Squall-type thunderstorm gale forecasting method and device

By constructing a near-surface thermal and dynamic disturbance precursor identification mechanism, combined with deep learning and potential diffusion models, the problems of short warning time, high false alarm rate and high missed alarm rate of thunderstorm and gale warnings have been solved, and high-precision, early warning and clear spatial structure forecast results have been achieved.

CN123113048APending Publication Date: 2026-11-17NANJING UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611315140.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-27
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing methods for issuing near-term warnings of thunderstorms and gales suffer from problems such as short warning time, high false alarm rate, high missed alarm rate, and ambiguous spatial structure of forecast results. In particular, it is difficult to accurately identify high-risk areas in squall line thunderstorm and gale forecasts.

Method used

A mechanism for identifying near-ground thermal and dynamic disturbance precursors is constructed. The near-ground precursor disturbance field is extracted using dual-polarization radar features and ground observation data. The field is then predicted using a deep learning prediction network. A potential diffusion model is introduced for post-processing, and risk weights are dynamically adjusted to improve the accuracy of early warning.

Benefits of technology

It enables early warning of squall line thunderstorms and strong winds, reduces the missed reporting rate and false alarm rate, improves the spatial structure clarity and interpretability of forecast results, extends the warning time, and enhances the value of business applications.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123113048A_ABST
    Figure CN123113048A_ABST
Patent Text Reader

Abstract

This invention provides a method and apparatus for predicting the imminent arrival of squall-type thunderstorms and strong winds. The method includes: extracting dual-polarization radar features from radar base data after quality control; obtaining radar radial wind vorticity and divergence at different radial distance scales using a least-squares fitting method based on radar radial velocity, as low-level dynamic background features; selecting ground observation samples that match radar observation records in time and space from near-surface antecedent disturbance observation data to construct a near-surface antecedent disturbance field; obtaining thunderstorm and strong wind imminent arrival prediction results using a deep learning prediction network based on the dual-polarization radar features, low-level dynamic background features, and near-surface antecedent disturbance field; and enhancing the thunderstorm and strong wind imminent arrival prediction results using a latent diffusion model. This invention improves the accuracy, reliability, and interpretability of thunderstorm and strong wind imminent arrival warning results.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of meteorological forecasting technology, and in particular to a method and apparatus for near-term forecasting of squall line thunderstorms and strong winds. Background Technology

[0002] Forecasting thunderstorms and strong winds is a challenging aspect of current severe convective weather warnings. The core reason is that severe convective wind systems develop rapidly, have short lifespans, and are highly localized. Their occurrence often involves rapid coupling between downdrafts within the storm, outflows from cold pools, gust front advancement, low-level divergence, and cloud microphysical processes.

[0003] Current thunderstorm and gale nowcasting technologies mainly fall into two categories: extrapolation methods based on radar and satellite observations, and physical models based on numerical weather prediction. However, these technologies have significant shortcomings in practical applications. First, traditional extrapolation methods have limited ability to characterize the rapid formation and dissipation of severe convective systems and their sudden local evolution, limiting the lead time for warnings and leaving insufficient time for public protection. Second, judgments based on a single wind field are prone to false alarms, resulting in a high false alarm rate for forecast models and affecting the reliability of the system. Furthermore, existing numerical weather prediction models are limited by spatial resolution and physical parameterization schemes, making it difficult to accurately characterize the near-surface disturbance evolution before the occurrence of local extreme gales. Finally, while existing artificial intelligence (AI) early warning methods can integrate multi-source meteorological data, they mostly remain at the level of treating meteorological elements as ordinary input variables. Moreover, existing AI models generally suffer from "inertia" due to "imbalanced samples." Since samples with no gales constitute the vast majority of the dataset, the model tends to predict all situations as "no gales" during training in pursuit of overall accuracy, thus missing truly high-risk, sparsely sampled gale events. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a method and apparatus for near-term prediction of squall line thunderstorms and strong winds.

[0005] This invention provides a method for near-term forecasting of squall line thunderstorms and strong winds, comprising:

[0006] After quality control of the radar base data, dual-polarization radar features are extracted from it;

[0007] Based on radar radial velocity, the least squares fitting method is used to invert radar radial wind vorticity and divergence at different radial distance scales, which are used as low-level dynamic background features.

[0008] Ground observation samples that match radar observation records in time and space are selected from near-ground frontal body disturbance observation data to construct near-ground frontal body disturbance field;

[0009] Based on the dual-polarization radar features, low-level dynamic background features, and near-ground frontal disturbance field, a deep learning prediction network is used to obtain the near-term prediction results of thunderstorms and strong winds.

[0010] The thunderstorm gale proximity prediction results are enhanced and post-processed based on the potential diffusion model.

[0011] The present invention also provides a squall line thunderstorm gale forecasting device, comprising:

[0012] The extraction module is used to extract dual-polarization radar features from the radar base data after quality control.

[0013] The inversion module is used to invert radar radial wind vorticity and divergence at different radial distance scales based on radar radial velocity using the least squares fitting method, as low-level dynamic background features.

[0014] The filtering module is used to select ground observation samples that match the radar observation records in time and space from the near-ground frontal body disturbance observation data, and to construct the near-ground frontal body disturbance field;

[0015] The prediction module is used to obtain the near-term prediction results of thunderstorms and strong winds based on the dual-polarization radar features, low-level dynamic background features and near-ground frontal disturbance field using a deep learning prediction network.

[0016] The post-processing module is used to enhance the thunderstorm gale proximity prediction results based on the potential diffusion model.

[0017] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement any of the above-described squall line thunderstorm gale near-term prediction methods.

[0018] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the squall line thunderstorm gale near-term prediction method as described above.

[0019] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements any of the above-described squall line thunderstorm gale near-term prediction methods.

[0020] To address the problems of short-term warning time, high false alarm rate, high missed alarm rate, and ambiguous spatial structure of forecast results in existing short-term warning methods for thunderstorms and strong winds, this invention proposes a squall line thunderstorm and strong wind nowcasting method and device that integrates near-surface forebody disturbance identification, dynamic risk weight adjustment, and potential diffusion model post-processing. This method focuses on the evolution mechanism of squall line strong convective systems, centering on key processes such as storm downdraft enhancement, cold pool outflow expansion, gust front advancement, and rapid enhancement of near-surface disturbances. It constructs a multi-source forebody disturbance field that characterizes the potential for thunderstorm and strong wind formation. This forebody disturbance field is not a simple superposition of ordinary meteorological elements, but a comprehensive characterization of the evolution characteristics of near-surface thermodynamic disturbances during the development of strong convective systems.

[0021] Building upon this foundation, this invention further generates dynamic risk weights based on the strength and spatial distribution of the precursor disturbance field. These weights enhance the model's ability to focus on potentially high-risk areas and low-probability strong wind samples, thereby improving the stability of identifying squall line-type thunderstorm wind risk areas. Subsequently, a Latent Diffusion Model (LDM) is introduced as a post-processing module for structural clarification of the forecast results. This module refines the boundaries of the wind risk field output by the AI ​​model, enhances its linear structure, and recovers local high-risk signals. This addresses the common problems of over-smoothing and spatial boundary ambiguity in traditional deep learning forecasts, thereby improving the accuracy, reliability, and interpretability of thunderstorm wind nowcasting warnings.

[0022] Improvement point 1: Construct a near-surface thermal disturbance precursor identification mechanism for squall line thunderstorms.

[0023] Traditional methods for identifying thunderstorms and strong winds primarily rely on radar echo intensity, historical wind observations, or single wind field characteristics. Warnings are often only issued when strong echoes have fully developed, downdrafts are approaching the ground, or strong surface winds are imminent, resulting in a short warning window. Furthermore, relying solely on strong echo or wind observations can easily misclassify ordinary severe convective processes as thunderstorms and strong winds, leading to a high false alarm rate.

[0024] This invention does not simply input surface temperature as a common meteorological variable into the model. Instead, it constructs a near-surface thermal disturbance precursor identification mechanism based on the cold pool outflow, gust front advancement, and rapid adjustment of near-surface thermal state before squall line thunderstorms. Specifically, in one embodiment, the changes in near-surface thermal observations between the current and historical times can be used to extract precursor features that characterize the enhancement of thermal disturbances. These precursor features are used to depict the evolution of near-surface disturbances caused by the expansion of cold pool outflow, the passage of gust fronts, and rapid changes in the thermal environment ahead of strong convective systems during squall line development.

[0025] Through the aforementioned near-surface thermal disturbance precursor identification mechanism, the model can capture thermal anomaly regions with enhanced potential for thunderstorm winds before strong echoes fully develop or surface strong winds actually occur. Unlike directly inputting raw temperature data, this invention emphasizes the extraction, organization, and constraint of near-surface thermal disturbance processes before thunderstorm winds occur, enabling the model to identify areas with squall line thunderstorm wind development potential in advance, rather than relying solely on already formed strong echoes or observed wind results.

[0026] Improvement point 2: Construct a near-surface dynamic disturbance precursor identification mechanism for squall line thunderstorms.

[0027] Thunderstorm winds, especially squall-type thunderstorm winds, are typically closely related to dynamic processes such as downdrafts within the storm, expansion of the cold pool outflow, and the advancement of gust fronts. When strong downdrafts reach the ground and spread outward, the near-surface dynamic state undergoes rapid adjustments, manifested as the advancement of the outflow boundary, enhanced low-level divergence, intensified surface disturbances, and the formation of thunderstorm high pressure. Therefore, the evolution of near-surface dynamic disturbances prior to the occurrence of thunderstorm winds is crucial precursor information for identifying the potential of squall-type thunderstorm winds.

[0028] This invention does not simply input surface air pressure as a common meteorological variable into the model. Instead, it constructs a near-surface dynamic disturbance precursor identification mechanism based on physical processes such as downdraft impact, cold pool outflow, and gust front advancement. Specifically, in one implementation, changes in near-surface dynamic correlation observations between the current and historical times can be used to extract precursor features that characterize the enhancement of dynamic disturbances. These precursor features are used to depict the advancement of thunderstorm outflow boundaries, the ground-reaching diffusion of downdrafts, and the rapid enhancement of near-surface disturbances.

[0029] Through the aforementioned near-surface dynamic disturbance precursor identification mechanism, the model can learn the nonlinear relationship between "near-surface dynamic disturbance enhancement" and "thunderstorm gale potential," thereby more effectively distinguishing between ordinary strong echo convection and squall line systems that may actually produce destructive surface winds. Unlike directly inputting raw air pressure data, this invention emphasizes the extraction, organization, and constraint of the dynamic disturbance process before thunderstorm gales occur, which helps reduce false alarms caused by relying solely on radar echo intensity and improves the reliability of near-term thunderstorm gale warnings.

[0030] Improvement point 3: Construct a fusion modeling mechanism for near-ground frontal body perturbation field with thermodynamic-dynamic coupling.

[0031] Traditional AI-based early warning methods often directly input variables such as temperature, air pressure, and wind speed into the model as ordinary meteorological elements, relying mainly on the model to learn the statistical relationships between multiple variables on its own. This lacks targeted extraction of key physical processes and their evolutionary characteristics before thunderstorms and strong winds occur. This invention does not simply superimpose multiple meteorological variables, but rather starts from the formation mechanism of squall line thunderstorms and strong winds, unifying near-surface thermal and dynamic disturbances into a precursor disturbance field that can characterize the potential for thunderstorm and strong wind formation.

[0032] Near-surface thermal disturbances are primarily used to characterize processes such as the influence of cold pool outflows, the passage of gust fronts, and the rapid adjustment of the thermal environment ahead of strong convective systems. Near-surface dynamic disturbances are primarily used to characterize processes such as the impact of descending airflows, thunderstorm high pressures, enhanced low-level divergence, and the advancement of outflow boundaries. These two processes do not exist in isolation but jointly reflect the evolutionary characteristics of enhanced near-surface thermal-dynamic coupling before the occurrence of squall line thunderstorms. By unifying these disturbance processes into a precursor disturbance field, the model can more accurately identify areas within the squall line system that truly possess surface wind potential, rather than simply equating strong echoes or ordinary strong convective clouds with thunderstorm winds.

[0033] This fusion modeling mechanism enhances the physical interpretability and operational usability of the model. When the model issues a high-risk warning for thunderstorms and strong winds, meteorological operators can trace the physical causes from the perspectives of enhanced near-surface thermal disturbances, strengthened dynamic disturbances, and their spatial synergistic development, thereby increasing their understanding and trust in the AI ​​warning results. In one specific implementation, the near-surface thermal disturbances and dynamic disturbances can be constructed from the short-term variation characteristics of surface thermal observations and dynamically related observations, respectively.

[0034] Improvement point four: Dynamic risk weight adjustment mechanism based on near-Earth frontal perturbation field.

[0035] Thunderstorms and strong winds are characterized by low probability but high impact. In terms of sample distribution, the number of strong wind samples is usually far less than that of non-strong wind samples. Traditional AI models are easily dominated by the majority class samples during training, resulting in insufficient recognition of strong wind events. That is, the model tends to predict "no strong wind" or low-risk conditions, thus causing missed reports.

[0036] To address this issue, this invention proposes a dynamic risk weight adjustment mechanism based on the near-surface forebody disturbance field. This mechanism does not apply fixed weights to all samples or regions, nor does it simply weight them based on a single meteorological variable. Instead, it dynamically adjusts the model's focus on different regions and samples based on the intensity of near-surface thermal-dynamic disturbances, spatial coordination characteristics, and the degree of matching with strong convective structures characterized by the forebody disturbance field. When a region exhibits significant enhancement of near-surface disturbances, outflow boundary advancement, low-level dynamic anomalies, or high-risk characteristics corresponding to strong echo structures, the model automatically increases the risk weight of that region or sample, giving it greater attention during training or prediction.

[0037] This mechanism essentially constructs a dynamic "thunderstorm and gale risk weight map" for the model. Under normal weather conditions, the model maintains a normal response; however, when the current volume disturbance field indicates that a certain area is under strong disturbance and high-risk conditions, the model will increase its sensitivity to identify that area and give a stronger response to areas that may develop into thunderstorms and gales. This can alleviate the model inertia problem caused by imbalanced samples, improve the ability to detect low-probability thunderstorm and gale events, and reduce the false negative rate.

[0038] Meanwhile, the dynamic risk weight adjustment mechanism does not significantly increase the structural complexity of the main prediction model, nor does it rely on a large number of additional network branches. Instead, it can dynamically adjust the model's attention at the level of feature input, attention adjustment, or loss function constraints. Therefore, this mechanism can improve early warning performance while maintaining a relatively simple model structure, high computational efficiency, and reducing the risk of overfitting.

[0039] Improvement point 5: Introduce LDM as a post-processing module for forecast result clarity.

[0040] Existing AI short-term forecasting models often suffer from issues such as smoothed spatial structure, blurred boundaries, and weakened local extrema in their output of thunderstorm and gale risk fields. This is especially true for squall line systems, where gale risk typically exhibits a banded, linear, or locally concentrated distribution along gust fronts, cold pool outflow boundaries, or strong convection zones. If the model output is too smooth, it becomes difficult to accurately depict the spatial organization of squall line gales, thus affecting the determination of the affected area and the identification of risk boundaries in operational early warning systems.

[0041] To address this issue, this invention further introduces a latent diffusion model (LDM) as a clarity post-processing procedure. LDM can learn the spatial distribution characteristics of the thunderstorm wind risk field in the latent space, enhancing the structure and restoring details of the blurred results output by the original AI model. This post-processing module can strengthen the boundary clarity of squall line wind risk areas, improve the spatial continuity of linear convective systems, and reduce the weakening of risk signals caused by excessive smoothing in forecast results.

[0042] Unlike directly increasing the complexity of the master forecasting model, the LDM, as a post-processing module, primarily refines the spatial representation of the model output without altering the core logic of precursor signal identification and gale probability prediction. Therefore, this invention retains the physical interpretability based on temperature and pressure precursor signals while utilizing a generative model to improve the spatial representation quality of the forecast results. This design contributes to obtaining clearer, more continuous, and more accurate gale risk forecast products that closely resemble actual squall line structures.

[0043] Improvement point six: Improve the timeliness of early warnings, reduce false alarms and missed reports, and enhance the value of business applications.

[0044] In summary, the core innovations of this invention are as follows: First, starting from the physical mechanism of squall line thunderstorms and strong winds, temperature difference and pressure difference precursor signals are constructed to identify key processes such as cold pool outflow, gust front advancement, downdraft impact, and near-surface disturbance enhancement in advance. Second, an adaptive weighting mechanism is used to enable the model to dynamically adjust the degree of attention to high-risk samples and high-risk areas according to the strength of the precursor signals, thereby improving the ability to identify low-probability thunderstorm and strong wind events. Third, LDM is introduced as a clarity post-processing module to refine the spatial structure and enhance the boundaries of the strong wind risk field output by the model, thereby improving the readability and operational usability of the forecast products.

[0045] Compared with traditional methods, this invention can not only capture thermal and dynamic anomalies before squall line thunderstorms and strong winds occur, extending the warning time, but also reduce the false alarm and missed alarm rates through precursor signal constraints and adaptive weighting mechanisms. Simultaneously, the LDM post-processing module further improves the ambiguity of AI forecast results, making the spatial distribution of thunderstorm and strong wind risk areas clearer, more continuous, and physically consistent. Therefore, this invention represents a significant technological advancement in short-term warnings of thunderstorms and strong winds, squall line process identification, refined forecasting of risk areas, and operational applications. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in this 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 some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0047] Figure 1 This is a flowchart illustrating the squall line thunderstorm gale prediction method provided by the present invention.

[0048] Figure 2 This is a flowchart illustrating the quality control process for radar-based data in the squall line thunderstorm gale near-term prediction method provided by this invention.

[0049] Figure 3 This is a schematic diagram of the inversion and interpolation process in the squall line thunderstorm gale prediction method provided by the present invention;

[0050] Figure 4 This is a schematic diagram of the deep learning prediction network structure in the squall line thunderstorm gale near-term prediction method provided by the present invention.

[0051] Figure 5 This is a schematic diagram of the squall line thunderstorm gale proximity prediction device provided by the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0053] The following is combined Figure 1 A method for near-term forecasting of squall line thunderstorms according to the present invention includes:

[0054] S1, after quality control of the radar base data, extract dual-polarization radar features from it;

[0055] S2, based on radar radial velocity, uses the least squares fitting method to invert radar radial wind vorticity and divergence at different radial distance scales, as low-level dynamic background features;

[0056] S3, select ground observation samples from near-ground antecedent disturbance observation data that match radar observation records in time and space, and construct near-ground antecedent disturbance field;

[0057] S4. Based on the dual-polarization radar characteristics, low-level dynamic background characteristics, and near-ground frontal disturbance field, a deep learning prediction network is used to obtain the near-term prediction results of squall-type thunderstorm winds.

[0058] S5. The prediction results of the approaching thunderstorm and strong wind are enhanced and post-processed based on the potential diffusion model.

[0059] Dual-polarization radar technology plays a crucial role in monitoring and providing imminent warnings for thunderstorms and strong winds. The formation of thunderstorms and strong winds is typically closely related to strong downdrafts within the storm, precipitation particle drag, cold pool outflow, and near-surface divergence. Dual-polarization radar can capture the microphysical and dynamic characteristics associated with these processes, such as the large number of precipitation particles indicated by high reflectivity factor regions, particle size and phase changes reflected by differential reflectivity and differential phase shift rate, and forecasting signals like mid-to-low-level convergence, echo core downdraft, and rapid evolution of strong echo structures. When the downdrafts within the storm intensify and reach the ground, the airflow rapidly spreads outward, easily forming destructive winds near the ground. Therefore, relying solely on a single radar echo intensity or a single wind field characteristic is insufficient to reliably distinguish between ordinary strong convective processes and thunderstorm systems with true surface wind potential.

[0060] During squall line thunderstorms, the downdrafts within the storm, the outflow of cold pools, and the advance of gust fronts often cause rapid disturbances in the near-surface thermodynamic state. These disturbances should not be simply interpreted as inputs of conventional meteorological elements, but rather as a comprehensive response to the storm's outflow boundary, the impact of downdrafts, and the enhancement of near-surface disturbances before the onset of thunderstorms. Based on this physical understanding, this embodiment combines the strong convective organization structure identified by dual-polarization radar with the near-surface forebody disturbance process reflected in ground observations to construct a multi-source forebody disturbance field oriented towards the potential for thunderstorm occurrence, providing physical constraints for subsequent risk identification and weight adjustment.

[0061] This embodiment first performs data collection, including:

[0062] Radar observation data: Observational variables related to the structure and microphysical processes of severe convective systems, including the horizontal reflectivity factor Z, were extracted from dual-polarization weather radar. H , differential phase shift rate K DP Differential reflectivity Z DR Differential propagation phase constant, correlation coefficient ρhv, etc., are used to characterize the features of precipitation particles inside the storm, the structure of strong echoes, and their evolution.

[0063] Ground observation data: Observational data related to the occurrence of thunderstorms and strong winds and their near-surface disturbance processes were collected using ground meteorological observation stations. Among them, the maximum gust observation data was used to construct samples or validate samples of thunderstorms and strong winds; the near-surface disturbance-related observations were used to characterize the precursor processes such as cold pool outflow, gust front advancement, downdraft impact, and rapid adjustment of near-surface thermodynamic state before the occurrence of thunderstorms and strong winds.

[0064] Then, data preprocessing is performed, including:

[0065] Radar data preprocessing: The raw dual-polarization radar data is subjected to quality control, bias correction, phase processing, attenuation correction and clutter suppression, and the radar microphysics and echo structure variables after quality control are output.

[0066] Calculation of low-level dynamic background features: Based on radar radial velocity data, the LLSD (Linear Least Squares Derivative) method is used to extract dynamic background features such as convergence, divergence and rotation in the low-level wind field to describe the low-level dynamic environment of strong convective systems, rather than relying solely on these variables for wind speed prediction.

[0067] Construction of near-surface frontal disturbance field: Temporal and spatial matching of ground observation data and radar observation data is performed to extract the changing characteristics that reflect the enhancement process of near-surface disturbance before the occurrence of thunderstorms and strong winds, and a near-surface frontal disturbance field consistent with the radar grid is generated by interpolation method.

[0068] Data interpolation and grid unification: Radar variables, low-level dynamic background features, near-surface frontal disturbance field, and gust sample labels are uniformly interpolated or mapped to a preset horizontal Cartesian coordinate grid to ensure consistency of multi-source data in terms of spatial resolution and coverage.

[0069] Normalization: The above multi-source variables are normalized, usually by scaling the variable values ​​to the range of [0,1] or [-1,1], in order to improve the stability of model training and the comparability between different variables.

[0070] In the process of constructing a strong convective wind identification and early warning model, this embodiment can use a deep learning prediction network with multi-source spatiotemporal feature extraction capability as the main prediction framework to carry the joint modeling of multi-source information such as radar microphysical structure, low-level dynamic background features and near-ground antecedent disturbance field, and extract the nonlinear correlation between radar microphysical features, dynamic background features and near-ground antecedent disturbance field.

[0071] The main network can be implemented using FURENet or other spatiotemporal prediction networks with an encoder-decoder structure. Its role is to extract the nonlinear correlation between the internal structural evolution of strong convective systems and the enhancement of near-surface disturbances, and to carry out the risk identification process driven by the precursor disturbance field, rather than simply superimposing existing model modules.

[0072] Specifically, the main prediction network captures radar echo morphology, intensity gradient, strong echo core evolution, and dual-polarization microphysical features through a spatial feature extraction structure, and combines this with a temporal modeling unit to learn the dynamic changes of the strong convective system over time. During multi-source feature fusion, the model performs branch extraction and deep fusion of radar reflectivity, dual-polarization parameters, low-level dynamic background field, and near-surface antecedent disturbance field to avoid feature interference caused by direct mixing of different modal information in the early stages. Through channel recalibration or attention adjustment mechanisms, the model can highlight key features closely related to the potential for thunderstorm and strong wind occurrence, while suppressing redundant or weakly correlated information.

[0073] The model learns the spatial semantic features of radar echo structure, dynamic background field and near-surface disturbance field through a multi-branch feature extraction structure, and fuses them in a deep network, enabling the model to simultaneously focus on strong echo organization inside the storm, low-level convergence and divergence features and near-surface disturbance enhancement regions.

[0074] To further improve the problems commonly found in deep learning forecasts, such as smoothing of spatial structure, blurring of boundaries, and weakening of local strong signals, a Latent Diffusion Model (LDM) is introduced as a de-sharpening post-processing module after the main prediction model outputs the initial thunderstorm and gale risk field.

[0075] LDM does not directly replace FURENet's main prediction function. Instead, it performs post-processing optimization on the wind risk field or forecast results output by FURENet. By learning the spatial structural characteristics of thunderstorm wind risk areas in the latent space, it enhances the details of blurred forecast boundaries, linear strong convective structures, and locally high-risk areas in the initial forecast results. It also restores and strengthens linear, banded, or locally concentrated risk structures that have been smoothed and weakened. Especially for squall-type thunderstorms, wind risk often presents a banded or linear distribution along gust fronts, cold pool outflow boundaries, or strong convective bands. Ordinary deep learning model outputs are prone to problems such as unclear boundary transitions, discontinuous linear structures, and weakened locally strong signals.

[0076] Introducing LDM post-processing can further enhance the linear and zonal strong convective structure characteristics in the thunderstorm and gale risk field, improve the spatial continuity and structural clarity of the forecast results, highlight local high-risk areas and their spatial distribution details, and make the optimized wind speed field or risk field closer to the actual organizational form of the strong convective system, thereby improving the precision and operational value of the thunderstorm and gale nowcasting results.

[0077] Through the above steps, the deep learning method in this embodiment combines radar data, physical information, and deep learning models to achieve high-precision and timely thunderstorm and gale forecasts, providing important support for emergency response and disaster early warning.

[0078] Based on the above embodiments, this embodiment performs quality control on radar base data, including:

[0079] The differential reflectivity in the radar base data is corrected for deviations using the vertical scanning method, the drizzle area method, or the raindrop spectrometer-assisted method.

[0080] The differential propagation phase in the radar base data is corrected by initial deviation estimation, phase defolding, noise suppression and smoothing.

[0081] Based on noise suppression, a variable step-size least squares method is used to robustly estimate the differential phase in the radar base data. At the same time, a linear programming method is introduced for the rainfall area. Combined with the physical constraints of the differential phase derived from the reflectivity factor and the differential factor, the optimal differential propagation phase sequence that satisfies the monotonically increasing characteristic is reconstructed. The differential phase is calculated through a differential filter to ensure that the differential phase is non-negative and conforms to the physical laws of precipitation.

[0082] Based on T-matrix scattering simulation and raindrop spectrum observation, the relationship between specific attenuation, specific differential attenuation and differential phase is established, thereby determining the attenuation characteristic parameters of the dual polarization parameters. Based on the attenuation characteristic parameters, an attenuation correction strategy is adopted to correct the reflectivity factor and differential reflectivity in the radar base data.

[0083] By utilizing the consistency characteristics of dual polarization parameters, combined with spectral width and phase variability, non-meteorological echoes in the radar base data are identified and eliminated.

[0084] Radar-based data quality control involves a complete quality control process for the acquired radar observation data, covering several key aspects such as bias correction, phase processing, attenuation compensation, and clutter suppression. For example... Figure 2 As shown, the entire process is divided into multiple functional modules, including: Z DR Deviation correction module, Φ DP Deviation correction module, Φ DP Processing with K DP Estimation module, reflectivity factor Z H With differential reflectivity Z DR The attenuation correction module and clutter suppression module.

[0085] In differential reflectivity (Z) DR In the deviation correction, a comprehensive approach was adopted, combining the light rain method, vertical pointing scan data, and statistical methods for weak precipitation areas below the zero-degree bright band, along with zero Z-axis data. DR The characteristic assumptions and raindrop spectrometer observation data were used to evaluate and calibrate the systematic bias.

[0086] Z DRThe deviation correction module uses observation data under light rain conditions to calibrate the system deviation introduced by the receiving channel. Common calibration methods mainly include the following three:

[0087] Vertical scanning method: When the radar has vertical pointing scanning capability, it can collect vertical observation data during stratiform cloud precipitation. Data is selected below the zero-degree layer height and with a high correlation coefficient (e.g., ρ). hv For regions with a concentration >0.99 and weak precipitation intensity, calculate the differential reflectance (Z0.99) of that region. DR The average value of the radar system is calculated. The deviation between this average and the theoretically expected value represents the system's deviation.

[0088] Drizzle area method: Based on Z in drizzle DR Taking advantage of the near-zero theoretical value, a region covered by drizzle was selected within the radar detection range, and the Z-value of that region was extracted. DR The average deviation of the observed values ​​is used as an estimate of the systematic bias.

[0089] Raindrop spectrometer-assisted method: Combining raindrop spectrometer data deployed within a 50km radius of the radar, the corresponding polarization parameters are simulated using the T-Matrix scattering model. By comparing this data with actual observations, the radar Z-axis is retrieved. DR Systematic bias.

[0090] This method can be used for the original Z DR The data was systematically corrected.

[0091] In differential propagation phase (Φ DP Processing and differential phase shift rate (K) DP In the estimation, the physical consistency of the KDP estimation is optimized by filtering and denoising, variable step size least squares fitting and linear programming methods.

[0092] Φ DP Deviation correction mainly includes the following steps:

[0093] Initial bias estimation: Using radar vertical scan data, in the region below the zero-degree layer of stratiform cloud precipitation, select regions with high correlation coefficients (ρ). hv >0.99) and statistical Φ of areas with weak precipitation DP The value is used to obtain the initial deviation of the system; simultaneously, based on Φ DP The physical characteristics accumulated from zero in the rainfall area, combined with multi-azimuth observations of the nearby precipitation area, are used to statistically obtain the initial phase deviation of the system, which is then fed back to the quality control module for initial phase correction.

[0094] Phase defolding: using Φ DP Spatial continuity is used to identify and correct phase folding caused by phase range limitations (such as ±180°) in signal processing.

[0095] Noise suppression and smoothing: Median filtering, mean filtering, and other methods are used to suppress Φ. DP Random noise in the data is eliminated, and phase jitter is further eliminated by combining FIR (Finite Impulse Response) iterative filtering, thereby improving data stability.

[0096] K DP The optimization estimation steps mainly include: based on noise suppression, using the variable step size least squares method to estimate K. DP Robust estimation is performed; simultaneously, a linear programming method is introduced in the rainfall area, combined with the reflectivity factor Z. H and differential reflectivity Z DR Exported K DP Physical constraints, reconstruct the optimal Φ that satisfies the monotonically increasing property. DP The sequence is then processed by a differential filter to calculate K. DP This ensures that the results are non-negative and conform to the physical laws of precipitation.

[0097] Reflectivity factor (Z) H The attenuation correction for ZDR is based on T-matrix scattering simulations and raindrop spectral observations, establishing the relationship between specific attenuation and K. DP The empirical relationship between them is determined, and path integral compensation is performed using linear or Z-PHI methods. The core idea of ​​the Z-PHI method is to utilize the propagation phase shift rate (K0). DP ) or its integral total propagation phase shift (Φ DP Path integral compensation is performed using a dual polarization parameter (such as Z) where Z represents the radar reflectivity factor and PHI represents the propagation phase shift. For clutter suppression, dual polarization parameters (such as Z) are used. DR K DP Correlation coefficient ρ hv Based on the consistency characteristics of spectral width and phase variability, a multivariate discrimination criterion is constructed to identify and eliminate non-meteorological echoes such as ground features and organisms.

[0098] The steps for determining the attenuation characteristic parameters include: calculating the scattering characteristics of raindrops of different sizes in the radar operating band based on the T-Matrix method; deriving the polarization variable by combining the observation data from the two-dimensional video raindrop spectrometer; and analyzing the specific attenuation AH and the specific differential attenuation A. DP With K DP The relationship between these parameters is used to determine the attenuation characteristic parameters of the dual polarization parameters at this radar wavelength.

[0099] Based on the characteristics of the radar's operating band, the following two attenuation correction strategies are adopted for the reflectivity factor (Z). H ) and differential reflectance (Z DR Correction is required:

[0100] First, linear correction and the Z-PHI method were used to process the data. The results show that, behind areas of heavy precipitation, the data corrected by these two methods outperformed the original signal processing algorithm in terms of both physical consistency and data reliability.

[0101] In the quantitative application of dual polarization variables (such as Z) DR and K DP When dealing with objects such as meteorological targets (e.g., ground features, organisms), a high degree of consistency between the observations of the two radar channels is required. However, the echoes from non-meteorological targets (e.g., ground features, organisms) typically exhibit poor consistency between the two channels, resulting in insufficient accuracy of the polarization parameters obtained from their inversion, making it difficult to meet the requirements of quantitative analysis.

[0102] Therefore, before performing quantitative inversion, unreliable dual-polarization data must be identified and removed using a clutter suppression module. Typical clutter characteristics include: dual-channel correlation coefficient (ρ... hv Low, differential phase (Φ) DP The variance is relatively large; the secondary echo region is often accompanied by an abnormal increase in spectral width; while at the same reflectivity factor (Z... H Under these conditions, the Z-wavelength of echoes from insects and other organisms... DR The value is usually significantly higher than that of precipitation echo.

[0103] By comprehensively utilizing various dual polarization variables and their derived parameters (such as ρ) hv Φ DP Variance, spectral width, Z DR -Z H By employing a multi-parameter optimal combination strategy (such as considering relationships), the system can effectively filter out the vast majority of unreliable non-meteorological echoes, thereby significantly improving the accuracy and reliability of subsequent quantitative applications.

[0104] Based on the above embodiments, in this embodiment, after quality control is completed, five core dual-polarization parameters are extracted from the preprocessed radar data cube, including: horizontal reflectivity factor (Z). H ), differential reflectance (Z) DR ), differential phase shift rate (K) DP ), correlation coefficient (ρ) hv ) and differential propagation phase constant (dBT).

[0105] To adapt to the application requirements of numerical models and construct a standardized dataset, the aforementioned dual polarization parameters were further resampled to a horizontal Cartesian grid using spatial interpolation methods. The default grid coverage is 400 km × 400 km, with a horizontal resolution of 1 km. Considering computational resource constraints and model input requirements, the actual interpolation area is centered on the radar site, with a radius of approximately 128 km, corresponding to a grid size of 256 km × 256 km. Simultaneously, to provide the dataset with a complete spatiotemporal reference framework, the latitude and longitude coordinates of the radar center for each volume scan were extracted synchronously, and the scan start time was used as the unified time identifier for the dataset.

[0106] Based on the above embodiments, this embodiment first uses bilinear interpolation to resample the radar radial velocity data into a Cartesian coordinate system. In this coordinate system, the distance *r* and azimuth angle *φ* of each grid point relative to the radar center are calculated. Defining the variable *s = rφ*, where *r* is the distance from the target to the radar and *φ* is its azimuth angle, the radar radial velocity at distance *r* and position *s* can be expressed as:

[0107] ;

[0108] Where r is the distance of the grid point relative to the radar center, and s is the position of the grid point. The radar radial velocity estimate for a grid point at a location s and a distance r relative to the radar center. Represents the known observed radial velocity. and These represent the partial derivatives of the radar radial velocity in the radial and azimuth directions, respectively, which are the wind shear in the corresponding directions.

[0109] Based on the above embodiments, the objective (cost) function inverted by the least squares fitting method in this embodiment, constructed from the estimated radial velocity and the actual observed values, is as follows:

[0110] ;

[0111] in, Let be the objective function. The distance relative to the radar center is 1 And the position is Radial velocity observed by radar at the grid points, The distance relative to the radar center is 1 And the position is Radar radial velocity estimates at the grid points, For locally symmetric weighting functions, The value of decreases as the distance between the grid point and the local coordinate center increases, where i and j are the local grid point indices.

[0112] By minimizing the value function, the radial shear u can be solved. r and azimuth shear u s .

[0113] Unlike traditional methods that rely on changes in velocity between adjacent distances to estimate azimuth shear, the advantage of using local functionals to estimate vorticity lies in its ability to fully utilize radial velocity information across the entire local area. Therefore, even with single-point or multi-point missing measurements, or interference from noise, this method maintains good robustness. r and u s These represent radial shear and azimuth shear, respectively, i.e., divergence along the radar radial direction and vorticity along the azimuth direction, u s The magnitude of the divergence directly reflects the strength of the vorticity. Finally, the divergence and vorticity in the Cartesian coordinate system are re-interpolated to the polar coordinate system centered on the radar to facilitate subsequent analysis and application. Figure 3 The main steps of the above inversion and interpolation process are shown.

[0114] After completing the radar base data quality control, a least-squares fitting inversion method is used to extract low-level dynamic background features. Specifically, the radial shear and azimuth shear in the radial velocity field are estimated through local least-squares fitting. The radial shear is used to characterize the convergence or divergence characteristics along the radar radial direction, and the azimuth shear is used to characterize the rotation or vortex characteristics along the azimuth direction. Through this inversion process, the following can be obtained: Four dynamic background variables represent the estimated azimuth and radial shear characteristics at radial distance scales of 500 m and 1500 m, respectively. These variables are mainly used to characterize the convergence, divergence, and rotation structure in the lower-level wind field of strong convective systems, providing dynamic background constraints for subsequent thunderstorm and gale risk identification.

[0115] It should be noted that this embodiment does not rely solely on vorticity and divergence variables for wind speed prediction, but rather uses them as a characterization of the low-level dynamic environment, which, together with the near-surface frontal disturbance field and radar microphysical structure characteristics, is used for thunderstorm wind potential identification.

[0116] Subsequently, spatial interpolation techniques were used to resample the aforementioned dynamic background features to a horizontal Cartesian grid with a coverage area of ​​400 km × 400 km and a horizontal resolution of 1 km. Considering the model input requirements and computational resource limitations, the analysis area was ultimately cropped into a square region with a side length of 256 km centered on the radar site, i.e., the area within approximately 128 km of the radar center.

[0117] Based on the above embodiments, the near-ground precursor disturbance observation data in this embodiment includes temperature differential precursor signals constructed based on temperature observations at the current time and multiple historical times, pressure differential precursor signals constructed based on pressure observations at the current time and multiple historical times, and the maximum wind speed observation results at the current time.

[0118] Near-Earth Antecedent Disturbance Field Construction and Gust Observation Sample Matching: Based on the latitude and longitude coordinates of the radar center and its corresponding timestamp information, when constructing the near-Earth antecedent disturbance field, it is necessary to extract near-surface observation data that matches the radar observation records in time and space from ground observation stations. Specific selection criteria are as follows: in terms of time, the observation time must be consistent with the radar timestamp or within the allowable error range; in terms of space, the distance between the observation station and the radar center must not exceed 100 kilometers.

[0119] Near-surface disturbance correlation observation data are used to characterize the evolution of precursor disturbances before thunderstorms and strong winds, with spatial coverage matched to gust observation data. Gust observation data is used to construct samples or validate samples for thunderstorms and strong winds, with an observation interval of 5 minutes and a spatial selection range within 100 kilometers of the radar station. Near-surface disturbance correlation observation data are used to characterize the evolution of precursor disturbances before thunderstorms and strong winds, such as ground thermal and dynamic disturbance correlation quantities, with an observation interval of 6 minutes and spatial coverage matched to gust observation data.

[0120] Based on the above embodiments, the custom interpolation strategy used in this embodiment is the Rbf class, based on the i-th two-dimensional spatial point (lat). i lon i ) and its corresponding precursor perturbation eigenvalue y i Construct a spatial interpolation function. This function is represented by a linear combination of radial basis functions:

[0121] ;

[0122] in, It is a linear combination of radial basis functions. Let x be the coordinates (latitude lat and longitude lon) of the i-th known data point, x be the coordinates of the data point to be interpolated, and n be the number of known data points. For radial basis functions, The weights are to be determined by solving a system of linear equations.

[0123] The specified radial basis function type is "multiquadric", and its mathematical form is:

[0124] ;

[0125] ;

[0126] in, Euclidean distance. It is a shape parameter that controls the decay rate or width of the radial basis function. Smaller Make the basis functions sharper and larger This makes the basis functions flatter and the interpolation surface smoother. A value of 0.1 here means the basis functions change more rapidly at smaller distances. It also results in a closer fit to the data points. When the smoothing parameter > 0, the interpolation curve is allowed not to completely pass through all data points, but rather to provide a smooth approximation, suitable for noisy observations. The larger the value, the smoother the fitted surface. A value of 0.01 here indicates slight smoothing, balancing data approximation and noise suppression. The Multiquadric function is a globally supported basis function with good smoothing and extrapolation capabilities.

[0127] Based on the filtered sparse station observation data, the changing characteristics reflecting the evolution of near-surface disturbances before thunderstorms and strong winds are extracted, and mapped onto a preset horizontal Cartesian coordinate grid using a custom interpolation strategy. This grid covers an area of ​​256 km × 256 km with a spatial resolution of 1 km. Through this processing, a gridded gust sample and near-surface forebody disturbance field consistent with the spatial scale of radar observations can be generated, making it suitable for subsequent forebody disturbance identification, dynamic risk weight calculation, and input into thunderstorm and strong wind prediction models.

[0128] Based on the above embodiments, this embodiment also includes:

[0129] Based on the near-surface anterior disturbance field, regions with obvious disturbance characteristics are identified, including regions with enhanced near-surface disturbances, low-level dynamic anomalies, or regions that match strong convective organization structures.

[0130] In the training loss or prediction response of the deep learning prediction network, the attention weight of regions with obvious perturbation features is increased, while the normal response is maintained for regions with indistinct perturbation features.

[0131] To address the model recognition inertia caused by the scarcity of thunderstorm and strong wind samples and the high proportion of non-strong wind samples, this embodiment does not simply use fixed category weights, but instead constructs a dynamic risk weight adjustment mechanism based on the near-ground forebody perturbation field to adaptively enhance areas where thunderstorms and strong winds may occur.

[0132] When the current volume disturbance field shows that a certain region has enhanced near-surface disturbance, low-level dynamic anomalies, or high-risk characteristics matching strong convective organization, the model increases the weight of that region or sample in the training loss or predicted response; when the current volume disturbance characteristics are not obvious, the model maintains a normal response.

[0133] Through this dynamic risk weighting mechanism, the model can enhance its ability to identify low-probability, high-impact thunderstorm and gale events, reduce the risk of missed reports, and thus more effectively identify potential risk areas of squall-type thunderstorms and gales, improving the accuracy, stability, and lead time of strong convective gale warnings.

[0134] This invention starts from the physical processes preceding squall-type thunderstorms, such as cold pool outflow, gust front propulsion, downdraft impact, and near-surface disturbance enhancement, to construct a precursor disturbance field that can characterize the potential for thunderstorms. Based on this precursor disturbance field, dynamic risk weights are generated to enhance the model's ability to identify high-risk areas and low-probability strong wind samples.

[0135] The construction of the near-surface forebody disturbance field is one of the key improvements of this method. This embodiment does not simply overlay historical and current ground observation data with radar data; instead, it extracts forebody disturbance features reflecting rapid changes in the near-surface thermodynamic state, starting from physical processes such as cold pool outflow, gust front advancement, downdraft impact, and near-surface disturbance enhancement before thunderstorms and strong winds. These features are then fused with radar microphysical structure characteristics and low-level dynamic background features. Through this processing, the model can more sensitively identify localized strong disturbance areas and their correspondence with strong convective organization structures, thereby improving its ability to identify potential thunderstorm and strong wind areas. This forebody disturbance information not only provides the model with physical constraints beyond ordinary meteorological variables but can also be further used for dynamic risk weight adjustment, enabling the model to pay stronger attention to high-risk areas during training and prediction. Therefore, this method can better adapt to the early warning needs of thunderstorms and strong winds, which are characterized by low probability, suddenness, and strong locality. In summary, this embodiment combines dual-polarization radar, physical information, and neural networks to form a high-precision and high-time-efficiency thunderstorm and gale forecasting system.

[0136] The selected model architecture is FURENet ( Figure 4 Its core design is based on the classic UNet convolutional neural network, which serves as the backbone framework. Given the input data resolution of 256×256 and the network setting of 31 layers (including 7 downsampling and 7 upsampling), FURENet fully leverages the advantages of UNet's inherent encoder-decoder symmetric structure in processing multi-scale information streams.

[0137] The network comprises a 7-level encoder / decoder module, which progressively compresses the spatial size from 256×256 to 16×16 through several downsampling operations with a stride of 2. This process achieves feature extraction from shallow to deep layers: shallow layers preserve high-resolution details at small scales (such as abrupt changes in local meteorological elements), while deep layers capture low-resolution global semantic information (such as large-scale circulation patterns) through a large receptive field. Subsequently, the decoder performs symmetrical upsampling using transposed convolutions and introduces skip connections. This mechanism directly concatenates high-resolution features from each layer of the encoder to the corresponding layer of the decoder, effectively compensating for the fine spatial structure lost during downsampling and ensuring that the final output accurately reconstructs the 256×256 high-resolution prediction field.

[0138] Furthermore, to address the core challenge of fusing multivariate meteorological physical field information, FURENet embeds a high-dimensional feature fusion technology squeeze-excitation module (SE-Block) after each level of the UNet backbone's encoding module. This module focuses on uncovering the intrinsic correlations between feature channels. Its operating mechanism is as follows: First, a 16×16 feature map is compressed into a 1×1×C channel descriptor through global average pooling; then, the nonlinear dependencies between channels are learned through a bottleneck structure composed of two fully connected layers, outputting a set of normalized weight coefficients; finally, the original feature map is recalibrated through channel-by-channel multiplication. This mechanism enables the network to dynamically strengthen feature channels that play a crucial role in the current task (such as specific physical variables or latent modes) based on the data context, while suppressing redundant information. Thanks to the spatial dimension compression of global pooling, this module introduces a very small number of parameters in the 7-layer network, yet significantly improves the model's discriminative feature extraction capability in complex meteorological data.

[0139] In summary, FURENet, through UNet's multi-scale feature extraction and reconstruction capabilities combined with SE-Block's lightweight channel attention mechanism, provides a solution that balances accuracy and efficiency for processing multivariable, multi-source meteorological data.

[0140] This invention establishes a two-stage optimization framework: "FURENet main prediction – LDM clarification post-processing". FURENet is responsible for extracting dynamic, thermodynamic, and microphysical features related to thunderstorm and gale formation from multivariate meteorological inputs and completing the initial prediction of the risk field. LDM further enhances the spatial structure and restores details of the prediction results. The combination of these two approaches improves the refined representation of thunderstorm and gale risk areas while ensuring prediction accuracy, providing clearer, more stable, and interpretable prediction products for short-term thunderstorm and gale warnings.

[0141] The following describes the squall line thunderstorm gale prediction device provided by the present invention. The squall line thunderstorm gale prediction device described below can be referred to in correspondence with the squall line thunderstorm gale prediction method described above.

[0142] like Figure 5 As shown, the device includes an extraction module 501, an inversion module 502, a screening module 503, a prediction module 504, and a post-processing module 505, wherein:

[0143] The extraction module 501 is used to extract dual-polarization radar features from the radar base data after quality control.

[0144] Inversion module 502 is used to invert radar radial wind vorticity and divergence at different radial distance scales based on radar radial velocity using the least squares fitting method, as low-level dynamic background features.

[0145] The screening module 503 is used to screen ground observation samples that match the radar observation records in time and space from the near-ground antecedent disturbance observation data, and to construct the near-ground antecedent disturbance field;

[0146] Prediction module 504 is used to obtain the near-term prediction results of thunderstorm and strong winds based on the dual-polarization radar features, low-level dynamic background features and near-ground frontal disturbance field using a deep learning prediction network.

[0147] The post-processing module 505 is used to enhance the thunderstorm gale proximity prediction results based on the potential diffusion model.

[0148] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for near-term forecasting of squall line thunderstorms and strong winds, characterized in that, include: After quality control of the radar base data, dual-polarization radar features are extracted from it; Based on radar radial velocity, the radar radial wind vorticity and divergence at different radial distance scales are obtained by least squares fitting method, which serve as low-level dynamic background features. Ground observation samples that match radar observation records in time and space are selected from near-ground frontal body disturbance observation data to construct near-ground frontal body disturbance field; Based on the dual-polarization radar features, low-level dynamic background features, and near-ground frontal disturbance field, a deep learning prediction network is used to obtain the near-term prediction results of thunderstorms and strong winds. The thunderstorm gale proximity prediction results are enhanced and post-processed based on the potential diffusion model.

2. The squall line thunderstorm gale prediction method according to claim 1, characterized in that, Quality control of radar base data includes: The differential reflectivity in the radar base data is corrected for deviations using the vertical scanning method, the drizzle area method, or the raindrop spectrometer-assisted method. The differential propagation phase in the radar base data is corrected by initial deviation estimation, phase defolding, and noise suppression and smoothing. Based on noise suppression, a variable step-size least squares method is used to robustly estimate the differential phase in the radar base data. At the same time, a linear programming method is introduced for the rainfall area. Combined with the physical constraints of the differential phase derived from the reflectivity factor and the differential factor, the optimal differential propagation phase sequence that satisfies the monotonically increasing characteristic is reconstructed. The differential phase is calculated through a differential filter to ensure that the differential phase is non-negative and conforms to the physical laws of precipitation. Based on T-matrix scattering simulation and raindrop spectrum observation, the relationship between specific attenuation, specific differential attenuation and differential phase is established, thereby determining the attenuation characteristic parameters of the dual polarization parameters. Based on the attenuation characteristic parameters, an attenuation correction strategy is adopted to correct the reflectivity factor and differential reflectivity in the radar base data. By utilizing the consistency characteristics of dual polarization parameters, combined with spectral width and phase variability, non-meteorological echoes in the radar base data are identified and eliminated.

3. The squall line thunderstorm gale prediction method according to claim 1, characterized in that, The characteristics of the dual-polarization radar include horizontal reflectivity factor, differential reflectivity, differential phase shift rate, correlation coefficient, and differential propagation phase constant.

4. The squall line thunderstorm gale prediction method according to claim 1, characterized in that, The radar radial velocity is expressed as: ; Where r is the distance of the grid point relative to the radar center, and s is the position of the grid point. The radar radial velocity estimate is given at a grid point at a distance r relative to the radar center and a position s. Represents the known observed radial velocity. and These represent the partial derivatives of the radar radial velocity in the radial and azimuth directions, respectively.

5. The squall line thunderstorm gale prediction method according to claim 4, characterized in that, The objective function obtained by the least squares fitting method is: ; in, Let be the objective function. The distance relative to the radar center is 1 And the position is Radial velocity observed by radar at the grid points, The distance relative to the radar center is 1 And the position is Radar radial velocity estimates at the grid points, For locally symmetric weighting functions, The value of decreases as the distance between the grid point and the local coordinate center increases, where i and j are the local grid point indices.

6. The squall line thunderstorm gale prediction method according to claim 1, characterized in that, The near-ground forebody disturbance observation data includes temperature differential forebody signals constructed based on temperature observations at the current time and multiple historical time points, pressure differential forebody signals constructed based on pressure observations at the current time and multiple historical time points, and the maximum wind speed observation results at the current time.

7. The squall line thunderstorm gale prediction method according to claim 1, characterized in that, When constructing the near-surface antecedent perturbation field, the antecedent perturbation field is interpolated by a linear combination of radial basis functions, as shown in the following formula: ; ; ; in, It is a linear combination of radial basis functions. Let x be the coordinates of the i-th known data point, x be the coordinates of the data point to be interpolated, and n be the number of known data points. For radial basis functions, The weights are to be determined by solving a system of linear equations; Euclidean distance. It is a shape parameter that controls the decay rate or width of the radial basis function.

8. The squall line thunderstorm gale prediction method according to claim 1, characterized in that, Also includes: Based on the near-surface anterior disturbance field, regions with obvious disturbance characteristics are identified, including regions with enhanced near-surface disturbances, low-level dynamic anomalies, or regions that match strong convective organization structures. In the training loss or prediction response of the deep learning prediction network, the attention weight of regions with obvious perturbation features is increased, while the normal response is maintained for regions with indistinct perturbation features.

9. A squall line thunderstorm gale prediction device, characterized in that, include: The extraction module is used to extract dual-polarization radar features from the radar base data after quality control. The inversion module is used to invert radar radial wind vorticity and divergence at different radial distance scales based on radar radial velocity using the least squares fitting method, as low-level dynamic background features. The filtering module is used to select ground observation samples that match the radar observation records in time and space from the near-ground frontal body disturbance observation data, and to construct the near-ground frontal body disturbance field; The prediction module is used to obtain the near-term prediction results of thunderstorms and strong winds based on the dual-polarization radar features, low-level dynamic background features and near-ground frontal disturbance field using a deep learning prediction network. The post-processing module is used to enhance the thunderstorm gale proximity prediction results based on the potential diffusion model.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the squall line thunderstorm gale near-term prediction method as described in any one of claims 1 to 8.