near-surface logarithmic law soft constraint, FiLM stability term and coupling matrix regularization term in combination.

[0100] This embodiment is suitable for high-precision research and offline reconstruction scenarios. Its advantages are that it has the strongest expressive power and the best adaptability to complex transition layers and extreme cases.

[0101] Example 2: Lightweight Deployment Example: This embodiment retains core structures such as anchored subtraction, conditional modulation, and dual-head output, but appropriately reduces the number of Fourier basis functions, the number of RBF centers, and the number of residual layers, and constrains the low-rank coupling matrix to... Or further reduce its amplitude.

[0102] This embodiment reduces the number of parameters and inference latency while ensuring strict consistency of 10 m anchor points and unchanged main physical constraints, making it more suitable for business deployment, edge computing platforms, or scenarios where real-time wind fields are generated rapidly.

[0103] Example 3: Extended Example of Multi-Source Observation: This embodiment, based on the original drop-sonde data, additionally introduces environmental features from radar wind profilers, marine buoys, microwave scatterometers, reanalysis data, or numerical weather prediction products as additional inputs for conditional branches.

[0104] The difference between this embodiment and Embodiment 1 is that the input condition vector x is more complex, which can improve robustness when the sample is sparse; however, its continuous wind profile is still output by the same type of implicit profile network, and the 10 m structural anchoring constraint is still retained.

[0105] Example 4: Missing Test Completion and Inversion Example: This embodiment is designed for scenarios where only limited information such as 10-meter observation wind, typhoon center location, maximum wind speed radius, and maximum wind speed is available. The system directly generates continuous wind profiles within a range of 10 to 300 meters using a pre-trained network, which can be used for rapid completion of missing height layers, playback of historical cases, or engineering safety assessments.

[0106] Compared with Example 1, this embodiment does not emphasize offline retraining, but emphasizes direct invocation and rapid generation during the inference stage; its core difference lies in the application purpose focusing on data inversion and profile completion.

[0107] The first embodiment emphasizes complete accuracy and full physical constraints; the second embodiment emphasizes computational efficiency and real-time deployment; the third embodiment emphasizes multi-source data fusion capabilities; and the fourth embodiment emphasizes missing data completion and rapid inversion capabilities. Although the above embodiments differ in input, network scale, or application scenarios, they all share the core inventive concept of "continuous height implicit modeling based on drop-in typhoon sounding data and environmental conditions, and ensuring near-surface consistency through 10 m structural anchoring," and should all fall within the protection scope of this invention.

[0108] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0109] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those 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 invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for generating marine typhoon profiles based on downdrop typhoon sounding data and deep learning, characterized in that, include: Acquire typhoon environmental parameters and surface wind components at preset reference heights from radiosonde observation points; The typhoon environmental parameters and the surface wind components are preprocessed to generate conditional input features; Determine the target generation height and map the target generation height to a normalized interval to obtain the normalized height; The conditional input features and the normalized height are input into the trained deep learning model, which outputs the radial wind component and tangential wind component at the target generation height. Based on the radial wind component and the tangential wind component, a typhoon wind profile corresponding to the target generation height is generated at sea. The deep learning model is constructed using an implicit profile network architecture, which includes: Conditional branching is used to map the conditional input features into a conditional vector. A height branch is used to map the normalized height to a high-dimensional latent space to obtain a height latent representation. This height branch employs a hybrid spectral basis function encoder, which fuses positional encoding, Fourier features, and radial basis functions to map the normalized height to the high-dimensional latent space. Its form is as follows: in, Normalized height; This represents the total number of Fourier features; For Fourier feature index; This represents the total number of radial basis functions; Radial basis function index; The first uniformly distributed in the interval [0,1] One radial basis function center; The width of the radial basis function is related to the dimensionless radius; the width of the radial basis function is positively correlated with the dimensionless radius. A shared backbone is used to receive and fuse the conditional vector and the high-level latent representation to generate a shared semantic representation of a continuous profile; the shared backbone includes multiple cascaded residual blocks, each residual block employing a characteristic linear modulation layer, the modulation parameters of which are generated by the conditional vector; Dual output heads are used to output radial wind increment and tangential wind increment respectively according to the shared semantic representation; A differential anchoring layer is used to perform differential anchoring processing based on the radial wind increment and tangential wind increment to determine the radial wind component and the tangential wind component; the radial wind component and the tangential wind component are represented as follows: in, This refers to either the radial wind component or the tangential wind component. 10m radial wind or 10m tangential wind; In order to achieve normalization Radial wind increment at the location or tangential wind increment ; In order to achieve normalization Radial wind increment at the location or tangential wind increment .

2. The method for generating marine typhoon profiles based on drop-in typhoon radiosonde data and deep learning as described in claim 1, characterized in that, The specific input features include: dimensionless radius, azimuth sine, azimuth cosine, maximum wind speed, maximum wind speed radius, absolute value of Coriolis parameter, sign of Coriolis parameter, latitude of typhoon center, longitude of typhoon center, radial wind at preset reference height, and tangential wind at preset reference height.

3. The method for generating marine typhoon profiles based on drop-in typhoon radiosonde data and deep learning as described in claim 1, characterized in that, The shared backbone comprises multiple cascaded residual blocks, each of which employs a characteristic linear modulation layer. The modulation parameters of the characteristic linear modulation layer are generated by the conditional vector and are used to dynamically modulate the shared semantic representation in the shared backbone layer by layer.

4. The method for generating marine typhoon profiles based on drop-in typhoon radiosonde data and deep learning as described in claim 1, characterized in that, The residual blocks of odd-numbered layers in the shared backbone also introduce gated jump connections to fuse the initial output of the height branch with the output of the current residual block.

5. The method for generating marine typhoon profiles based on drop-in typhoon radiosonde data and deep learning as described in claim 1, characterized in that, The dual output head includes a low-rank coupling term, which constructs a cross-component physical correlation between the radial wind increment and the tangential wind increment through a low-rank matrix.

6. The method for generating marine typhoon profiles based on drop-in typhoon radiosonde data and deep learning as described in claim 1, characterized in that, The total loss function of the deep learning model Represented as: in, For data fitting loss; For curvature regularization loss; For tangential wind sign consistency loss; For the inflow angle prior loss; For logarithmic soft prior loss; This is due to the loss of modulation parameter stability; This is due to the low-rank coupling stability loss; For slope regularization loss; , , , , , , , These are the weighting coefficients for the corresponding loss terms; This represents the total number of sampling points; For interval The total number of valid sampling points involved in the calculation of the prior loss of the inflow angle; For interval The total number of valid sampling points involved in the calculation of logarithmic-law soft prior loss; For the first The height of each sampling point; For the first The weight of each sampling point; For robust penalty functions; Indicates a threshold parameter Robust penalty function; Indicates a threshold parameter Robust penalty function; and They are respectively height Predicted and observed values ​​of radial wind at the location; and They are respectively height Predicted and observed values ​​of tangential wind; For height The predicted wind vector at that location, and ; For height Curvature weight at the location; Coriolis parameter symbol; It is a second-order difference operator along the height direction; This represents the change in the inflow angle between adjacent heights; For height The reference wind speed modulus value is obtained based on the logarithmic law of the near-surface layer. This represents the total number of FiLM modulation layers; For the first One FiLM modulation layer; For the first A vector of scaling modulation parameters for each FiLM modulation layer; It is a low-rank coupling matrix; This represents the Frobenius norm.

7. The method for generating marine typhoon profiles based on drop-in typhoon radiosonde data and deep learning as described in claim 1, characterized in that, Also includes: A lightweight model is obtained by reducing the number of Fourier features, the number of residual blocks, or the rank of low-rank coupling terms in the deep learning model; the lightweight model is then used for inference deployment on edge computing platforms or in real-time wind field generation scenarios.

Images