Intelligent early warning method for severe convective weather based on phased array weather radar cooperative networking

Abstract

The application discloses a severe convective weather intelligent early warning method based on phased array weather radar cooperative networking, and belongs to the field of weather warning. The method comprises the following steps: S1, multiple phased array (dual-polarization) weather radars are used for cooperative networking observation to obtain high space-time resolution observation data; S2, quality control is performed on the obtained dual-polarization radar observation base data, including ground object clutter suppression, electromagnetic wave interference suppression, radial velocity deblurring and missing data filling; S3, the observation data of the multiple phased array weather radars are projected into three-dimensional grid point data in a Cartesian coordinate system, and data fusion processing is performed. The application has the beneficial effects that the severe convective weather intelligent early warning method based on phased array weather radar cooperative networking is provided, dual-polarization detection data and characteristic quantities thereof are added, and the recognition accuracy and suppression rate of ground object clutter are significantly improved.

Description

Technical Field

[0001] This application relates to the field of weather warning, and more specifically, to a method for intelligent early warning of severe convective weather based on a phased array weather radar collaborative network. Background Technology

[0002] Severe convective weather (such as hail, heavy rain, and thunderstorms) is characterized by its small spatial scale, short lifespan, sudden onset, rapid development and evolution, and high destructive power. Doppler weather radar, with its high spatiotemporal resolution and rapid scanning mechanism, has become a unique and effective means of real-time monitoring and early warning of small-scale severe convective weather. The WSR-88D Severe Storm Laboratory in the United States has developed the SCIT and TITAN storm identification algorithm systems. These systems, developed by the National Center for Atmospheric Research based on radar observation systems, are storm identification, tracking, analysis, and forecasting systems and have been widely used in the United States and other countries and regions worldwide. In recent years, with the rapid development of artificial intelligence technology, its greatest advantage lies in its ability to effectively solve the difficulties in forecasting and early warning of the nonlinear development and evolution of small-scale severe convective weather disasters under complex terrain conditions. Deep learning-based severe convective weather early warning models, represented by CNNs, significantly outperform traditional single-threshold judgment methods in terms of warning timeliness and accuracy.

[0003] Compared to single-polarization weather radars using mechanical scanning, phased array (dual-polarization) weather radars offer the advantage of higher spatiotemporal resolution. The time to detect a complete three-dimensional structure of severe convective weather is significantly reduced from 5-6 minutes to approximately 1 minute. Furthermore, its dual-polarization parameters contain richer information on precipitation process variations. In addition to physical quantities such as reflectivity factor, radial velocity, and spectral width, they also include polarization physical quantities such as differential reflectivity, differential propagation phase, differential propagation phase shift rate, correlation coefficient, and depolarization polarization ratio. These parameters reflect the scale differences of condensate particles in the horizontal and vertical directions, enabling more accurate estimation of condensate size, phase state, and number concentration distribution in three-dimensional space, thus providing more information on the characteristics of severe convective weather such as hail, heavy rain, and thunderstorms.

[0004] Furthermore, compared to CNN deep learning models, deep learning models based on the ViT (Vision Transformer) architecture have the following advantages in recognizing and predicting image features: First, through the self-attention mechanism, they have a stronger ability to capture global features; second, their lower layers contain global information, and features are evenly propagated and processed between layers of the neural network, which helps the model to understand sample data information more comprehensively; third, they have higher computational efficiency, can handle large multi-channel sample datasets, and can achieve dynamic adjustment of the model.

[0005] In summary, from the current technological perspective, future deep learning-based identification and early warning models based on the ViT architecture need to combine observation data from multiple phased array (dual-polarization) weather radars in a collaborative network, along with the dual-polarization structural characteristics of various severe convective weather events, to better address issues such as missed reports, false alarms, and delays in early warning of severe convective weather under complex climate and terrain conditions. Current traditional single-weather radar detection methods suffer from limitations such as limited data types, low temporal resolution, and blind spots in the boundary layer at medium to long distances, making it difficult to fully capture the formation and evolution characteristics of severe convective weather. The triggering mechanisms and cloud precipitation microphysical characteristics of severe convective weather vary significantly under different climates, seasons, and complex terrain conditions, exhibiting nonlinear evolution. Traditional identification and early warning methods such as SCIT and TITAN have low early warning timeliness and accuracy. Existing deep learning models based on the CNN architecture suffer from problems such as limited training sample data, poor global feature capture capabilities, uneven feature representation, and low computational efficiency. Consequently, their intelligent early warning models for severe convective weather have low early warning performance scores. Summary of the Invention

[0006] The summary section of this application is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description section below. This summary section is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0007] The purpose of this application is to address the shortcomings of existing technologies by providing an intelligent early warning method for severe convective weather based on a phased array weather radar collaborative network. This method aims to solve the problems of traditional single weather radar detection methods, such as limited data types, low temporal resolution, and blind spots in the boundary layer at medium to long distances.

[0008] To achieve the above objectives, this application provides the following technical solution: The intelligent early warning method for severe convective weather based on phased array weather radar collaborative networking includes the following steps: S1. Multiple phased array (dual polarization) weather radars are used for collaborative network observation to obtain high spatiotemporal resolution observation data; S2. Perform quality control on the acquired dual-polarization radar observation base data, including ground clutter suppression, electromagnetic interference suppression, radial velocity de-ambiguity, and missing data filling. S3. Project the observation data from multiple phased array weather radars into three-dimensional grid point data in the Cartesian coordinate system, and perform data fusion processing. S4. Extract typical feature sample datasets for classifying severe convective weather, including feature samples of hail, rainstorm, and strong wind weather; S5. Construct an intelligent early warning model for severe convective weather based on the ViT-Large neural network architecture, and use a multi-channel input-single-channel output mechanism for model training. S6. Use the trained model to provide intelligent early warnings for severe convective weather and output the warning results.

[0009] Optionally, the quality control steps for the dual-polarization radar observation base data include: A classification-based fuzzy logic method is used to incorporate the dual-polarization physical quantities "correlation coefficient / differential phase" for ground clutter suppression; Electromagnetic interference suppression is achieved using a multi-layer convolutional neural network. A dual retrieval method along the zero velocity line and along the radial direction is used to identify the radial velocity ambiguity region, and adaptive de-velocity ambiguity is performed; Kalman filtering is used to fill in large areas of missing data caused by beam obstruction or hardware failure.

[0010] Optionally, the collaborative networking observation steps of the multiple phased array weather radars include: Based on phased array radar observation data, we use moving average filtering and linear interpolation methods to integrate geographic information data to obtain three-dimensional spatial resolution cooperative and consistent observation data. The observation time of multiple phased array weather radars was unified using the IFNet residual neural network. Using an altimeter formula that takes into account the curvature of the Earth and the refraction of the standard atmosphere, and a polar-terrestrial coordinate transformation, combined with an interpolation method, the original observation data is projected into three-dimensional grid point data in a Cartesian coordinate system.

[0011] Optionally, the step of extracting typical feature samples for classifying severe convective weather includes: For hail weather, reflectivity factor, vertical cumulative liquid water content density and hail index are extracted as feature samples; For rainstorm weather, reflectivity factor, rainfall intensity and equivalent diameter of raindrops are extracted as feature samples; For windy weather, horizontal wind field and strong wind index are extracted as feature samples.

[0012] Optionally, the steps for constructing the intelligent early warning model for severe convective weather based on the ViT-Large neural network architecture include: The training sample dataset is standardized using z-score data preprocessing techniques. The image rotation method is used to increase the original training sample size; A multi-channel sample data input mechanism is adopted, including rainstorm characteristic data, hail characteristic data, and strong wind characteristic data; The "cat" stitching method is used to restore the output sub-image patches to feature data, and the warning results are obtained through convolution and sigmoid activation functions.

[0013] Optionally, the early warning model effectiveness evaluation step includes: Data from ground-based automatic weather stations, rain gauges, and raindrop spectrometers were used as true values. The Barness objective analysis method is used to interpolate the discrete true values ​​of the observations to two-dimensional uniform grid points; The Critical Success Index (CSI), False Alarm Rate (FAR), and Missed Alarm Rate (MAR) are used to quantitatively compare and evaluate the early warning results. The Continuous Ranking Probability Score (CRPS) is introduced to conduct a comprehensive quantitative assessment of the spatial reliability of the early warning results.

[0014] Optionally, the computer language and hardware environment used include: The raw data quality control algorithm of phased array (dual polarization) weather radar, the collaborative networking data fusion and inversion method, and the method for constructing typical structural feature samples of classified severe convective weather were compiled using standard C++ language; The ViT-Large deep learning severe convective weather early warning model is trained using Python and related environments; The training equipment consists of servers with high-performance CPUs and GPUs, and the operating system is LINUX / Ubuntu based on the x86 architecture.

[0015] Optionally, in the collaborative network observation of the multiple phased array weather radars, the data fusion processing adopts the exponential distance weighting method to perform data fusion on the observation overlap area to ensure the continuity of the spatial structure of precipitation.

[0016] Optionally, in the ViT-Large neural network architecture's intelligent early warning model for severe convective weather, a Gaussian kernel filter and discrete wavelet transform are introduced to remove noise and redundant information from the initial probability values.

[0017] Optionally, this method can improve the timeliness and accuracy of early warnings for severe convective weather under complex climate and terrain conditions, and reduce the rates of missed and false alarms.

[0018] The beneficial effects of this application are as follows: it increases the amount of dual-polarization detection data and its characteristic quantities, significantly improving the identification accuracy and suppression rate of ground clutter. Kalman filtering is employed to effectively fill large areas of missing data caused by electromagnetic beam obstruction due to complex mountainous terrain. The zero-velocity line retrieval method and the radial sliding window detection method are used in combination to determine the location of ambiguous velocities and perform adaptive velocity de-ambiguity to obtain true radial velocity data.

[0019] 2. A large-sample dataset of characteristic data for classifying severe convective weather is constructed using observation data from a phased array dual-polarization weather radar network based on an array-type antenna electronic scanning mechanism. Compared to traditional single-unit mechanically scanned single-polarization weather radars, the sample data from this patented technology has higher spatiotemporal resolution (horizontal range resolution up to 30m, time interval approximately 1min), and contains more complete and richer structural feature information (e.g., effectively covering the long-range boundary layer precipitation blind zone of the radar, inverting a large-scale, high-precision true horizontal wind field through the overlapping area of ​​multiple radars, and providing dual-polarization quantity classification information for the structural features of severe convective weather).

[0020] 3. Compared with existing severe convective weather warning methods, the patented technology of this invention uses an improved ViT-large deep learning warning model with multi-channel input and single-channel output, which significantly improves the timeliness and accuracy of warnings for various severe convective weather conditions under complex weather backgrounds and terrain conditions. Attached Figure Description

[0021] The accompanying drawings, which form part of this application, are used to provide a further understanding of the application and to make other features, objects, and advantages of the application more apparent. The illustrative embodiments and descriptions of this application are used to explain the application and do not constitute an undue limitation of the application.

[0022] Furthermore, throughout the accompanying drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and the elements are not necessarily drawn to scale.

[0023] In the attached diagram: Figure 1 A schematic diagram of the overall process of a deep learning-based early warning model for severe convective weather; Figure 2 This is a schematic diagram of polarization weather radar data quality control. Figure 3 This is a schematic diagram of ground clutter suppression in a dual-polarization weather radar. Figure 4 This is a schematic diagram illustrating the suppression of electromagnetic interference signals from dual-polarization weather radar. Figure 5 A schematic diagram of radial velocity deblurring for an X-band phased array weather radar; Figure 6 A schematic diagram illustrating the filling of missing data from a dual-polarization weather radar. Figure 7 A schematic diagram of a collaborative consistency reconstruction scheme for data from multiple phased array weather radars with ultra-high spatiotemporal resolution; Figure 8 A schematic diagram of the collaborative networking of reflectivity factor data from three phased array weather radars during a severe convective precipitation event in western Sichuan Basin. Figure 9Schematic diagram a of the collaborative network observation results of X-band phased array dual polarization radar for typical characteristics of hail weather; Figure 10 Schematic diagram b of the collaborative network observation results of X-band phased array dual polarization radar for typical characteristics of hail weather; Figure 11 Schematic diagram c: X-band phased array dual polarization radar collaborative network observation results for typical characteristic quantities of hail weather; Figure 12 a) Observation results of X-band phased array dual polarization radar collaborative networking for typical characteristics of short-term heavy precipitation weather; Figure 13 b) is the observation result of X-band phased array dual polarization radar collaborative network, which represents the typical characteristics of short-term heavy precipitation weather. Figure 14 c; Observation results of X-band phased array dual polarization radar collaborative networking for typical characteristics of short-term heavy precipitation weather. Figure 15 a) Observation results of X-band phased array (dual polarization) radar collaborative networking for typical characteristic quantities of thunderstorm and strong wind weather; Figure 16 b) is the observation result of X-band phased array (dual polarization) radar collaborative networking for typical characteristic quantities of thunderstorm and strong wind weather; Figure 17 A schematic diagram of the training architecture for the ViT-Large intelligent early warning model for classifying severe convective weather; Figure 18 Schematic diagram a1 of a localized severe convective weather event in the southwestern part of the Sichuan Basin; Figure 19 Schematic diagram a2 showing a localized severe convective weather event in the southwestern part of the Sichuan Basin; Figure 20 Schematic diagram a3 of a localized severe convective weather event in the southwestern part of the Sichuan Basin; Figure 21 Schematic diagram b1 of a localized severe convective weather event in the southwestern part of the Sichuan Basin; Figure 22 Schematic diagram b2 of a localized severe convective weather event in the southwestern part of the Sichuan Basin; Figure 23 Schematic diagram b3 of a localized severe convective weather event in the southwestern part of the Sichuan Basin; Figure 24 Schematic diagram c1 of a localized severe convective weather event in the southwestern part of the Sichuan Basin; Figure 25 Schematic diagram c2 of a localized severe convective weather event in the southwestern part of the Sichuan Basin; Figure 26 Schematic diagram c3 of a localized severe convective weather event in the southwestern part of the Sichuan Basin; Figure 27 A schematic diagram showing the CSI / FAR / MAR assessment results of 824 classified severe convective weather events in southwestern Sichuan Basin; Figure 28 The results of the CRPS assessment of 824 severe convective weather events in southwestern Sichuan Basin. (a) CRPS scores for hail warning results using three different methods; (b) CRPS scores for rainstorm warning results using three different methods; (c) Schematic diagram of CRPS scores for gale warning results using three different methods. Detailed Implementation

[0024] The intelligent early warning method for severe convective weather based on phased array weather radar cooperative networking provided in this application will be described in detail below with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more detailed, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit this application.

[0025] It should be noted that the use of terms such as "an embodiment," "an embodiment," "an exemplary embodiment," and "some embodiments" in the specification indicates that the described embodiment may include a specific feature, structure, or characteristic, but not every embodiment necessarily includes that specific feature, structure, or characteristic. Furthermore, when a specific feature, structure, or characteristic is described in connection with an embodiment, implementing such a feature, structure, or characteristic in conjunction with other embodiments (whether explicitly described or not) should be within the knowledge of those skilled in the art.

[0026] Generally, terms can be understood at least partly from their use in context. For example, depending at least partly on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in a singular sense, or a combination of features, structures, or characteristics in a plural sense. Additionally, the term "based on" can be understood not necessarily to convey an exclusive set of factors, but rather, alternatively, depending at least partly on the context, to allow for the presence of other factors that are not necessarily explicitly described.

[0027] The present application will now be described in detail with reference to the accompanying drawings and embodiments.

[0028] Reference Figure 1-15 As shown, this invention provides a smart early warning method for severe convective weather based on phased array weather radar collaborative networking, including the following steps: S1. Multiple phased array (dual polarization) weather radars are used for collaborative network observation to obtain high spatiotemporal resolution observation data; S2. Perform quality control on the acquired dual-polarization radar observation base data, including ground clutter suppression, electromagnetic interference suppression, radial velocity de-ambiguity, and missing data filling. S3. Project the observation data from multiple phased array weather radars into three-dimensional grid point data in the Cartesian coordinate system, and perform data fusion processing. S4. Extract typical feature sample datasets for classifying severe convective weather, including feature samples of hail, rainstorm, and strong wind weather; S5. Construct an intelligent early warning model for severe convective weather based on the ViT-Large neural network architecture, and use a multi-channel input-single-channel output mechanism for model training. S6. Use the trained model to provide intelligent early warnings for severe convective weather and output the warning results.

[0029] The quality control steps for the dual-polarization radar observation base data include: A classification-based fuzzy logic method is used to incorporate the dual-polarization physical quantities "correlation coefficient / differential phase" for ground clutter suppression; Electromagnetic interference suppression is achieved using a multi-layer convolutional neural network. A dual retrieval method along the zero velocity line and along the radial direction is used to identify the radial velocity ambiguity region, and adaptive de-velocity ambiguity is performed; Kalman filtering is used to fill in large areas of missing data caused by beam obstruction or hardware failure.

[0030] The collaborative network observation steps of the multiple phased array weather radars include: Based on phased array radar observation data, we use moving average filtering and linear interpolation methods to integrate geographic information data to obtain three-dimensional spatial resolution cooperative and consistent observation data. The observation time of multiple phased array weather radars was unified using the IFNet residual neural network. Using an altimeter formula that takes into account the curvature of the Earth and the refraction of the standard atmosphere, and a polar-terrestrial coordinate transformation, combined with an interpolation method, the original observation data is projected into three-dimensional grid point data in a Cartesian coordinate system.

[0031] The steps for extracting typical feature samples for classifying severe convective weather include: For hail weather, reflectivity factor, vertical cumulative liquid water content density and hail index are extracted as feature samples; For rainstorm weather, reflectivity factor, rainfall intensity and equivalent diameter of raindrops are extracted as feature samples; For windy weather, horizontal wind field and strong wind index are extracted as feature samples.

[0032] The steps for constructing the intelligent early warning model for severe convective weather based on the ViT-Large neural network architecture include: The training sample dataset is standardized using z-score data preprocessing techniques. The image rotation method is used to increase the original training sample size; A multi-channel sample data input mechanism is adopted, including rainstorm characteristic data, hail characteristic data, and strong wind characteristic data; The "cat" stitching method is used to restore the output sub-image patches to feature data, and the warning results are obtained through convolution and sigmoid activation functions.

[0033] The evaluation steps for the effectiveness of the early warning model include: Data from ground-based automatic weather stations, rain gauges, and raindrop spectrometers were used as true values. The Barness objective analysis method is used to interpolate the discrete true values ​​of the observations to two-dimensional uniform grid points; The Critical Success Index (CSI), False Alarm Rate (FAR), and Missed Alarm Rate (MAR) are used to quantitatively compare and evaluate the early warning results. The Continuous Ranking Probability Score (CRPS) is introduced to conduct a comprehensive quantitative assessment of the spatial reliability of the early warning results.

[0034] The computer languages ​​and hardware environments it uses include: The raw data quality control algorithm of phased array (dual polarization) weather radar, the collaborative networking data fusion and inversion method, and the method for constructing typical structural feature samples of classified severe convective weather were compiled using standard C++ language; The ViT-Large deep learning severe convective weather early warning model is trained using Python and related environments; The training equipment consists of servers with high-performance CPUs and GPUs, and the operating system is LINUX / Ubuntu based on the x86 architecture.

[0035] In the collaborative network observation of the multiple phased array weather radars, the data fusion processing adopts the exponential distance weighting method to fuse data in the overlapping observation areas to ensure the continuity of the spatial structure of precipitation.

[0036] In the ViT-Large neural network architecture-based intelligent early warning model for severe convective weather, Gaussian kernel filters and discrete wavelet transforms are introduced to remove noise and redundant information from the initial probability values. This method can improve the timeliness and accuracy of early warnings for severe convective weather under complex climate and terrain conditions, and reduce the rates of missed and false alarms.

[0037] This invention employs an improved radar data quality control method for dual-polarization radar observation data, compared to traditional single-polarization radar data quality control methods. The main improvements include ground clutter suppression, electromagnetic interference suppression, radial velocity de-ambiguity correction, and missing data imputation. The dual-polarization weather radar data quality control technology scheme is as follows: Figure 2 As shown. From Figure 2 It can be seen that this invention incorporates the dual-polarization physical quantity "correlation coefficient / differential phase" and its characteristic quantity into the classification fuzzy logic method, which significantly improves the accuracy of ground clutter identification and effectively removes ground clutter signal interference. Figure 3 This invention, based on electromagnetic interference echo characteristics, uses a multi-layer convolutional neural network to improve the electromagnetic interference suppression rate. Figure 4 ), Figure 4 This is a schematic diagram of electromagnetic interference signal suppression for dual-polarization weather radar (left: before suppression; right: after suppression). The invention employs a dual retrieval method along the zero-velocity line and along the radial direction to identify radial velocity ambiguity regions and adaptively de-ambiguously de-velocity signals for short-wavelength phased array radar (X-band). Figure 5 ), Figure 5 This is a schematic diagram of radial velocity deblurring for an X-band phased array weather radar (left: before deblurring; right: after deblurring). Furthermore, this invention employs Kalman filtering to effectively fill in large areas of missing data caused by beam obstruction and hardware failure. Figure 6 ).

[0038] This invention utilizes an improved method for reconstructing the consistency of spatiotemporal data from radar detection to achieve coordinated observation by multiple phased array (dual polarization) radars in a coordinated network. Figure 7 A collaborative consistency reconstruction scheme for data from multiple phased array weather radars with superior spatiotemporal resolution is proposed. Figure 7 It can be seen that: based on phased array radar observation data, methods such as moving average filtering and linear interpolation are used to integrate geographic information data to obtain three-dimensional spatial resolution coordinated and consistent observation data; based on the IFNet residual neural network, the observation time of multiple phased array weather radars is unified.

[0039] This invention uses the altimeter formula (Equation 10) that takes into account the curvature of the Earth and standard atmospheric refraction, and the polar-Earth coordinate transformation, combined with Cressman / nearest neighbor interpolation methods, to project the raw observation data of the phased array (dual polarization) weather radar into three-dimensional grid point data in the Cartesian coordinate system.

[0040] (10) In formula 10, To account for the actual vertical height of radar targets relative to the ground due to the curvature of the Earth and atmospheric refraction, For antenna altitude, Indicates the radar radial distance. Indicates the radar detection elevation angle. This represents the equivalent Earth radius (approximately 8500 km). This invention employs a method of 11 to project the coordinates of grid points at different altitudes of a single radar into a network area, thereby completing a three-dimensional network of multiple phased array (dual polarization) weather radars.

[0041] (11) In equation 11, ( ) represents the coordinates of two-dimensional planar grid points at different altitudes in the three-dimensional grid data of a single radar. These are their actual coordinates projected onto the network area. )and( () respectively represent the latitude, longitude and coordinates of the center of the network area of ​​multiple radars. and( ) represent the center coordinates and latitude and longitude of a single radar on a two-dimensional plane grid, respectively. This indicates the radial range resolution of the phased array (dual polarization) weather radar after coordination and consistency. It is a constant, representing the Earth's radius (approximately 6371 km).

[0042] After networking multiple phased array radars using the 11-type method, for the detection blind spot of a certain radar, the non-invalid values ​​of another radar at the same location can be used to fill in the gap. In addition, this invention uses the exponential distance weighting method to perform data fusion on the overlapping areas of observations from multiple radars (except for radial velocity), which can ensure the continuity of the original spatial structure of precipitation as much as possible.

[0043] Figure 8 It was a strong convective precipitation event that occurred in the western part of the Sichuan Basin. Figure 8 The top right image shows the reflectivity factor at a height of 2km after observation and interpolation by the Chengdu phased array (dual polarization) radar; the middle right image shows the reflectivity factor at a height of 2km after observation and interpolation by the Ya'an phased array (dual polarization) radar; and the bottom right image shows the reflectivity factor at a height of 2km after observation and interpolation by the Leshan phased array (dual polarization) radar. Figure 8 The left side shows the fused reflectivity factor data at a height of 2km after the three radars were networked together. From Figure 8 It can be seen that the reflectivity factor coverage area is significantly increased after networking, and the entire strong convective precipitation echo area is more complete and continuous. Data gaps near radar stations caused by interpolation are filled. Figure 8 Within the black box, data missing from three radars due to ground obstruction has been filled in. Figure 8Within the radar network, in the overlapping area of ​​the three radars, the data fusion processing effectively suppressed the significantly high reflectivity factor and enhanced the previously low reflectivity factor due to attenuation. Figure 8 Inside the frame.

[0044] This invention extracts typical feature samples of severe convective weather based on three-dimensional grid point data from a collaborative network of multiple phased array weather radars. It categorizes severe convective weather into three types: rainstorms (short-duration heavy precipitation), hail, and strong winds. Based on the typical structural characteristics of each type of severe convective weather, it optimizes the observation operators to obtain feature sample datasets for the above three types of typical severe convective weather.

[0045] Hail Feature Sample Extraction Method (1) Reflectivity factor feature extraction: Based on different terrain conditions (plains, plateaus or mountains), two-dimensional reflectivity factor grid data of three-dimensional grid points of multiple phased array weather radars at a specified vertical height layer relative to the ground are extracted as one of the hail feature samples.

[0046] (2) Extraction of vertically accumulated liquid water content density features: vertically accumulated liquid water content density The definitions are as follows (Form 12): (12) In formula 12, Represents the reflectivity factor at different vertical heights. This represents the maximum height layer (adjustable according to different regions and precipitation types). Two-dimensional grid points can be obtained using equation 12. The data serves as the second sample of hail characteristics.

[0047] (3) Hail index feature extraction: Hail index of phased array (dual polarization) weather radar The definition is as follows (Form 13): (13) In formula 13, These represent the reflectivity factor and differential reflectivity at a specified relative vertical height (adjusted according to different regions and terrain conditions), respectively. Two-dimensional grid points can be obtained using Equation 13. The data serves as the third sample of hail characteristics.

[0048] Figure 9 , Figure 10 , Figure 11 It involves multiple X-band phased array (dual polarization) radars working together in a network to observe and extract the reflectivity factor of a hailstorm at a vertical height of 1.5 km above the ground. Feature quantity, and Characteristic quantity.

[0049] (a) Reflectivity factor at a vertical height of 1.5 km; (b) Reflectivity factor at a vertical height of 1.5 km (c)

[0050] 6.3.2 Method for Extracting Feature Samples of Rainstorms (1) Extraction of reflectance factor features: The method of extracting hail reflectance factor features is basically the same as that in Section 6.3.1, and is used as one of the feature samples of rainstorm.

[0051] (2) Rainfall intensity (hourly rainfall rate) feature extraction: Rainfall intensity The definition is as follows (Form 14): (14) In formula 14, These represent the reflectivity factor and differential reflectivity at a specified relative vertical height (adjusted according to different regions and terrain conditions), respectively. for The adjustable parameters of the relationship need to be retrieved by combining rain gauge / raindrop spectrometer and weather radar observation data under complex terrain and precipitation types in different regions. Two-dimensional grid points can be obtained through Equation 14. The data serves as the second sample of characteristics of rainstorms.

[0052] (3) Extraction of equivalent diameter features of raindrops: equivalent diameter of liquid raindrops The definition is as follows (Form 15): (15) In the 15th formula, It represents the differential reflectance at a specified relative vertical height (adjusted according to different regions and terrain conditions). To ensure adjustable parameters, inversion is required under complex terrain and precipitation types in different regions, using a combination of raindrop spectrometer and weather radar observation data. Two-dimensional grid points can be obtained using equation 15. The data serves as the third sample of characteristics of rainstorms.

[0053] Figure 12 , Figure 13 , Figure 14 This is the reflectivity factor of a short-duration heavy precipitation event observed and extracted at a vertical height of 1.5 km above the ground by multiple X-band phased array (dual polarization) radars working together in a network. and Characteristic quantity.

[0054] (1) Horizontal wind field feature extraction: Based on the radial velocity data at a specified relative vertical height layer (adjusted according to different regions and terrain conditions) in the overlapping area by multiple phased array radars in a coordinated network, the radial velocity-horizontal wind field projection equation set can be formed by every two phased array radars as follows (Equation 16): (16) In the 16th style This indicates the different radial velocities observed by two radars at the same location within the overlapping area of ​​a cooperative network. It is a horizontal wind field. These represent the azimuth and elevation angles of the observation point relative to the two radars, respectively. Since only the influence of boundary layer winds on the ground is considered, vertical wind... The magnitude is much smaller than that of horizontal winds; furthermore, the boundary layer is typically located in the lower to middle elevation levels for radar detection. The component contributes very little to the radial velocity; therefore, the value in equation 16 can be ignored. The term is obtained by solving a simplified system of equations to obtain the horizontal wind field at a two-dimensional grid. The data serves as one of the characteristic samples of strong winds.

[0055] (2) Strong wind index feature extraction: Strong wind index The definition is as follows (Form 17): (17) In the 15th formula, These represent the reflectivity factor, differential reflectivity, and correlation coefficient at a specified relative vertical height (adjusted according to different regions and terrain conditions). Adjustable weighting coefficients are assigned under complex terrain and precipitation types in different regions. Two-dimensional grid points can be obtained using Equation 17. The data serves as the second sample of characteristics of rainstorms.

[0056] Figure 15 , Figure 16 This is a data set of multiple X-band phased array (dual polarization) radars that collaboratively networked and observed and extracted the horizontal wind field at a vertical height of 1.5 km during a thunderstorm with strong winds. Characteristic quantity.

[0057] Scheme for Convective Weather Intelligent Early Warning Model Based on ViT-Large Neural Network Architecture This invention patent technology establishes a severe convective weather early warning model based on the ViT-large neural network architecture, defined as follows (Equation 18): (18) In the 18th formula, It is a feature sample dataset of classified severe convective weather at a two-dimensional grid point coordinate at a specified height layer at the current time, with multi-channel input (see Section 6.3). The future of single-channel output Minutes later ( The early warning results for various severe convective weather conditions are based on the coordinates of two-dimensional grid points (with different values ​​for different severe convective weather conditions).

[0058] The large-sample training dataset for the deep learning early warning model selected in this invention comes from nearly six years of observation data and their features from X-band phased array (dual polarization) weather radar, automatic weather stations, and raindrop spectrometers used to classify severe convective weather in Chengdu, Ya'an, and Leshan from 2019 to the present. The dataset has a distance resolution of 250m, a horizontal coverage area of ​​256km*256km for the networked area of ​​multiple radars, a relative ground height of 1.5km for the specified altitude layer, and a 1024*1024 grid of points for constructing the two-dimensional image.

[0059] Figure 17 This invention relates to the ViT-Large intelligent early warning model training architecture for classified severe convective weather, a patented technology. The model employs z-score data preprocessing to standardize the training sample dataset to a data distribution with a mean of 0 and a standard deviation of 1, which improves the stability and convergence speed of the model training. Simultaneously, an image rotation method is used to increase the original training sample size. Furthermore, a multi-channel sample data input mechanism is employed, primarily including three categories: rainstorm feature data, hail feature data, and strong wind feature data. Each type of severe convective weather contains M channels, each containing N*N two-dimensional grid points, resulting in a total sample dataset of 3*M*N*N samples (M=3, N=1024 in this patented technology). These data are divided into 256 sub-image blocks at 16-distance intervals, each image block carrying 3*64*64 data points. These sub-image blocks are flattened and linearly transformed before entering the encoder. Each encoding layer consists of two sub-layers: a multi-attention head and a feedforward neural network. The embedded sub-image patch data is processed by layer normalization and then fed into a multi-attention head. After residual connection and layer normalization, it is fed into a feedforward neural network and then subjected to another residual connection before outputting the processed sub-image patch.

[0060] This invention uses a "cat" stitching method to restore the 256 sub-image blocks (each carrying 3*64*64 data points) output above into 3-channel feature data, with each channel containing 1024*1024 data points. Convolution is used to compress the 3-channel feature data into single-channel data, and an improved sigmoid activation function is added to normalize the data to probability values ​​between 0 and 1. Furthermore, this invention introduces a 3*3 Gaussian kernel filter (Equation 19) and discrete wavelet transform (Equation 20) to remove noise and redundant information from the initial probability values, obtaining the final warning result.

[0061] (19) (20) 19th style For the data points to be processed, The neighborhood of the data point to be processed is 8 points. is the standard deviation (an adjustable parameter). In equation 20, The data to be processed is divided into four sub-bands. , It is a series decomposition. It is a low-frequency coefficient. It is high-frequency detail in the horizontal direction. It refers to high-frequency details in the vertical direction. This refers to high-frequency details in the diagonal direction. (Using this technique for high-frequency details...) The function removes high-frequency noise, and then the denoised data is obtained through inverse wavelet transform (IDWT). .

[0062] This invention patent technology employs a binary cross-entropy loss function, the purpose of which is to find the minimum cost and update the model parameters through gradient descent, as defined below (Equation 21): (twenty one) In equation 21, This is the output of the warning result. It is the true value of the actual observation (takes a value of 0 or 1). This refers to the number of samples. Applying the chain rule to equation 21, we can reverse-engineer the weight gradients of the output layer neural network. Matrix and bias gradient The matrix is ​​used to obtain a new one by applying weight decay. (Form 22): (twenty two) In formula 22, It is an adjustable weight decay coefficient (as defined in this invention patent). =0.01), These are the weights and biases updated after the previous gradient descent. This invention further introduces... The first moment and the second moment (Equation 23): (twenty three) In formula 23, and It is an adjustable attenuation rate of the first and second moments (as described in this invention patent). , ). yes The first moment (in this invention patent, the initial value is set to) ). yes The second moment (in this invention patent, the initial value is set to) ).

[0063] Combining equations 22 and 23, we can further update the neural network layer. (24-style): (twenty four) In the 24th formula, It's the learning rate. For the iteration round (initial round) ), It is a very small adjustable positive number (as defined in this invention patent). The final updated neural network layer is obtained. Then, the gradient descent method (Equation 25) is used to gradually optimize the weight and bias parameter matrices in the ViT multilayer neural network. (Form 25): (25) This invention patent technology uses observation data from ground automatic stations, rain gauges, and raindrop spectrometers as true values, and uses the Barness objective analysis method (Equation 26) to interpolate the discrete true values ​​of the observations to two-dimensional uniform grid points in order to quantitatively evaluate the warning results of the ViT-Large severe convective weather warning model.

[0064] (26) In the 26th formula, This indicates the point to be interpolated, which is the center point (0,0) of the 5*5 window, and there is no actual observation data at this location; This represents the actual observation data of the 24 neighboring points around the point to be interpolated (if there is no actual observation data, the value is 0). The weighting coefficients are represented by the distances between each neighboring point and the point to be interpolated. and filter coefficients This invention patent technology uses the root mean square error of the distance between the interpolation point and its 24 neighboring points, along with an adjustment coefficient, to jointly determine the method. Sure . The value range is typically between 0.8 and 1.5. This invention patent technology, in order to highlight the influence of nearby sites, takes... .

[0065] Figures 18-26 The following is ground-based observation data of a localized severe convective weather event that occurred in the southwestern part of the Sichuan Basin: Figure 18 a1 represents the actual ground observation results of hail; Figure 19 a2 represents the actual ground observation results of short-term heavy precipitation; Figure 20 a3 represents the actual ground observation results of strong winds; Figure 21 b1- Figure 23 b3 is correct Figure 18 The a1-20a3 field observation data uses two-dimensional uniform grid point data after interpolation using the 26-type interpolation method; Figure 24 c1- Figure 26 c3 represents the prediction results of the aforementioned three types of severe convective weather using the ViT-Large deep learning early warning model, which utilizes the patented technology of this invention. From Figure 13 It can be seen that the warning results for hail and short-duration heavy precipitation (24c1-) Figure 26 The c3 landing area is relatively accurate, but the warning area is less accurate than the actual landing area. Figure 21 b1- Figure 23 b3) is too high; the wind warning result is ( Figure 26 The c3 landing area has a certain deviation, and the warning area is different from the actual area ( Figure 23 b3) is too small.

[0066] Figure 27 This is a localized severe convective weather event located in the southwestern part of the Sichuan Basin. This invention employs three methods (a1-a3) to score hail warnings using CSI / FAR / MAR; three methods (b1-b3) to score rainstorm warnings using CSI / FAR / MAR; and three methods (c1-c3) to score gale warnings using CSI / FAR / MAR. The patented technology uses the critical success index (CSI, equation 27), false alarm rate (FAR, equation 27), and miss alarm rate (MAR, equation 27) to quantitatively compare and evaluate severe convective weather warning results. An improved Continuous Ranked Probability Score (CRPS, equation 28) is introduced to provide a comprehensive quantitative assessment of the spatial reliability of severe convective weather warning results.

[0067] (27) In formula 27, The table is divided into three sub-tables: the number of grid points with successful warnings (i.e., a grid point's warning result was severe convective weather, and severe convective weather actually occurred at that location), the number of grid points with missed warnings (i.e., a grid point's warning result was non-severe convective weather, and severe convective weather actually occurred at that location), and the number of false alarms (i.e., a grid point's warning result was severe convective weather, and severe convective weather actually did not occur at that location).

[0068] (28) In the 28th formula, It is the cumulative distribution function. Let be the indicator function. In this invention patent technology, the early warning result is 1024*1024 two-dimensional discrete grid point data, where each grid is binary data (i.e., 0 or 1). Therefore, discretizing and simplifying equation 28 yields equation 29: (29) In formula 29, This indicates the number of data points for which the warning result is not severe convective weather. This represents the probability of non-severe convective weather in a 1024*1024 two-dimensional grid. This represents the total number of actual severe convective weather events in the two-dimensional grid data. The data value (i.e., 1) corresponds to the actual occurrence of severe convective weather at each grid point. This represents the total number of instances in the two-dimensional grid data where no severe convective weather actually occurred. This represents the data value (i.e., 0) for each grid point where no severe convective weather occurred. This refers to the CRPS score of the severe convective weather warning results.

[0069] We selected 197 severe convective weather events in southwestern Sichuan Basin from 2022 to the present, including approximately 824 individual cases of hail, short-duration heavy rainfall, and strong winds. Figure 27 To quantitatively evaluate the CSI / FAR / MAR of the early warning results of the aforementioned 824 sample datasets using three methods: SCIT, CNN, and VIT (the technology of this invention). Figure 27a1-a3 represent the CSI / FAR / MAR scores of the three methods for hail weather warning results. As can be seen from the figures: the CSI score of the ViT method of this invention is 0.47, higher than the CNN method's 0.39, and significantly higher than the SCIT method's 0.21; the FAR score of the ViT method of this invention is 0.29, lower than the CNN method's 0.33, and significantly lower than the SCIT method's 0.43; the MAR score of the ViT method of this invention is 0.19, lower than the CNN method's 0.31, and significantly lower than the SCIT method's 0.54. Figure 27 b1-b3 represent the CSI / FAR / MAR scores of the three methods for rainstorm weather warning results. As can be seen from the figures: the CSI score of the ViT method of this invention is 0.53, higher than the CNN method's 0.45, and significantly higher than the SCIT method's 0.38; the FAR score of the ViT method of this invention is 0.24, lower than the CNN method's 0.26, and significantly lower than the SCIT method's 0.41; the MAR score of the ViT method of this invention is 0.19, lower than the CNN method's 0.25, and significantly lower than the SCIT method's 0.32. Figure 27 c1-c3 represent the CSI / FAR / MAR scores of the three methods for strong wind weather warnings. As shown in the figure: the ViT method of this invention has a CSI score of 0.33, significantly higher than the CNN method's 0.19 and the SCIT method's 0.12; the ViT method's FAR score is 0.24, lower than the CNN method's 0.39 and significantly lower than the SCIT method's 0.45; and the ViT method's MAR score is 0.37, significantly lower than the CNN method's 0.54 and the SCIT method's 0.62. In summary, the ViT warning method of this invention significantly outperforms the traditional CNN and SCIT methods in warning of three types of severe convective weather. Furthermore, the ViT method is more effective in warning of hail and heavy rain than in warning of strong winds, primarily because the ViT method has a much higher false negative rate for strong winds than for hail and heavy rain.

[0070] Figure 28 The CRPS score of the early warning results for all two-dimensional grid points within the observation range of the phased array radar network is presented based on the aforementioned 824 sample datasets. Figure 28 Figure a shows the quantitative evaluation of CRPS of hail weather warning results by three methods. As can be seen from the figure, the ViT method of this invention has lower CRPS scores than the two traditional methods CNN and SCIT, whether it is the maximum CRPS score of a single hail weather event (ViT(Max)) or the average CRPS score of all hail weather samples (ViT(Aver)). Moreover, its warning timeliness (20 minutes) is also better than the traditional CNN method (10 minutes) and SCIT method (2 minutes). Figure 28b shows the quantitative evaluation of CRPS of the three methods for rainstorm weather warning results. As can be seen from the figure, the ViT method of this invention has lower CRPS scores than the two traditional methods CNN and SCIT, whether it is the maximum CRPS score of a single rainstorm weather (ViT(Max)) or the average CRPS score of all rainstorm weather samples (ViT(Aver)). Moreover, its warning timeliness (30 minutes) is also better than the traditional CNN method (15 minutes) and SCIT method (5 minutes). Figure 28 Figure c shows the quantitative CRPS evaluation of the three methods for strong wind weather warning results. As can be seen from the figure, the ViT method of this invention has lower CRPS scores than the traditional CNN and SCIT methods, both for the maximum CRPS score (ViT(Max)) of a single strong wind weather event and the average CRPS score (ViT(Aver)) of all rainstorm weather samples. Furthermore, its warning timeliness (20 minutes) is superior to the traditional CNN method (15 minutes) and SCIT method (10 minutes). In summary, the ViT warning method of this invention outperforms the traditional CNN and SCIT methods in terms of warning timeliness and accuracy for three types of severe convective weather, and its warning timeliness and accuracy for rainstorm weather are superior to those for hail and strong wind weather.

[0071] Figure 27 The CSI / FAR / MAR assessment results for 824 severe convective weather events in southwestern Sichuan Basin are presented. The CSI / FAR / MAR scores for hail warnings are given by three methods (a1-a3); the CSI / FAR / MAR scores for rainstorm warnings are given by three methods (b1-b3); and the CSI / FAR / MAR scores for gale warnings are given by three methods (c1-c3).

[0072] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

[0073] It should be noted that this application covers any alternatives, modifications, equivalent methods, and solutions made within the spirit and scope of this application. To provide the public with a thorough understanding of this application, specific details are described in detail in the preferred embodiments, while those skilled in the art can fully understand this application without these details. Furthermore, to avoid unnecessary confusion regarding the substance of this application, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0074] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for intelligent early warning of severe convective weather based on phased array weather radar collaborative networking, characterized in that: Includes the following steps: S1. Multiple phased array weather radars are used for collaborative network observation to acquire high spatiotemporal resolution observation data; the horizontal range resolution of the high spatiotemporal resolution is 30m, and the time interval is 1min; S2. Perform quality control on the acquired dual-polarization radar observation base data, including ground clutter suppression, electromagnetic interference suppression, radial velocity de-ambiguity, and missing data filling. S3. Project the observation data from multiple phased array weather radars into three-dimensional grid point data in the Cartesian coordinate system, and perform data fusion processing. S4. Extract typical feature sample datasets for classifying severe convective weather, including feature samples of hail, rainstorm, and strong wind weather; S5. Construct an intelligent early warning model for severe convective weather based on the ViT-Large neural network architecture, and use a multi-channel input-single-channel output mechanism for model training. S6. Use the trained model to provide intelligent early warnings for severe convective weather and output the warning results; The steps for extracting typical feature samples for classifying severe convective weather include: For hail weather, reflectivity factor, vertical cumulative liquid water content density and hail index are extracted as feature samples; For rainstorm weather, reflectivity factor, rainfall intensity and equivalent diameter of raindrops are extracted as feature samples; For windy weather, horizontal wind field and strong wind index are extracted as feature samples; The steps for constructing the intelligent early warning model for severe convective weather based on the ViT-Large neural network architecture include: The training sample dataset is standardized using z-score data preprocessing techniques. The image rotation method is used to increase the original training sample size; A multi-channel sample data input mechanism is adopted, including rainstorm characteristic data, hail characteristic data, and strong wind characteristic data; The cat concatenation method is used to restore the output sub-image patches to feature data, and the warning results are obtained through convolution and sigmoid activation functions.

2. The intelligent early warning method for severe convective weather based on phased array weather radar collaborative networking as described in claim 1, characterized in that: The quality control steps for the dual-polarization radar observation base data include: A classification fuzzy logic method is adopted to incorporate dual-polarization physical quantities, and the correlation coefficient / differential phase is used to suppress ground clutter; Electromagnetic interference suppression is achieved using a multi-layer convolutional neural network. A dual retrieval method along the zero velocity line and along the radial direction is used to identify the radial velocity ambiguity region, and adaptive de-velocity ambiguity is performed; Kalman filtering is used to fill in large areas of missing data caused by beam obstruction or hardware failure.

3. The intelligent early warning method for severe convective weather based on phased array weather radar collaborative networking as described in claim 1, characterized in that: The collaborative network observation steps of the multiple phased array weather radars include: Based on phased array radar observation data, we use moving average filtering and linear interpolation methods to integrate geographic information data to obtain three-dimensional spatial resolution cooperative and consistent observation data. The observation time of multiple phased array weather radars was unified using the IFNet residual neural network. Using an altimeter formula that takes into account the curvature of the Earth and the standard atmospheric refraction conditions, and a polar-Earth coordinate transformation, combined with an interpolation method, the original observation data is projected into three-dimensional grid point data in a Cartesian coordinate system.

4. The intelligent early warning method for severe convective weather based on phased array weather radar collaborative networking as described in claim 1, characterized in that: The evaluation steps for the effectiveness of the early warning model include: Data from ground-based automatic weather stations, rain gauges, and raindrop spectrometers were used as true values. The Barness objective analysis method is used to interpolate the discrete true values ​​of the observations to two-dimensional uniform grid points; The critical success index, false alarm rate, and false alarm rate are used to quantitatively compare and evaluate the early warning results. A continuous sorting probability score is introduced to conduct a comprehensive quantitative assessment of the spatial reliability of the early warning results.

5. The intelligent early warning method for severe convective weather based on phased array weather radar collaborative networking according to any one of claims 1-4, characterized in that: The computer languages ​​and hardware environments it uses include: The raw data quality control algorithm of phased array weather radar, the collaborative networking data fusion and inversion method, and the method for constructing typical structural feature samples of classified severe convective weather were compiled using standard C++ language. The ViT-Large deep learning severe convective weather early warning model is trained using Python and related environments; The training equipment consists of servers with high-performance CPUs and GPUs, and the operating system is LINUX / Ubuntu based on the x86 architecture.

6. The intelligent early warning method for severe convective weather based on phased array weather radar collaborative networking as described in claim 1, characterized in that: In the collaborative network observation of the multiple phased array weather radars, the data fusion processing adopts the exponential distance weighting method to fuse data in the overlapping observation areas to ensure the continuity of the spatial structure of precipitation.

7. The intelligent early warning method for severe convective weather based on phased array weather radar collaborative networking as described in claim 1, characterized in that: In the ViT-Large neural network architecture's intelligent early warning model for severe convective weather, a Gaussian kernel filter and discrete wavelet transform are introduced to remove noise and redundant information from the initial probability values.

8. The intelligent early warning method for severe convective weather based on phased array weather radar collaborative networking as described in claim 1, characterized in that: This method can improve the timeliness and accuracy of early warnings for severe convective weather under complex climate and terrain conditions, and reduce the rates of missed and false reports.

Images