CN-120995856-B - Ultra-low orbit 100-300km single-station atmospheric wind speed modeling method based on transfer learning

CN120995856BCN 120995856 BCN120995856 BCN 120995856BCN-120995856-B

Abstract

The invention relates to an ultra-low orbit 100-300km single-site atmospheric wind speed modeling method based on transfer learning, which comprises the steps of collecting multi-source wind field data at different heights, preprocessing the multi-source wind field data, carrying out fusion processing on the preprocessed multi-source wind field data to obtain fusion wind field data, carrying out wavelet multi-scale decomposition on the fusion wind field data to obtain multi-scale different-height characteristic data, inputting the multi-scale different-height characteristic data into a pre-built transfer learning frame, extracting domain independent characteristics and target domain generalization modeling, inputting the extracted domain independent characteristics and target domain generalization modeling result into a pre-built DeepONet model, obtaining a wind speed predicted value, and simultaneously obtaining gravitational wave amplitude and tidal phase parameters. The invention effectively solves the problem of height expansion under the condition of small samples, and is suitable for the scenes of ultralow orbit aircraft orbit control, space weather disturbance early warning, atmospheric density modeling and the like.

Inventors

CAI BING
HE YUBIN
XU QINGCHEN
YAN ZHAOAI
YANG JUNFENG
LI WEN
SUN ZHIBIN
WU YONGKUN

Assignees

中国科学院国家空间科学中心

Dates

Publication Date: 20260508
Application Date: 20250807

Claims (10)

1. An ultra-low orbit 100-300km single-site atmospheric wind speed modeling method based on transfer learning is characterized by comprising the following steps: Collecting multi-source wind field data at different heights, and preprocessing the multi-source wind field data; carrying out fusion processing on the preprocessed multi-source wind field data to obtain fusion wind field data; Performing wavelet multi-scale decomposition on the fused wind field data to obtain multi-scale characteristic data with different heights; Inputting the multi-scale feature data with different heights into a pre-constructed transfer learning frame, extracting field independent features, and simultaneously carrying out target field generalization modeling; Inputting the extracted domain independent characteristics and the target domain generalization modeling result into a pre-constructed DeepONet model to obtain a wind speed predicted value, simultaneously obtaining gravitational wave amplitude based on high-frequency components of the multi-scale different-height characteristic data, and obtaining tidal phase parameters based on tidal components.
2. The ultra-low track 100-300km single-site atmospheric wind speed modeling method based on transfer learning of claim 1, wherein preprocessing the multi-source wind field data comprises: And performing quality control and space-time alignment processing on the multi-source wind field data.
3. The ultra-low orbit 100-300km single-site atmospheric wind speed modeling method based on transfer learning of claim 1, wherein the fusion processing of the preprocessed multi-source wind field data comprises the following steps: for the height covered by the data, fusing the average value and standard deviation of the wind field to obtain first wind field data; filling the height without data coverage by adopting a low-order spherical harmonic fitting method to acquire second wind field data; and carrying out weighted fusion on the first wind field data and the second wind field data.
4. The ultra-low orbit 100-300km single-site atmospheric wind speed modeling method based on transfer learning of claim 1, wherein performing wavelet multi-scale decomposition on the fused wind field data comprises: And performing discrete wavelet transformation on the fused wind field data by using a Meyer wavelet basis, and adopting a multiscale selective extraction strategy according to the time resolution and the observed height distribution difference of each data source.
5. The ultra-low orbit 100-300km single-site atmospheric wind speed modeling method based on transfer learning according to claim 1, wherein the transfer learning framework comprises a feature extraction network, a pre-constructed DeepONet model, a gradient inversion layer, a domain discriminator and a joint loss function; Inputting the multi-scale different-height characteristic data into a pre-constructed migration learning framework comprises the following steps: Firstly, generating hidden representations through a feature extraction network, and then inputting the hidden representations to a DeepONet model and a gradient inversion layer simultaneously; The gradient inversion layer performs countermeasure training with the domain discriminator by inverting the gradient sign, forces the characteristics to have high invariance, and finally outputs the field-independent characteristics for DeepONet models to perform multitask prediction through coordination optimization of a joint loss function, wherein the joint loss function comprises prediction loss, domain countermeasure loss and physical constraint terms.
6. The ultra-low orbit 100-300km single-site atmospheric wind speed modeling method based on transfer learning of claim 5, wherein the joint loss function is: L total ＝L pred +λL phys +γL domain +μL phase Where L total is the joint loss, L pred is the predicted loss, L phys is the physical constraint, L domain is the domain countermeasure loss, and L phase is the phase consistency loss.
7. The ultra-low orbit 100-300km single-site atmospheric wind speed modeling method based on transfer learning of claim 1, wherein the pre-built DeepONet model comprises: the branch network adopts a multi-layer cascade structure and comprises a residual error connection and layer normalization processing module, and realizes the characteristic abstraction of a historical wind speed function through a nonlinear activation function; And the main network integrates a space-time coordinate coding mechanism, constructs a coordinate joint semantic mapping module based on the principle of multi-head self-attention, and captures the high-time coupling characteristic through sine position coding.
8. The ultra-low orbit 100-300km single-site atmospheric wind speed modeling method based on transfer learning of claim 7, wherein the pre-built DeepONet model further comprises an operator mapping function, employing a DeepONet framework, defined as follows: Where u is the input wind speed function, y is the space-time coordinates, b k and t k are the branch and trunk network outputs, G (u) (y) is the operator mapping of function space to real numbers, p is the hidden space dimension, and k is the feature summation index.
9. The ultra-low orbit 100-300km single-site atmospheric wind speed modeling method based on transfer learning of claim 8, wherein inputting the extracted domain independent features and the target domain generalization modeling result into a pre-constructed DeepONet model, obtaining a wind speed predicted value comprises: Inputting DeepONet the extracted domain independent characteristics into a branch network to replace an original wind speed function, initializing a backbone network coordinate encoder based on parameters obtained by target domain generalization modeling, incorporating DeepONet a physical constraint term into a loss function, and finally generating a wind speed predicted value through operator mapping.
10. The ultra-low track 100-300km single-site atmospheric wind speed modeling method based on transfer learning of claim 1, wherein obtaining gravitational wave amplitude based on the multi-scale different height feature data, and obtaining tidal phase parameters based on tidal components comprises: the quantitative characterization of the gravity wave characteristics is realized by carrying out nonlinear transformation on the high-frequency components of the multi-scale different-height characteristic data and combining a standard deviation scale constraint mechanism, and the gravity wave amplitude is obtained; Extracting tidal components by adopting a frequency domain analysis method, introducing a phase drift correction term which changes along with the height, establishing a vertical propagation model of the thermal layer tide, and obtaining the tide phase parameters.

Description

Ultra-low orbit 100-300km single-station atmospheric wind speed modeling method based on transfer learning Technical Field The invention relates to the technical field of ultra-low orbit atmospheric environment modeling, in particular to an ultra-low orbit 100-300km single-site atmospheric wind speed modeling method based on transfer learning. Background The ultra-low orbit (100-300 km) region is located at the interface of the thermal layer and the ionosphere and has unique aerodynamic characteristics that gravitational waves cause up to 25% of atmospheric density fluctuation at this height, magnetic storm events can cause thermal layer densities to surge by more than 50% within hours, and strong spatial shearing effects (wind speed peaks up to 400 m/s) exist in wind fields, forming complex multi-scale coupling systems (tidal waves, planetary waves and gravitational wave superposition). The extreme environment forms a serious challenge for ultra-low orbit aircraft orbit prediction, spatial weather research and monitoring and early warning, and high-precision wind speed modeling method support is needed. The existing technical system has obvious limitations that transient disturbance is difficult to capture due to calculation complexity of a physical model (such as TIE-GCM), an empirical model (such as HWM 14) can only reflect weather average state and cannot predict daily change, and the traditional deep learning method (LSTM/CNN and the like) is limited by low coverage rate of ultralow-orbit actual measurement data, so that the problem of generalized failure caused by sample scarcity is difficult to overcome. Although a hybrid decomposition method, such as VMD-LSTM, a modeling method for forecasting the atmospheric wind speed of 80-100km in the near space based on VMD-PSO-LSTM in the prior art can model under the condition of sufficient data of meteor radar, the modeling method is dependent on experience selection, is easy to cause modal aliasing problem and can not predict higher areas, and the modeling challenge of the cross-scale physical process peculiar to an ultra-low rail is not solved in the prior art based on a deep learning fusion radar based strong wind forecasting method and system for focusing the ground surface to 20 km high wind field forecast. With the remarkable strategic value of ultra-low orbit satellites in the fields of military reconnaissance and high-precision earth observation, the prior art cannot meet core requirements such as high-precision prediction, multi-scale parametric synchronous inversion, extreme event adaptability modeling and the like under the condition of small samples. Especially, under the condition of lacking an effective cross-height knowledge transfer mechanism, the construction of an intelligent model integrating a physical mechanism and data driving becomes a key path for breaking through the modeling bottleneck of the ultralow-orbit wind field. Disclosure of Invention Aiming at three technical bottlenecks of data scarcity, multi-scale disturbance separation difficulty and weak cross-height generalization capability existing in ultra-low rail wind field modeling, the invention provides an ultra-low rail 100-300km single-site atmospheric wind speed modeling method based on transfer learning, which is characterized in that the characteristic transfer of a 100-300km high layer is realized by utilizing the transfer learning through the explicit separation of tide and gravitational wave components by wavelet multi-scale decomposition, and a wind field mapping operator is constructed based on a depth operator network (DeepONet), so that the synchronous output of wind speed prediction and fluctuation parameters is finally realized. Compared with the prior art, the method has effective progress in data utilization rate, modeling accuracy and physical consistency. In order to achieve the above object, the present invention provides the following solutions: An ultra-low orbit 100-300km single-site atmospheric wind speed modeling method based on transfer learning comprises the following steps: Collecting multi-source wind field data at different heights, and preprocessing the multi-source wind field data; carrying out fusion processing on the preprocessed multi-source wind field data to obtain fusion wind field data; Performing wavelet multi-scale decomposition on the fused wind field data to obtain multi-scale characteristic data with different heights; Inputting the multi-scale characteristic data with different heights into a pre-constructed transfer learning frame, extracting domain independent characteristics for eliminating distribution differences among different heights/data sources, and simultaneously carrying out target domain generalization modeling to improve the generalization capability of the model at the non-visible height; Inputting the extracted domain independent features and the target domain generalization modeling result into a pre-constructed Deep