Jet identification method fusing LLM numerical model analysis and radar wind field retrieval
Patent Information
- Application Number
- CN202611017503.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-09
- Publication Date
- 2026-09-29
AI Technical Summary
探空站时间分辨率低、空间分布稀疏，难以捕捉急流的精细结构与突发变化；数值预报模式虽能提供空间连续的结果，但本身存在系统偏差，且在复杂地形区域的模拟能力十分有限：一是急流识别往往依赖预报员人工分析天气图，通过目视判断等高线密集区或风矢量强梯度带来划定急流轴，主观性较强，不同预报员之间易产生判断差异，且难以实现识别过程的自动化与标准化；二是单一类型的天气雷达在低空区域存在波束遮挡和近地面探测盲区，噪点过滤能力不足，区域风场中常有大量孤立强风噪点，导致低空急流的底部结构无法被清晰探测
1）本发明将大语言模型的认知能力与多源雷达遥感技术的高精度探测能力相结合，克服了传统气象业务中人工分析的主观局限性和单一数据源的物理局限性；
Smart Images
Figure CN122836742A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of three-dimensional wind field construction technology, and in particular to a jet stream identification method that integrates LLM numerical model analysis and radar wind field inversion. Background Technology
[0002] The southwest jet stream is a strong and narrow low-level airflow active in southwestern my country and areas to its east. It exhibits a significant southerly or southwesterly component and is a crucial dynamic and thermal condition influencing the onset of the summer monsoon, the formation of torrential rains during the flood season, and the development of severe convective weather. Accurately identifying and forecasting the location, intensity, and evolution of the southwest jet stream is essential for improving early warning capabilities for severe weather events such as torrential rains and floods.
[0003] Traditional jet stream identification methods mainly rely on observations from conventional radiosonde stations or numerical weather prediction models. Radiosonde stations have low temporal resolution and sparse spatial distribution, making it difficult to capture the fine structure and sudden changes of jet streams. While numerical weather prediction models can provide spatially continuous results, they inherently have systematic biases and their simulation capabilities in complex terrain areas are very limited. First, jet stream identification often relies on forecasters manually analyzing weather maps and visually determining the jet stream axis based on areas of dense contour lines or strong wind vector gradients. This is highly subjective, and differences in judgment can easily arise between different forecasters, making it difficult to automate and standardize the identification process. Second, single-type weather radars suffer from beam obstruction and near-surface detection blind spots in low-altitude regions, and their noise filtering capabilities are insufficient. There are often many isolated strong wind noise points in the regional wind field, making it impossible to clearly detect the bottom structure of low-altitude jet streams.
[0004] Therefore, providing a jet stream identification method that integrates LLM numerical model analysis and radar wind field inversion to address the difficulties in existing technologies is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0005] In view of this, the present invention provides a jet stream identification method that integrates LLM numerical model analysis and radar wind field inversion, which improves the accuracy and precision of southwest jet stream identification under complex terrain.
[0006] To achieve the above objectives, the present invention adopts the following technical solution: A jet stream identification method integrating LLM numerical model analysis and radar wind field inversion includes the following steps: Collect wind field data from numerical models, perform data preprocessing on the wind field data to obtain wind direction and wind speed, and generate a gridded wind field dataset; The grid data is converted into structured text, the criteria for jet stream identification are defined, and the jet stream region is identified by combining the trained LLM, thus determining the approximate location and shape of the southwest jet stream. The rough location and morphology of the detected southwest jet stream were corrected with high precision using the dual-Doppler weather radar inversion method to obtain the three-dimensional wind field of the weather radar. Based on the horizontal and vertical wind data obtained by wind profiler radar and wind lidar in the region, a three-dimensional wind field in the region is constructed by networking wind profiler radars and using a range-weighted method. By combining the three-dimensional wind field of weather radar with the regional three-dimensional wind field and setting weights based on the coverage of dual-Doppler weather radar, the identified southwest jet stream is obtained.
[0007] Optional, generate a gridded wind field dataset: Construct a multi-source data access interface to collect wind field data output from global or regional numerical weather prediction models as raw format data; Wind speed is obtained by analyzing and calculating the U and V components in the original format data. With wind direction : , ; By combining wind speed and direction with the specified isobaric surface layer of the target area, gridded wind field data is formed.
[0008] Optionally, identifying rapid flow areas includes: Convert the gridded wind field dataset into structured text containing latitude, longitude, altitude, wind direction, wind speed, and gradient information; Wind speed at different air pressures is defined as a criterion for jet stream identification and used as a prompt word. By inputting structured text and prompts into the LLM model, isolated strong wind noise is filtered out, and meteorologically significant jet stream areas are identified.
[0009] Optionally, the LLM model includes a sparse coding framework and a transformer-decoder framework. The sparse coding framework is set before the input layer of the transformer-decoder framework, specifically as follows: Grid wind field sample data is obtained, and the sample data is dimensionality reduced by using a sparse coding framework to form a compact sequence representation. Then, the Transformer decoder framework is trained with the dimensionality-reduced data to complete self-supervised pre-training and obtain the LLM model. Based on historically calibrated jet stream area data, common data features are extracted as historical data features. The specific calculation methods used to extract these historical data features are sorted out and transformed into natural language descriptions. Using the natural language descriptions as instructions, combined with the corresponding wind field data, a fine-tuning dataset is constructed and input into the LLM model for instruction fine-tuning to obtain a trained LLM model.
[0010] Optional, obtaining horizontal and vertical wind data includes: The local coordinate system of each wind profiler radar is determined based on the known location information of the wind profiler radar sites. The local coordinate system of one of the wind profiler radars is used as the reference coordinate system, and the local coordinate systems of the other wind profiler radars are transformed to the reference coordinate system to correct the wind speed of each wind profiler radar. Adaptive resampling is performed on the wind speed vectors on the vertical wind profiles of each station after correction, and the vertical resolution and time reference are unified to form a multi-station wind profile dataset containing both horizontal and vertical wind data.
[0011] Optionally, constructing a regional three-dimensional wind field includes: Horizontal and vertical wind data at different altitudes are obtained from multi-site wind profile datasets. Distance-inverse weighted or Kriging spatial interpolation algorithms are used to grid the discrete site data into a regional three-dimensional wind field.
[0012] Optionally, identifying the southwest jet stream includes: A weighting function that varies with altitude and horizontal distance is defined. In the core area covered by dual-Doppler radar, the wind field retrieved by weather radar is given a high weight. In the edge area or near-surface blind area, the wind field of the wind profiler radar network is given a relatively high weight, so as to obtain the three-dimensional fine structure of the southwest jet stream after multi-source collaborative identification.
[0013] As can be seen from the above technical solution, compared with the prior art, the present invention provides a jet stream identification method that integrates LLM numerical model analysis and radar wind field inversion, which has the following beneficial effects: 1) This invention combines the cognitive ability of large language models with the high-precision detection capability of multi-source radar remote sensing technology, overcoming the subjective limitations of manual analysis and the physical limitations of a single data source in traditional meteorological operations. 2) This invention achieves rapid and intelligent initial screening of massive pattern data through an LLM model that combines sparse coding and Transformer, and realizes high-precision characterization of the core area of the jet stream by using dual Doppler radar inversion, significantly improving the accuracy and refinement of the identification of the southwest jet stream under complex terrain. Attached Figure Description
[0014] 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.
[0015] Figure 1This is a flowchart of the jet stream identification method that integrates LLM numerical model analysis and radar wind field inversion disclosed in this invention. Detailed Implementation
[0016] 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.
[0017] Reference Figure 1 As shown, this invention discloses a jet stream identification method that integrates LLM numerical model analysis and radar wind field inversion, comprising the following steps: Collect wind field data from numerical models, perform data preprocessing on the wind field data to obtain wind direction and wind speed, and generate a gridded wind field dataset; The grid data is converted into structured text, the criteria for jet stream identification are defined, and the jet stream region is identified by combining the trained LLM, thus determining the approximate location and shape of the southwest jet stream. The rough location and morphology of the detected southwest jet stream were corrected with high precision using the dual-Doppler weather radar inversion method to obtain the three-dimensional wind field of the weather radar. Based on the horizontal and vertical wind data obtained by wind profiler radar and wind lidar in the region, a three-dimensional wind field in the region is constructed by networking wind profiler radars and using a range-weighted method. By combining the three-dimensional wind field of weather radar with the regional three-dimensional wind field and setting weights based on the coverage of dual-Doppler weather radar, the identified southwest jet stream is obtained.
[0018] Furthermore, a gridded wind field dataset is generated: Construct a multi-source data access interface to collect wind field data output from global or regional numerical weather prediction models as raw format data; Wind speed is obtained by analyzing and calculating the U and V components in the original format data. With wind direction : , ; By combining wind speed and direction with the specified isobaric surface layer of the target area, gridded wind field data is formed.
[0019] Furthermore, the interface adopts a modular design, supporting the parsing and compatibility of multiple data formats and adapting to output data with different numerical modes. The core functions of the data access interface include: automatic data capture, format parsing, data validation, exception handling, and storage management. Among these: The automatic data capture system supports scheduled and real-time capture to ensure data timeliness. The format parsing module automatically identifies raw data in different formats and extracts wind field-related information. The data verification module performs integrity and validity checks on the captured data, removing missing and outlier values. The exception handling module automatically triggers a retry mechanism for data capture failures, format parsing errors, etc. If a retry fails, it logs and issues an alarm to ensure the stability of data acquisition. The storage management module adopts a distributed storage architecture, classifying and storing raw and preprocessed data, supporting fast data query, retrieval, and backup. The storage cycle can be set according to business needs.
[0020] Specifically, the scope of raw data collection needs to cover the target area and a certain surrounding area. The core parameters include: the U-component and V-component of the wind field output by the numerical model, latitude and longitude grid information, altitude information, and timestamp.
[0021] The analytical calculations are based on U and V components to determine wind speed and direction, and the results are validated to ensure that the wind direction value is within the range of 0–360° and the wind speed value is non-negative. Outliers occurring during the calculation process are corrected using linear interpolation between adjacent grid points. Simultaneously, a parallel computing architecture is employed to process large-scale grid data in batches to shorten computation time and meet real-time requirements.
[0022] Furthermore, the gridded wind field data includes: Determine the latitude and longitude range of the target area, match the latitude and longitude grid of the original numerical model data with the latitude and longitude range of the target area, and extract the grid data within the target area; Based on the specified isobaric surface layer, extract the wind speed and wind direction data corresponding to each isobaric surface layer to ensure that the height information of each grid point accurately corresponds to the isobaric surface layer. The extracted data is standardized to unify the data format and units, and metadata such as timestamps, isobaric surface identifiers, and grid numbers are added. The processed gridded data is organized according to the dimensions of time-isobaric surface-latitude and longitude to generate a gridded wind field dataset. The dataset is then checked to see if it meets the quality control standards. If not, the above steps are repeated. If it does, it is stored in a distributed database to support subsequent calls and analysis of LLM models.
[0023] Furthermore, identifying rapid flow areas includes: Convert the gridded wind field dataset into structured text containing latitude, longitude, altitude, wind direction, wind speed, and gradient information; Wind speed at different air pressures is defined as a criterion for jet stream identification and used as a prompt word. By inputting structured text and prompts into the LLM model, isolated strong wind noise is filtered out, and meteorologically significant jet stream areas are identified.
[0024] Furthermore, converting to structured text includes: The structured text of each grid point is organized in the form of key-value pairs, containing the following core information: grid point number, latitude and longitude, altitude, wind speed, wind direction, wind speed gradient, and timestamp; A batch processing script is used to convert each grid point in the grid wind field dataset into text, generating a structured text fragment for each grid point. The fragments are then organized in the manner of isobaric surface layer-region block to form a complete structured text dataset. Each region block contains structured text fragments of all grid points within that region, and the fragments are separated by delimiters. The converted structured text is optimized by removing redundant information, standardizing data format, and supplementing contextual information. The optimized structured text dataset needs to be quality checked to ensure that the text information is consistent with the gridded wind field data, thus forming structured text.
[0025] Furthermore, the LLM model includes a sparse coding framework and a transformer-decoder framework. The sparse coding framework is set before the input layer of the transformer-decoder framework, specifically as follows: Grid wind field sample data is obtained, and the sample data is dimensionality reduced by using a sparse coding framework to form a compact sequence representation. Then, the Transformer decoder framework is trained with the dimensionality-reduced data to complete self-supervised pre-training and obtain the LLM model. Based on historically calibrated jet stream area data, common data features are extracted as historical data features. The specific calculation methods used to extract these historical data features are sorted out and transformed into natural language descriptions. Using the natural language descriptions as instructions, combined with the corresponding wind field data, a fine-tuning dataset is constructed. This dataset is then used as the training set to input into the base model for fine-tuning, resulting in a trained LLM model.
[0026] Furthermore, the LLM model includes an input layer, a sparse coding framework, a transformer-decoder framework, and an output layer; among which, The sparse coding framework includes: acquiring structured text transformed from grid wind fields; determining a set of basic data as a dictionary from the structured text by optimizing and solving an objective function, and training and learning the dictionary. The objective function can be expressed as: , in, This represents the structured text input. For the dictionary that is being tried to determine, For the coefficient vector, To balance the parameters of reconstruction error and sparsity, To balance the parameters of reconstruction error and periodicity, It is the L2 norm. Describing the L1 norm, To perform a Fourier transform on X, To calculate the entropy value of the spectrum; The learned dictionary is used to encode the structured text to obtain a sparse coefficient vector, which is then used as a sequence representation.
[0027] Furthermore, the Transformer-decoder framework is composed of multiple stacked decoder blocks with identical structures. Each decoder block contains, in sequence, a masked multi-head self-attention layer, a first residual connection and layer normalization, a multi-head cross-attention layer, a second residual connection and layer normalization, a feedforward layer, and a third residual connection and layer normalization. The masked multi-head self-attention layer introduces a lower triangular mask, forcing the current sequence to focus only on itself and previous historical information; each residual connection and layer normalization adds the input and output of the previous layer, and then uses layer normalization to stabilize the deep network training and accelerate convergence; the feedforward layer consists of two linear transformations and a nonlinear activation function, performing a deep nonlinear transformation on the representation of each sequence; finally, the top layer output is linearly projected and the Softmax function is used to obtain the predicted probability distribution, realizing the identification of the jet stream region.
[0028] Furthermore, performance testing was conducted on the fine-tuned LLM model, with metrics including accuracy, recall, F1 score, and noise filtering rate, to ensure that the model can accurately identify the southwest jet stream area and effectively filter isolated strong wind noise.
[0029] Furthermore, identifying rapid flow areas specifically includes: The LLM model is used to analyze the information of each grid point in the structured text, extract core parameters such as latitude and longitude, altitude, wind speed, wind direction, and wind speed gradient, and establish the spatial correlation of wind field data. Based on the wind speed threshold and wind direction range in the prompt words, each grid point is judged and candidate grid points that meet the conditions are selected. Spatial continuity analysis is performed on candidate grid points to filter out isolated strong wind noise points and retain candidate grid point clusters with spatial continuity. Further analysis is conducted on the retained candidate grid clusters to determine whether they meet the preset conditions. Clusters that meet the conditions are identified as the southwest jet stream region, and their latitude and longitude range, isobaric surface layer, average wind speed, wind direction range, and approximate shape are extracted. The LLM model outputs the identification results of the southwest jet stream regions, including detailed information for each jet stream region and the confidence level of the identification results.
[0030] Furthermore, the inversion using dual-Doppler weather radar includes: Two Doppler weather radars with identical performance and coverage that can completely cover the rough area of the southwest jet stream identified by LLM were selected. The radar installation area was set up to ensure that the radar beams could effectively overlap, forming a dual radar observation network. The radar observation mode was set according to the height distribution characteristics of the southwest jet stream. The radar elevation angle, scanning range and scanning cycle were adjusted. The two radars were calibrated, including radial velocity calibration, azimuth calibration and elevation calibration, to ensure that the observation data of the two radars are consistent and accurate. According to the set observation mode, two Doppler weather radars were activated to conduct synchronous observations and collect radar echo data and radial velocity data in the jet stream area; The collected radar observation data is preprocessed to remove abnormal data and correct errors, ensuring the quality and usability of the data; The least squares method is used as the core algorithm for wind field inversion by dual Doppler radar. The error three-dimensional wind field data is obtained by minimizing the observed values and simulated values. The wind field data is then optimized by processing the inverted three-dimensional wind field data.
[0031] Furthermore, obtaining horizontal and vertical wind data includes: The local coordinate system of each wind profiler radar is determined based on the known location information of the wind profiler radar sites. Using the local coordinate system of one wind profiler radar as the reference coordinate system, the local coordinate systems of the remaining wind profiler radars are transformed to the reference coordinate system. The wind speed of each wind profiler radar is then corrected. This includes using a comparative calibration method, selecting a benchmark station in the area, comparing the observed wind speed of each radar station with the measured wind speed of the benchmark station, calculating the correction coefficient, and correcting the wind speed data of each radar station to ensure that the wind speed data of all stations are consistent. Adaptive resampling is performed on the wind speed vectors on the vertical wind profiles of each station after correction, and the vertical resolution and time reference are unified to form a multi-station wind profile dataset containing both horizontal and vertical wind data.
[0032] Specifically, the resampling uses an adaptive interpolation method, which adjusts the interpolation step size according to the spatial rate of change of wind speed to ensure that the resampled data can completely retain the vertical distribution characteristics of wind speed.
[0033] Furthermore, constructing the regional three-dimensional wind field includes: Horizontal and vertical wind data at different altitudes are obtained from multi-site wind profile datasets. Distance-inverse weighted or Kriging spatial interpolation algorithms are used to grid the discrete site data into a regional three-dimensional wind field.
[0034] Specifically, the method of using distance inverse weighting to obtain the regional three-dimensional wind field includes: , in, The wind speed value at the unknown grid point P. Let be the wind speed value at the i-th known station. Let P be the horizontal distance from the unknown grid point P to the i-th known station, p be the weight exponent, and N be the number of known stations participating in the interpolation.
[0035] Furthermore, the southwest jet stream was identified as including: A weighting function that varies with altitude and horizontal distance is defined. In the core area covered by dual-Doppler radar, the wind field retrieved by weather radar is given a high weight. In the edge area or near-surface blind area, the wind field of the wind profiler radar network is given a relatively high weight, so as to obtain the three-dimensional fine structure of the southwest jet stream after multi-source collaborative identification.
[0036] Furthermore, the weighting function expression is as follows: , in, It is a height weighting function, which takes a value close to 1 within the effective detection altitude range of the dual-Doppler radar and approaches 0 in the near-ground blind zone. It is a horizontal range weighting function, with a value of 1 in the core area of radar coverage and a linear decrease in the edge area.
[0037] 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 jet stream identification method integrating LLM numerical model analysis and radar wind field inversion, characterized in that, Includes the following steps: Collect wind field data from numerical models, perform data preprocessing on the wind field data to obtain wind direction and wind speed, and generate a gridded wind field dataset; The grid data is converted into structured text, the criteria for jet stream identification are defined, and the jet stream region is identified by combining the trained LLM, thus determining the approximate location and shape of the southwest jet stream. The rough location and morphology of the detected southwest jet stream were corrected with high precision using the dual-Doppler weather radar inversion method to obtain the three-dimensional wind field of the weather radar. Based on the horizontal and vertical wind data obtained by wind profiler radar and wind lidar in the region, a three-dimensional wind field in the region is constructed by networking wind profiler radars and using a range-weighted method. By combining the three-dimensional wind field of weather radar with the regional three-dimensional wind field and setting weights based on the coverage of dual-Doppler weather radar, the identified southwest jet stream is obtained.
2. The jet stream identification method integrating LLM numerical model analysis and radar wind field inversion as described in claim 1, characterized in that, Generate a gridded wind field dataset: Construct a multi-source data access interface to collect wind field data output from global or regional numerical weather prediction models as raw format data; Wind speed is obtained by analyzing and calculating the U and V components in the original format data. With wind direction : ， ； By combining wind speed and direction with the specified isobaric surface layer of the target area, gridded wind field data is formed.
3. The jet stream identification method integrating LLM numerical model analysis and radar wind field inversion as described in claim 1, characterized in that, Identifying rapid flow areas includes: Convert the gridded wind field dataset into structured text containing latitude, longitude, altitude, wind direction, wind speed, and gradient information; Wind speed at different air pressures is defined as a criterion for jet stream identification and used as a prompt word. By inputting structured text and prompts into the LLM model, isolated strong wind noise is filtered out, and meteorologically significant jet stream areas are identified.
4. The jet stream identification method integrating LLM numerical model analysis and radar wind field inversion according to claim 3, characterized in that, The LLM model includes a sparse coding framework and a transformer-decoder framework. The sparse coding framework is placed before the input layer of the transformer-decoder framework, specifically: Grid wind field sample data is obtained, and the sample data is dimensionality reduced by using a sparse coding framework to form a compact sequence representation. Then, the Transformer decoder framework is trained with the dimensionality-reduced data to complete self-supervised pre-training and obtain the LLM model. Based on historically calibrated jet stream area data, common data features are extracted as historical data features. The specific calculation methods used to extract these historical data features are sorted out and transformed into natural language descriptions. Using the natural language descriptions as instructions, combined with the corresponding wind field data, a fine-tuning dataset is constructed and input into the LLM model for instruction fine-tuning to obtain a trained LLM model.
5. The jet stream identification method integrating LLM numerical model analysis and radar wind field inversion according to claim 1, characterized in that, Obtaining horizontal and vertical wind data includes: The local coordinate system of each wind profiler radar is determined based on the known location information of the wind profiler radar sites. The local coordinate system of one of the wind profiler radars is used as the reference coordinate system, and the local coordinate systems of the other wind profiler radars are transformed to the reference coordinate system to correct the wind speed of each wind profiler radar. Adaptive resampling is performed on the wind speed vectors on the vertical wind profiles of each station after correction, and the vertical resolution and time reference are unified to form a multi-station wind profile dataset containing both horizontal and vertical wind data.
6. The jet stream identification method integrating LLM numerical model analysis and radar wind field inversion according to claim 5, characterized in that, Constructing a regional three-dimensional wind field includes: Horizontal and vertical wind data at different altitudes are obtained from multi-site wind profile datasets. Distance-inverse weighted or Kriging spatial interpolation algorithms are used to grid the discrete site data into a regional three-dimensional wind field.
7. The jet stream identification method integrating LLM numerical model analysis and radar wind field inversion according to claim 1, characterized in that, The southwest jet stream is identified as including: A weighting function that varies with altitude and horizontal distance is defined. In the core area covered by dual-Doppler radar, the wind field retrieved by weather radar is given a high weight. In the edge area or near-surface blind area, the wind field of the wind profiler radar network is given a relatively high weight, so as to obtain the three-dimensional fine structure of the southwest jet stream after multi-source collaborative identification.