A method for modeling atmospheric wind speed at a single site at 100-300 km altitude based on transfer learning

Abstract

The application relates to a kind of ultra-low orbit 100-300km single site atmospheric wind speed modeling methods based on transfer learning, comprising: collecting multi-source wind field data at different heights, and pre-processing the multi-source wind field data;The multi-source wind field data after pre-processing is fused and processed, and fusion wind field data is obtained;The fusion wind field data is wavelet multiscale decomposition, and multi-scale different height feature data is obtained;The multi-scale different height feature data is input into the pre-constructed transfer learning framework, and the field-independent feature and target domain generalization modeling are extracted;The extracted field-independent feature and target domain generalization modeling result is input into the pre-constructed DeepONet model, and wind speed prediction value is obtained, while gravity wave amplitude and tidal phase parameter are obtained.The application effectively overcomes the height expansion problem under small sample conditions, and is suitable for ultra-low orbit vehicle orbit control, space weather disturbance early warning and atmospheric density modeling and the like scenes.

Description

Technical Field

[0001] This invention relates to the field of ultra-low orbit atmospheric environment modeling technology, and in particular to a method for modeling atmospheric wind speed at a single ultra-low orbit station at 100-300km based on transfer learning. Background Technology

[0002] The very low Earth orbit (UEO) region (100-300 km) lies at the boundary between the thermosphere and the ionosphere, exhibiting unique atmospheric dynamic characteristics: gravity waves at this altitude cause atmospheric density fluctuations of up to 25%, and geomagnetic storms can lead to a surge in thermospheric density of over 50% within hours. Simultaneously, the wind field experiences strong spatial shear effects (peak wind speeds reaching 400 m / s), forming a complex multi-scale coupled system (superposition of tidal waves, planetary waves, and gravity waves). This extreme environment poses a severe challenge to UEO spacecraft trajectory prediction, space weather research, and monitoring and early warning, urgently requiring high-precision wind speed modeling methods.

[0003] Current technological systems have significant limitations: physical models (such as TIE-GCM) struggle to capture transient disturbances due to computational complexity; empirical models (such as HWM14) can only reflect the average climate state and cannot predict daily variations; traditional deep learning methods (LSTM / CNN, etc.) are limited by the low coverage of ultra-low orbit (ULO) measured data, making it difficult to overcome the generalization failure problem caused by sample scarcity. Although hybrid decomposition methods, such as VMD-LSTM, and existing patented technologies, such as a VMD-PSO-LSTM-based modeling method for near-space atmospheric wind speed forecasting at 80-100 km altitude, can perform modeling under conditions of sufficient meteor radar data, they rely on empirical selection, are prone to mode aliasing problems, and cannot predict higher regions. Existing patented technologies, such as deep learning-based radar fusion wind forecasting methods and systems focusing on wind field forecasting from the Earth's surface to a height of 20 km, have not yet solved the modeling challenges of cross-scale physical processes unique to ULE.

[0004] As the strategic value of ultra-low orbit (ULO) satellites in military reconnaissance and high-precision Earth observation becomes increasingly apparent, existing technologies are unable to meet core requirements such as high-precision prediction under small sample conditions, simultaneous inversion of multi-scale parameters, and adaptive modeling for extreme events. Especially in the absence of an effective cross-altitude knowledge transfer mechanism, constructing intelligent models that integrate physical mechanisms and data-driven approaches has become a key path to overcome the bottleneck in ULE wind field modeling. Summary of the Invention

[0005] The purpose of this invention is to address three major technical bottlenecks in ultra-low orbit (ULE) wind field modeling: data scarcity, difficulty in separating multi-scale disturbances, and weak generalization ability across altitudes. This invention proposes a single-site atmospheric wind speed modeling method for ULE winds at altitudes of 100-300 km based on transfer learning. Its core innovation lies in: explicitly separating tidal and gravity wave components through wavelet multi-scale decomposition; utilizing transfer learning to achieve feature transfer at altitudes of 100-300 km; and constructing a wind field mapping operator based on a deep operator network (DeepONet), ultimately achieving simultaneous output of wind speed prediction and fluctuation parameters. Compared to existing technologies, this method achieves significant progress in data utilization, modeling accuracy, and physical consistency.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A method for modeling atmospheric wind speed at a single site in the 100-300km ultra-low orbit based on transfer learning includes:

[0008] Collect multi-source wind field data at different altitudes and preprocess the multi-source wind field data;

[0009] The preprocessed multi-source wind field data is fused to obtain fused wind field data.

[0010] The fused wind field data is decomposed using wavelet multi-scale decomposition to obtain multi-scale feature data at different heights.

[0011] The multi-scale feature data at different heights are input into a pre-built transfer learning framework to extract domain-independent features to eliminate distribution differences between different heights / data sources. At the same time, target domain generalization modeling is performed to improve the model's generalization ability at unseen heights.

[0012] The extracted domain-independent features and target domain generalization modeling results are input into a pre-built DeepONet model to obtain wind speed predictions; simultaneously, gravity wave amplitudes are obtained based on the multi-scale, different-height feature data; and tidal phase parameters are obtained based on tidal components.

[0013] Optionally, preprocessing the multi-source wind field data includes:

[0014] The multi-source wind field data is subjected to quality control and spatiotemporal alignment processing.

[0015] Optionally, the fusion processing of the preprocessed multi-source wind field data includes:

[0016] For altitudes with data coverage, the mean and standard deviation of the wind field are fused to obtain the first wind field data;

[0017] For heights lacking data coverage, a low-order spherical harmonic function fitting method is used to fill in the gaps and obtain the second wind field data.

[0018] The first wind field data and the second wind field data are weighted and fused.

[0019] Optionally, performing wavelet multi-scale decomposition on the fused wind field data includes:

[0020] The fused wind field data were subjected to discrete wavelet transform using the Meyer wavelet basis, and a multi-scale selective extraction strategy was adopted based on the differences in time resolution and observation height distribution of each data source.

[0021] Optionally, the transfer learning framework includes a feature extraction network, a pre-built DeepONet model, a gradient reversal layer, a domain discriminator, and a joint loss function;

[0022] Inputting the multi-scale, height-different feature data into a pre-built transfer learning framework includes:

[0023] Multi-scale feature data at different heights are first processed by a feature extraction network to generate hidden representations, which are then simultaneously input into the DeepONet model and the gradient inversion layer.

[0024] The gradient inversion layer reverses the gradient sign and engages in adversarial training with the domain discriminator, forcing the features to have high invariance. Finally, it coordinates and optimizes through a joint loss function, outputting domain-independent features for the DeepONet model to perform multi-task prediction. The joint loss function includes prediction loss, domain adversarial loss, and physical constraint term.

[0025] Optionally, the joint loss function is:

[0026] L total =L pred +λL phys +γL domain +μL phase

[0027] Among them, L total For joint losses, L pred To predict losses, L phys For physical constraint terms, L domain For domain adversarial losses, L phase This represents the loss of phase consistency.

[0028] Optionally, the pre-built DeepONet model includes:

[0029] The branch network adopts a multi-layer cascaded structure, including residual connection and layer normalization processing modules, and realizes the feature abstraction of historical wind speed function through nonlinear activation function;

[0030] The backbone network integrates a spatiotemporal coordinate encoding mechanism and constructs a coordinate joint semantic mapping module based on the multi-head self-attention principle. It captures the high-temporal coupling characteristics through sinusoidal position encoding.

[0031] Optionally, the pre-built DeepONet model further includes: an operator mapping function, which adopts the DeepONet framework and is defined as follows:

[0032]

[0033] Where u is the input wind speed function, y is the spatiotemporal coordinate, and b is the input wind speed function. k and t k , respectively, are the outputs of the branch network and the backbone network, G(u)(y) is the operator mapping from the function space to real numbers, p is the dimension of the latent space, and k is the feature summation index.

[0034] Optionally, the extracted domain-independent features and the target domain generalization modeling results are input into a pre-built DeepONet model to obtain wind speed prediction values, including:

[0035] The extracted domain-independent features are input into the DeepONet branch network to replace the original wind speed function. At the same time, the backbone network coordinate encoder is initialized based on the parameters obtained from the target domain generalization modeling, and the physical constraint terms are incorporated into the DeepONet loss function. Finally, the wind speed prediction value is generated through operator mapping.

[0036] Optionally, obtaining the gravity wave amplitude based on the multi-scale height characteristic data and obtaining the tidal phase parameters based on the tidal components include:

[0037] By performing nonlinear transformation on the high-frequency components of the multi-scale, different-height characteristic data, and combining the scale constraint mechanism of standard deviation, the characteristics of transient gravity waves are quantitatively characterized, and the amplitude of the gravity waves is obtained.

[0038] The dominant period of the tidal component is extracted using a frequency domain analytical method. A phase drift correction term that varies with altitude is introduced to establish a vertical propagation model of the thermosphere tides and obtain the tidal phase parameters.

[0039] The beneficial effects of this invention are as follows:

[0040] In this invention, the reliability and physical consistency of ultra-low orbit wind field data are significantly improved by multi-source data fusion and physical-driven preprocessing, eliminating systematic errors from different detection methods. A wavelet multi-scale decomposition mechanism is used to achieve accurate separation of gravity wave and tidal components, solving the mode aliasing problem of traditional methods. A cross-altitude transfer learning framework reduces the number of training samples required under conditions of scarce target domain data, overcoming the bottleneck of small-sample generalization. The DeepONet multi-task operator design simultaneously outputs wind speed predictions, gravity wave amplitude, and tidal phase parameters, simplifying the complex process of traditional multi-model concatenation.

[0041] This invention effectively overcomes the challenge of altitude expansion under small sample conditions and is applicable to scenarios such as orbit control of ultra-low orbit vehicles, early warning of space weather disturbances, and atmospheric density modeling. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a schematic diagram of a method for modeling atmospheric wind speed at a single site in the 100-300km ultra-low orbit based on transfer learning, according to an embodiment of the present invention.

[0044] Figure 2 This is a comparison diagram of the domain adaptation feature distribution alignment in an embodiment of the present invention;

[0045] Figure 3 This is a diagram of the DeepONet network architecture according to an embodiment of the present invention;

[0046] Figure 4 This is a feature response diagram of an embodiment of the present invention;

[0047] Figure 5 This is a diagram of the multi-task output structure according to an embodiment of the present invention. Detailed Implementation

[0048] 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.

[0049] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0050] like Figure 1 As shown, this embodiment proposes a method for modeling atmospheric wind speed at a single site in the 100-300km ultra-low orbit based on transfer learning, including:

[0051] Collect multi-source wind field data at different altitudes and preprocess the multi-source wind field data;

[0052] The preprocessed multi-source wind field data is fused to obtain fused wind field data.

[0053] The fused wind field data is decomposed using wavelet multi-scale decomposition to obtain multi-scale feature data at different heights.

[0054] The multi-scale, high-height feature data is input into a pre-constructed transfer learning framework to extract domain-independent features and simultaneously perform target domain generalization modeling.

[0055] The extracted domain-independent features and target domain generalization modeling results are input into the pre-built DeepONet model to obtain wind speed prediction values; at the same time, gravity wave amplitude is obtained based on the high-frequency components of the multi-scale different height feature data, and tidal phase parameters are obtained based on the tidal components.

[0056] Furthermore, the preprocessing of the multi-source wind field data includes:

[0057] The multi-source wind field data is subjected to quality control and spatiotemporal alignment processing.

[0058] Specifically, in this embodiment, data acquisition and preprocessing are performed as follows:

[0059] Acquire wind field data from meteor radar at altitudes of 90-110km and FPI interferometers at altitudes of 100km and 250km, as well as wind field data from ICON satellites at altitudes of 90-110km and 200-300km and wind field data from reanalysis models at altitudes of 100-150km, and perform preprocessing such as quality control and spatiotemporal alignment;

[0060] Quality control is performed on different detection data. Outliers are removed through information range checks, with the maximum possible range of wind speed and its changes used as the criterion. Specifically, this is represented by F. min ≤F≤F max , of which F min and F max These are the minimum and maximum threshold values ​​for wind speed, respectively. Values ​​where the wind speed is outside the range of variation are invalidated. Here, the minimum threshold value F for wind speed is... min Set to 0 m / s, maximum threshold Fmax Set to 200m / s.

[0061] Spatiotemporal alignment of the probe data is performed. Using Coordinated Universal Time (UTC) as the reference, linear interpolation is used to generate hourly resolution sequences for discontinuous observation data to achieve time alignment. Data from different probe sources (ICON satellite / reanalysis mode) are aligned by latitude and longitude gridding, and missing grids are filled by Lagrange interpolation to achieve spatial matching with the station probe data.

[0062] Furthermore, the fusion processing of the preprocessed multi-source wind field data includes:

[0063] For altitudes with data coverage, the mean and standard deviation of the wind field are fused to obtain the first wind field data;

[0064] For heights lacking data coverage, a low-order spherical harmonic function fitting method is used to fill in the gaps and obtain the second wind field data.

[0065] The first wind field data and the second wind field data are weighted and fused.

[0066] Specifically, in this embodiment, multi-source data fusion is performed:

[0067] For multi-source wind field data at different altitudes, the mean and standard deviation of the wind field in areas with data coverage are fused, and a low-order spherical harmonic function is used to fit the data where data is missing. For the transition area (the fused data may have a jump at the seam), a local weighted regression method is used to smooth it, and then weighted fusion is performed.

[0068] For altitudes with data coverage, the mean and standard deviation of the wind field are fused using the following formula:

[0069]

[0070] Where X is the variable to be statistically analyzed, namely wind speed; X i Let σ be the statistical average of the i-th type of data at that height. i Let M be the statistical standard deviation of the i-th type of data at this altitude, where M = 1, 2, 3, 4 represent that ICON data, meteor radar data, FPI data, and reanalysis mode data are covered at this altitude, respectively; N is the total number of data of type M at this altitude. σ represents the fusion mean at that height, and σ represents the fusion standard deviation at the corresponding height.

[0071] For heights lacking data coverage, a low-order spherical harmonic function fitting method is used to fill in the gaps. The fitting function is as follows:

[0072]

[0073] There may be abrupt changes at the seams in the merged data. A locally weighted regression method (LOESS) is used for smoothing and data imputation.

[0074] Furthermore, the wavelet multi-scale decomposition of the fused wind field data includes:

[0075] The fused wind field data were subjected to discrete wavelet transform using the Meyer wavelet basis, and a multi-scale selective extraction strategy was adopted based on the differences in time resolution and observation height distribution of each data source.

[0076] Specifically, in this embodiment, the Meyer wavelet basis is used to perform Discrete Wavelet Transform (DWT) on the fused wind field data. Based on the differences in temporal resolution and observation height distribution of each data source, a multi-scale selective extraction strategy is adopted. For example, ICON data is mainly used to extract mid-to-low frequency background wind field and tidal features, while high temporal resolution radar data is preferentially used to capture high-frequency gravity wave components, thereby achieving complementary modeling of different source data across multiple frequency bands.

[0077] This embodiment uses the Meyer wavelet basis to perform a 5-level discrete wavelet transform:

[0078] coeffs = pywt.wavedec(s w ,'meyer',level=5);

[0079] The fifth layer low-frequency approximation component is used to extract background wind field and tidal components; the first to fourth layers high-frequency detail components are used to capture gravity wave disturbances; the differences in the fused data structure (such as ICON / FPI / radar / mode) at different altitudes determine the priority of feature extraction in multi-scale space. Since data is missing in the 150-200km range, inference is performed by training the model using features from upper and lower scales. The specific feature scale selection strategy and feature response diagram are shown below. Figure 4 As shown.

[0080] Furthermore, the transfer learning framework includes a feature extraction network, a pre-built DeepONet model, a gradient reversal layer, a domain discriminator, and a joint loss function;

[0081] Inputting the multi-scale, height-different feature data into a pre-built transfer learning framework includes:

[0082] Multi-scale feature data at different heights are first processed by a feature extraction network to generate hidden representations, which are then simultaneously input into the DeepONet model and the gradient inversion layer.

[0083] The gradient inversion layer reverses the gradient sign and engages in adversarial training with the domain discriminator, forcing the features to have high invariance. Finally, it coordinates and optimizes through a joint loss function, outputting domain-independent features for the DeepONet model to perform multi-task prediction. The joint loss function includes prediction loss, domain adversarial loss, and physical constraint term.

[0084] More specifically, using multi-scale feature data at different heights as the source domain, DeepONet is trained to learn the spatiotemporal wind field mapping operator. A gradient inversion layer is introduced to construct a feature alignment mechanism. This gradient inversion layer is located between the DeepONet branch network and the domain discriminator, receiving the hidden representation output by the feature extraction network and achieving domain-invariant feature alignment between the source and target domains by inverting the gradient direction of the hidden representation. A joint loss function is used for domain-independent feature extraction and target domain generalization modeling. A comparison of domain-adaptive feature distribution alignment is shown in the figure below. Figure 2 As shown.

[0085] Specifically, in this embodiment, the transfer learning framework is constructed as follows:

[0086] Source domain pre-training: DeepONet is trained to learn the spatiotemporal wind field mapping operator using multi-scale feature data of different heights as the source domain;

[0087] Domain Adaptive Transfer: A gradient inversion layer (GRL) is introduced to construct a feature alignment mechanism. Located between the DeepONet branch network and the domain discriminator, the GRL receives the hidden representation output by the feature extraction network and, by inverting its gradient direction, achieves domain-invariant feature alignment between the source and target domains, improving the model's cross-domain generalization ability. Domain-independent feature extraction and target domain generalization modeling are performed through a joint loss function (including prediction loss, physical constraints, adversarial loss, and phase consistency loss).

[0088] The loss function for adaptive migration of the domain is:

[0089]

[0090] in, To predict losses, For physical constraints, For domain confrontation losses, For phase consistency loss, here It is the model's predicted wind speed, u i These are actual observed values. Let represent the physical constraint function, h1 and h2 be the heights of the source and target domains respectively, φ0 be the initial phase, and α be the phase change rate.

[0091] Furthermore, the pre-built DeepONet model includes:

[0092] The branch network adopts a multi-layer cascaded structure, including residual connection and layer normalization processing modules, and realizes the feature abstraction of historical wind speed function through nonlinear activation function;

[0093] The backbone network integrates a spatiotemporal coordinate encoding mechanism and constructs a coordinate joint semantic mapping module based on the multi-head self-attention principle. It captures the high-temporal coupling characteristics through sinusoidal position encoding.

[0094] Specifically, in this embodiment, DeepONet models:

[0095] Network Structure: The branch network adopts a multi-layer fully connected structure, introducing residual connections and layer normalization mechanisms. Nonlinear activation functions are used to abstract historical wind speed function features. The backbone network combines spatiotemporal coordinate embedding and location encoding mechanisms. Temporal information uses sine-cosine location encoding to represent periodic features, while spatial location (altitude, latitude, and longitude) uses trainable embedding vectors to represent spatial continuity. A joint semantic representation integrating temporal and spatial distributions is constructed. The DeepONet network architecture is as follows: Figure 3 As shown.

[0096] Operator mapping construction: Define the wind speed prediction mapping as:

[0097] Where u is the input wind speed function, y is the spatiotemporal coordinate, and b is the input wind speed function. k and t k The outputs are for the branch network and the backbone network, respectively.

[0098] The branch network adopts a multi-layer cascaded structure and uses a nonlinear activation function to abstract the features of the historical wind speed function. Its output feature vector satisfies:

[0099] b = LayerNorm(f b,5 (Swish(f b,4 (...f b,1 (u)...))))

[0100] Where f b,i This indicates the i-th fully connected layer, with residual connections implemented between the third and fifth layers.

[0101] The backbone network integrates a spatiotemporal coordinate encoding mechanism, constructs a coordinate joint semantic mapping module based on the multi-head self-attention principle, and captures the height-temporal coupling characteristics through sinusoidal position encoding.

[0102] t = Transformer([PE(t); Ε)); λ ; E φ ; E h ])

[0103] Where PE(t) is the time location code, and E is the embedding vector of latitude, longitude and altitude.

[0104] Specifically, in this embodiment, such as Figure 5 As shown, multi-task output: synchronously outputs the predicted future wind speed, gravity wave amplitude, and tidal phase parameters at the target height point;

[0105] Multi-task output adopts a shared encoding-task decoding structure:

[0106] The main structure is a shared DeepONet backbone and branch network;

[0107] In the output stage, three parallel fully connected output heads are set up, each corresponding to the predicted wind speed value. Gravity wave amplitude estimate and tidal phase parameters

[0108] Each task head consists of a 2-layer MLP, and a task-specific normalization (LayerNorm) layer is introduced to improve generalization ability. At the same time, collaborative training is carried out through a joint loss function to enable different tasks to share features and enhance the learning effect.

[0109] Furthermore, the extracted domain-independent features and the target domain generalization modeling results are input into the pre-built DeepONet model to obtain wind speed prediction values, including:

[0110] The extracted domain-independent features are input into the DeepONet branch network to replace the original wind speed function. At the same time, the backbone network coordinate encoder is initialized based on the parameters obtained from the target domain generalization modeling, and the physical constraint terms are incorporated into the DeepONet loss function. Finally, the wind speed prediction value is generated through operator mapping.

[0111] Wind speed prediction is based on the generalized output of the operator mapping function. It utilizes the latent space interaction between the feature vectors of the branch network and the coordinate encoding of the backbone network to generate wind speed distribution in a continuous spatiotemporal domain.

[0112] Furthermore, obtaining the gravity wave amplitude based on the multi-scale height characteristic data and obtaining the tidal phase parameters based on the tidal components include:

[0113] By performing nonlinear transformation on the high-frequency components obtained from wavelet transformation of wind field data at different heights at multiple scales, and combining the scale constraint mechanism of standard deviation, the characteristics of gravity waves are quantitatively characterized, and the amplitude of the gravity waves is obtained.

[0114] Tidal components are extracted using a frequency domain analytical method. For example, the dominant period of semi-diurnal tides is 12 hours and the dominant period of diurnal tides is 24 hours. A phase drift correction term that varies with altitude is introduced to establish a vertical propagation model of thermospheric tides and obtain the tidal phase parameters.

[0115] Gravity wave amplitude is quantitatively characterized by nonlinear transformation of the high-frequency wavelet components and a scale constraint mechanism based on standard deviation.

[0116] The dominant period of the tidal components was extracted using a frequency domain analytical method, and a phase drift correction term varying with altitude was introduced to establish a vertical propagation model of thermospheric tides. Where T is the tidal period.

[0117] In this embodiment, model validation is also performed: the root mean square error (RMSE) and the correlation coefficient are used as evaluation indicators of the model's prediction accuracy and consistency to verify the model's spatiotemporal generalization ability.

[0118] In this embodiment, the reliability and physical consistency of ultra-low orbit wind field data are significantly improved by multi-source data fusion and physical-driven preprocessing, eliminating systematic errors from different detection methods. A wavelet multi-scale decomposition mechanism is used to achieve accurate separation of gravity waves and tidal components, solving the mode aliasing problem of traditional methods. A cross-altitude transfer learning framework reduces the number of training samples required under conditions of scarce target domain data, overcoming the bottleneck of small-sample generalization. DeepONet's multi-task operator design simultaneously outputs wind speed predictions, gravity wave amplitude, and tidal phase parameters, simplifying the complex process of traditional multi-model concatenation.

[0119] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for modeling atmospheric wind speed at a single site in the 100-300km ultra-low orbit based on transfer learning, characterized in that, include: Collect multi-source wind field data at different altitudes and preprocess the multi-source wind field data; The preprocessed multi-source wind field data is fused to obtain fused wind field data. The fused wind field data is decomposed using wavelet multi-scale decomposition to obtain multi-scale feature data at different heights. The multi-scale, high-height feature data is input into a pre-constructed transfer learning framework to extract domain-independent features and simultaneously perform target domain generalization modeling. The extracted domain-independent features and target domain generalization modeling results are input into the pre-built DeepONet model to obtain wind speed prediction values; at the same time, gravity wave amplitude is obtained based on the high-frequency components of the multi-scale different height feature data, and tidal phase parameters are obtained based on the tidal components.

2. The method for modeling atmospheric wind speed at a single site in the 100-300km ultra-low orbit based on transfer learning as described in claim 1, characterized in that, Preprocessing of the multi-source wind field data includes: The multi-source wind field data is subjected to quality control and spatiotemporal alignment processing.

3. The method for modeling atmospheric wind speed at a single site in the 100-300km ultra-low orbit based on transfer learning as described in claim 1, characterized in that, The fusion processing of the preprocessed multi-source wind field data includes: For altitudes with data coverage, the mean and standard deviation of the wind field are fused to obtain the first wind field data; For heights lacking data coverage, a low-order spherical harmonic function fitting method is used to fill in the gaps and obtain the second wind field data. The first wind field data and the second wind field data are weighted and fused.

4. The method for modeling atmospheric wind speed at a single site in the 100-300km ultra-low orbit based on transfer learning as described in claim 1, characterized in that, The wavelet multi-scale decomposition of the fused wind field data includes: The fused wind field data were subjected to discrete wavelet transform using the Meyer wavelet basis, and a multi-scale selective extraction strategy was adopted based on the differences in time resolution and observation height distribution of each data source.

5. The method for modeling atmospheric wind speed at a single site in the 100-300km ultra-low orbit based on transfer learning as described in claim 1, characterized in that, The transfer learning framework includes a feature extraction network, a pre-built DeepONet model, a gradient reversal layer, a domain discriminator, and a joint loss function. Inputting the multi-scale, height-different feature data into a pre-built transfer learning framework includes: Multi-scale feature data at different heights are first processed by a feature extraction network to generate hidden representations, which are then simultaneously input into the DeepONet model and the gradient inversion layer. The gradient inversion layer reverses the gradient sign and engages in adversarial training with the domain discriminator, forcing the features to have high invariance. Finally, it coordinates and optimizes through a joint loss function, outputting domain-independent features for the DeepONet model to perform multi-task prediction. The joint loss function includes prediction loss, domain adversarial loss, and physical constraint term.

6. The method for modeling atmospheric wind speed at a single site in the 100-300km ultra-low orbit based on transfer learning as described in claim 5, is characterized in that... The joint loss function is: L total =L pred +λL phys +γL domain +μL phase Among them, L total For joint losses, L pred To predict losses, L phys For physical constraint terms, L domain For domain adversarial losses, L phase This represents the loss of phase consistency.

7. The method for modeling atmospheric wind speed at a single site in the 100-300km ultra-low orbit based on transfer learning as described in claim 1, characterized in that, The pre-built DeepONet model includes: The branch network adopts a multi-layer cascaded structure, including residual connection and layer normalization processing modules, and realizes the feature abstraction of historical wind speed function through nonlinear activation function; The backbone network integrates a spatiotemporal coordinate encoding mechanism and constructs a coordinate joint semantic mapping module based on the multi-head self-attention principle. It captures the high-temporal coupling characteristics through sinusoidal position encoding.

8. The method for modeling atmospheric wind speed at a single site in the 100-300km ultra-low orbit based on transfer learning as described in claim 7, characterized in that, The pre-built DeepONet model also includes: operator mapping functions, which adopt the DeepONet framework and are defined as follows: Where u is the input wind speed function, y is the spatiotemporal coordinate, and b is the input wind speed function. k and t k , respectively, are the outputs of the branch network and the backbone network, G(u)(y) is the operator mapping from the function space to real numbers, p is the dimension of the latent space, and k is the feature summation index.

9. The method for modeling atmospheric wind speed at a single site in the 100-300km ultra-low orbit based on transfer learning as described in claim 8, characterized in that, The extracted domain-independent features and the target domain generalization modeling results are input into the pre-built DeepONet model to obtain wind speed prediction values, including: The extracted domain-independent features are input into the DeepONet branch network to replace the original wind speed function. At the same time, the backbone network coordinate encoder is initialized based on the parameters obtained from the target domain generalization modeling, and the physical constraint terms are incorporated into the DeepONet loss function. Finally, the wind speed prediction value is generated through operator mapping.

10. The method for modeling atmospheric wind speed at a single site in the 100-300km ultra-low orbit based on transfer learning according to claim 1, characterized in that, The acquisition of gravity wave amplitude based on the multi-scale, different height characteristic data, and the acquisition of tidal phase parameters based on tidal components, include: By performing nonlinear transformation on the high-frequency components of the multi-scale, different-height characteristic data, and combining the scale constraint mechanism of standard deviation, the characteristics of gravity waves are quantitatively characterized, and the amplitude of the gravity waves is obtained. The tidal components are extracted using a frequency domain analytical method, and a phase drift correction term that varies with altitude is introduced to establish a vertical propagation model of the thermosphere tides, thereby obtaining the tidal phase parameters.

Images