Thunderstorm and gale identification method based on multi-dimensional characteristic parameter fusion

Abstract

The invention relates to a thunderstorm and gale identification method based on multi-dimensional characteristic parameter fusion, and belongs to the field of radio, and the method comprises the steps: employing a neighborhood value method to remove the electromagnetic interference of dual-line polarization Doppler weather radar data, carrying out the rollback fuzzy processing, and then carrying out the meshing; dual-polarization radar data, wind profile radar data, sounding data and ground automatic station data are respectively obtained, and respective characteristic parameters are extracted from the data; carrying out normalization processing on the characteristic parameters, and obtaining corresponding parameters; constructing a multi-parameter integrated early warning model based on the plurality of fusion parameters, and training and verifying the model; and inputting parameters into the trained model for operation, and outputting the thunderstorm gale probability P. According to the thunderstorm gale identification method, through combination of multi-source data fusion and multiple physical mechanism fusion and integration of innovation of a machine learning model, the technology of thunderstorm gale identification from single index dependence to multi-dimensional collaboration and from static threshold value to dynamic self-adaption is realized.

Description

Technical Field

[0001] The present invention relates to the field of radio, and in particular to a thunderstorm and gale identification method based on multi-dimensional feature parameter fusion. Background Art

[0002] Thunderstorm gale is a type of non-tornadic straight-line gale weather generated by a convective system. The China Meteorological Administration's judgment standard for thunderstorm gale is that the maximum gust wind speed generated by convection is greater than 17.2m / s. However, since thunderstorm gale is mainly generated by small and medium-scale convective systems, it is highly sudden and localized, making it difficult to forecast. Currently, China's meteorological business departments generally give low scores to this type of weather forecast, and existing technical concepts and methods need to be further improved.

[0003] However, existing technologies suffer from the following drawbacks: First, the use of a small number of characteristic physical quantities leads to a high underreporting rate. Most existing methods rely solely on radar reflectivity thresholds, resulting in a high underreporting rate for hidden gale-force winds such as downbursts. Second, multiple data sources are used in isolation, without deep integration of radar, wind profiler, and ground station data. For example, these methods focus solely on common indicators such as radar reflectivity and wind speed, without fully analyzing their formation mechanisms. Thunderstorm formation involves the complex interactions of multiple factors, including water vapor, vertical airflow, and unstable energy. Simple principles are difficult to fully consider. In complex weather conditions, such as areas with strong confluence of cold and warm air and significant topographic influences, key factors can be easily overlooked, leading to identification errors. Third, warning lead time is insufficient. Existing patents fail to integrate and analyze numerous complex influencing factors when forecasting thunderstorms. They fail to fully consider the temporal and spatial variations and interactions of different meteorological elements, such as the impact of atmospheric stability and vertical wind shear on thunderstorm intensity and path. This fails to meet the high demands for accurate meteorological warnings, such as those for aviation and power facility protection. Summary of the Invention

[0004] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a thunderstorm gale identification method based on multi-dimensional feature parameter fusion, which solves the shortcomings of the existing technology.

[0005] The object of the present invention is achieved by the following technical solution: a thunderstorm gale identification method based on multi-dimensional feature parameter fusion, the identification method comprising:

[0006] Step 1: The dual-polarization Doppler weather radar data is subjected to the neighborhood value method to remove electromagnetic interference, and then subjected to de-velocity blurring and gridding.

[0007] Step 2: Obtain dual-polarization radar data, wind profiler radar data, sounding data, and ground automatic station data respectively and extract their respective characteristic parameters;

[0008] Step 3: Normalize the characteristic parameters and obtain the corresponding parameters from the vertical-dynamic fusion module, the microphysics-environmental fusion module and the dynamic threshold adjustment module;

[0009] Step 4: Build a multi-parameter integrated early warning model based on multiple fusion parameters, and train and verify the model;

[0010] Step 5: Input the parameters into the trained model for calculation and output the probability P of thunderstorm and gale.

[0011] The use of the neighborhood value method to remove electromagnetic interference specifically includes the following contents:

[0012] A1. Use the 5×5 neighborhood threshold method to calculate whether the number of valid numerical points around the center point in the 5×5 neighborhood exceeds the set value. If not, the center point is marked as invalid.

[0013] A2. Remove banner-shaped interference echoes, count the total number of bins in each radial direction and the number of continuous valid echo bins, locate the radial direction N with the maximum valid echo bin number, and calculate the radial echo continuity index C. If C > 20%, radial direction N is considered an interference radial direction. If C ≤ 20%, radial direction N is considered not an interference radial direction.

[0014] A3. If radial N is an interference radial, calculate the absolute value ΔS of the difference between it and the adjacent radial effective echo library. When ΔS ≤ 30, determine the starting radial N1 and ending radial N2 of the interference band. The adjacent radial is the radial N ± 20. Perform edge effective echo detection in the area N1-1 to N2+1. If there is no effective echo, remove the interference echo in this area.

[0015] A4. Repeat operations A2 and A3 until there is no interfering radial direction that meets the conditions.

[0016] The de-speed fuzzy processing specifically includes the following contents:

[0017] Traverse the set elevation layer radials, screen the radials that meet the adjacent distance reservoir wind shear less than or equal to 0.8m / s as good radials, use the radial velocity zero area to locate the weak wind area, and give priority to the radials with a continuous non-fuzzy reservoir number greater than or equal to 40 and a velocity average less than or equal to 0.3 times the Nyquist velocity as the initial reference. If the initial screening fails, relax the condition to use the radials with a continuous non-fuzzy reservoir number greater than or equal to 35 and a velocity average less than or equal to 0.3 times the Nyquist velocity as the initial reference for iterative search;

[0018] Select 3 adjacent radials to calculate the reference speed, process the adjacent radials in the clockwise or counterclockwise direction, compare the difference between the current radial speed and the reference speed, and when the difference exceeds the threshold, use V 真实 =V 观测 +2nVN Correct blur and mark processing status;

[0019] Search for the starting point that meets the set conditions at both ends of the radial direction, and process each library inward from the starting point. Use the average of the speeds of the first three libraries that meet the set conditions as a reference. If the data is interrupted, search for the starting point again, and search the radial direction for the nearest three consecutive processed libraries for complete correction.

[0020] The correction success rate in complex wind fields is improved by dynamically adjusting the wind shear threshold and the minimum number of continuous bins, the bidirectional traversal mechanism, and the marking priority mechanism.

[0021] The marking processing status includes: 1 for high confidence, 2 for low confidence, and 3 for folding or missing;

[0022] The setting conditions include: there are continuous non-missing reservoirs greater than or equal to 10 km on both sides and there are greater than or equal to 2 neighboring reservoirs marked as 1, and there are greater than or equal to 3 reservoirs marked as 1 in the three adjacent radial directions.

[0023] The second step specifically includes the following contents:

[0024] The differential reflectivity column height and vertically integrated liquid water content density were extracted from the dual-polarization radar data. The vertical column heights with differential reflectivity columns ≥1.5dB above the 0°C layer were screened. When the vertical column height obtained by screening was >1.5km, it indicated a strong correlation with downdrafts. When the vertically integrated liquid water content density was >3g / m, it indicated a sudden drop in water vapor, triggering downdrafts. In the low-elevation radial velocity field, the area with a velocity >17.2m / s was marked, and the area of continuous large value areas was counted. When the area of continuous large value areas was >50km, the area of continuous large value areas was >17.2m / s. 2 Directly related to the ground high wind risk, where the low elevation angle is ≤3°; the centroid height is calculated for the reflectivity REF≥45dBZ area , the rate of change of center of mass height When the time is >1km / 6min, it is judged as a strong sinking trend;

[0025] Extract vertical velocity and 0km-3km wind shear based on wind profiler radar data, obtain vertical velocity of wind profiler radar, and filter turbulence noise through 5-point sliding average. A strong downdraft warning is triggered when the wind speed is less than or equal to -2m / s; the horizontal wind speed components at the ground level of 0km and 3km are extracted to calculate the wind shear between 0km and 3km. ,when When the speed is greater than 15m / s, it is conducive to the development of bow echo and increases the risk of strong winds;

[0026] The convective effective potential energy and 0℃ layer height are extracted from the sounding data by the formula Calculate the convective effective potential energy CAPE, where It is due to the heat of the air mass. is the ambient virtual temperature, LFC is the free convection height, is the equilibrium height, R d represents the dry adiabatic rate, d is the sign of the differential, ln(p) represents the logarithm of pressure, when CAPE>1000J / kg, it indicates that the atmospheric stratification is extremely unstable, which is conducive to the formation of strong thunderstorms; the 0℃ layer height is located by linear interpolation of the temperature profile, when When the distance is >4km, the melting of hail releases latent heat, which significantly enhances the downdraft. is the height of the 0℃ layer;

[0027] The temperature and humidity mutation rate and maximum wind speed are extracted using ground automatic station data, and the temperature drop rate and humidity rise rate within 10 minutes are calculated. When they meet the standards, it means that the cold pool front has arrived and strong winds are about to occur on the ground. The maximum instantaneous wind speed within 10 minutes is counted. When the maximum wind speed is ≥17.2m / s, it meets the thunderstorm strong wind judgment standard and is used as the ground reality verification indicator.

[0028] The specific implementation process of the vertical-power fusion module includes:

[0029] Calculation of the change of the centroid height of the reflectivity: Track the change of the centroid height of the area with reflectivity ≥ 45dBZ. When the speed is greater than 1km / 6min, it is judged as a strong sinking trend;

[0030] Vertical speed collaborative judgment: combined with wind profiler radar vertical speed And the vertically integrated liquid water density VILD, construct the trigger condition 、 and , in, It is a parameter determined by the collaborative judgment of the reflectivity core height change and vertical velocity, which can reflect the situation in the vertical direction. is the parameter converted from the vertically integrated liquid water content according to the threshold value, For combination and The two parameters are used to obtain the parameters reflecting the velocity and water vapor content in the vertical direction;

[0031] Fusion weight calculation: Quantify the dual-source data to obtain the downdraft-related weights: ,when When it is greater than 0.7, it is determined to be a high-confidence downdraft signal.

[0032] The specific implementation process of the microphysics-environment fusion module includes:

[0033] Differential emissivity column height threshold detection: When the differential reflectivity column centroid height meets the following conditions When the hail melting potential signal is triggered, It is a risk parameter derived from the combination of the differential reflectivity column height and the zero-degree layer, reflecting the state at the microphysical and environmental levels. is the height of the differential reflectivity column;

[0034] Downdraft enhancement index: when the sounding 0℃ layer height When the microphysical height characteristics and dynamic signals are combined, the convection enhancement index is calculated as .

[0035] The specific implementation process of the dynamic threshold adjustment module includes:

[0036] Vertical speed threshold correction: based on convective effective potential energy CAPE and 0km-3km wind shear 0km-3km Dynamically adjust vertical speed threshold ;

[0037] Non-thunderstorm strong wind suppression: When the temperature and humidity mutation rate and wind speed do not meet the thunderstorm conditions, the non-strong weather suppression index is , where WS_max is the maximum wind speed, ΔT is the temperature drop rate, and ΔRH is the humidity rise rate.

[0038] The step 4 specifically includes the following contents:

[0039] B1. Obtain the downdraft-related weights obtained by the vertical-power fusion module , convection enhancement index obtained by microphysics-environment fusion module , dynamic threshold adjustment fusion module to obtain non-severe weather suppression index , vertically integrated liquid water content density , Convective Effective Potential Energy (CAPE) and 0-3km wind shear ;

[0040] B2. Normalize the parameters obtained in B1 to the same magnitude range, combine the CNN feature vectors obtained by ICA dimensionality reduction with the parameters obtained in B1, and construct a linear combination model to calculate the probability of strong winds. Output as the probability of strong winds, where is the logistic function, which maps the linear combination result to the [0,1] interval. is the probability that the output belongs to the positive class, is the normalized convective effective potential energy, is the normalized 0km-3km wind shear, is the vertically integrated liquid water density after normalization;

[0041] B3. Divide the dataset into training, validation, and test sets in proportion. Train the model on the training set using a stochastic gradient descent optimizer with a learning rate of 0.001 and a momentum of 0.9. Adjust the model parameters based on the validation set loss and metrics. Train the model after reducing the dimension of the features output by the validation set. Determine the optimal parameters based on the validation set performance. Integrate the models and evaluate and adjust the parameters using the test set.

[0042] The CNN feature vector of the ICA dimension reduction specifically includes the following:

[0043] The collected data is preprocessed by data cleaning, normalization, data annotation and parameter calculation;

[0044] Constructing the CNN feature extraction layer: including the network structure and input function, the network structure includes input layer, convolution layer, pooling layer, fully connected layer and output layer; the input layer determines the input size according to the data dimension and inputs multiple meteorological parameters as tensors. The convolution layer is used to capture features, and the pooling layer is used to reduce the data dimension and reduce the amount of calculation. The fully connected layer has two layers of neurons with 512 and 256 neurons respectively. Dropout is used to prevent overfitting. The output layer converts the feature information processed by the previous layers into the final output result;

[0045] Construct the SVM classification layer: Use independent component analysis to reduce the dimension of the feature vector output by the fully connected layer in CNN, extract independent feature components, and remove high-order correlations between data; set the sum function and parameters, use the Gaussian sum function, and search for the optimal parameter combination through the Bayesian optimization algorithm to improve the classification performance. Finally, construct the classification hyperplane. Based on the features after dimensionality reduction, solve the quadratic programming problem to determine the classification hyperplane, determine the optimal classification hyperplane, and transform the original problem into a dual problem through the Lagrange multiplier method. After obtaining the optimal solution, classify the test samples and output the probability of belonging to the positive class. .

[0046] The present invention has the following advantages: a thunderstorm and gale identification method based on multi-dimensional feature parameter fusion, by combining multi-source data fusion, multiple physical mechanism fusion, and integrating innovations in machine learning models, realizes the technology of thunderstorm and gale identification from single indicator dependence to multi-dimensional collaboration, and from static threshold to dynamic self-adaptation. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of the present application provided below in conjunction with the drawings is not intended to limit the scope of protection of the present application for which protection is claimed, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present application. The present invention is further described below in conjunction with the drawings.

[0049] like Figure 1 As shown, the present invention specifically relates to a thunderstorm gale identification method based on multi-dimensional feature parameter fusion, which specifically includes the following contents:

[0050] 1. Data preprocessing;

[0051] The radar data of the dual-polarization Doppler weather radar was gridded (resolution 0.25 km × 0.25 km), deinterferenced (5 × 5 neighborhood threshold method, threshold 55%), and defuzzified (based on adjacent radial continuity constraints) to facilitate subsequent parameter extraction. The details are as follows:

[0052] The first step is radar data preprocessing. Based on polar coordinates, to account for the variability of the number of radials per PPI scan plane and the initial scanning azimuth angle in the S-band dual-polarization radar base data, each layer of elevation radar data must contain 360 radials, spaced at 1° intervals. First, isolated interfering echoes are removed. A 5×5 neighborhood threshold method is used to calculate whether the number of valid values around the center point within this range exceeds 55%. If not, the center point is marked as invalid. Next, strip-shaped interference echoes are removed. The total bin counts and the number of continuous valid echo bins for each radial are counted. The radial N with the maximum total valid echo bin count is located. The radial echo continuity index, C, is calculated (the percentage of the valid echo variation range to the minimum value). If C is greater than 20% (radial valid echo bin count ≥ 20), radial N is initially identified as an interference radial. If so, the absolute value ΔS of the difference between the effective echo bin counts and those of the adjacent radials (i = N ± 20) is calculated. When ΔS ≤ 30, the starting and ending radials N1 and N2 of the interference band are determined. The region N1-1 to N2+1 is tested for valid edge echoes. If no valid echoes are found, the interference echoes in that region are removed. The above process of searching for interference radials is repeated until no interference radials meet the criteria. If C ≤ 20%, radial N is not an interference radial.

[0053] The second step is to perform velocity defuzzification. Based on the adjacent radial continuity constraint in polar coordinates, the radial direction of the specific elevation layer is first traversed to screen the "good radial" that satisfies the adjacent distance bin wind shear ≤ 0.8m / s. The weak wind area is located using the area with zero radial velocity (the true wind direction is perpendicular to the radial direction). Priority is given to the area with ≥40 consecutive non-fuzzy bins and an average velocity ≤ 0.3 times the Nyquist speed ( ) radial as the initial reference. If the first screening fails, relax the condition to ≥35 and iterate the search. Then select 3 adjacent good radials to calculate the reference speed, process the adjacent radials in the clockwise / counterclockwise direction, compare the current radial speed with the reference speed, and if the difference exceeds the threshold (such as 1.5 times ) when passing V 真实 =V 观测 +2nV N The fuzzy data is corrected and the processing status is marked (1 for high confidence, 2 for low confidence, and 3 for collapsed / missing; reservoirs already marked 1 are not processed again). A search is then performed at both ends of the radial direction for a starting point that meets the following conditions: "There are ≥10 km of continuous non-missing reservoirs on each side, ≥2 neighboring reservoirs marked 1, and ≥3 reservoirs marked 1 in three adjacent radial directions." The data is processed one reservoir at a time, starting from the starting point and working inward. The average of the speeds of the first three reservoirs that meet the set conditions is used as a reference. If data is interrupted, the starting point is searched again, and the correction is completed by searching the radial direction for the nearest three continuous processed reservoirs for the remaining reservoirs. Finally, the correction success rate in complex wind fields is improved by dynamically adjusting the wind shear threshold (default 0.8 m / s) and the minimum number of continuous reservoirs (default 40), using a bidirectional traversal mechanism, and a marking priority mechanism (prioritizing reservoirs marked 1).

[0054] The set conditional speed is a processed high-confidence speed in a radial direction that satisfies the conditions of small wind shear, a large number of non-missing speed measurement points, and a small average value of all non-missing speed measurement points.

[0055] Finally, gridding is performed to interpolate the processed polar-coordinate radar data into a 0.25 km × 0.25 km Cartesian grid, and bilinear interpolation is used to improve the accuracy of edge areas.

[0056] 2. Feature extraction;

[0057] The following parameters are extracted from dual-polarization radar data:

[0058] Height of differential reflectivity column (ZDR column): The height of the vertical column with ZDR ≥ 1.5dB above the 0°C layer of the environment is screened. When the value is greater than 1.5km, it indicates a strong correlation with downdrafts.

[0059] The vertically integrated liquid water density (VILD) is the standardized vertically integrated liquid water content (VIL), which is the ratio of the VIL to its corresponding echo thickness. Literature shows that when the VILD is greater than 3g / m³, it indicates a sudden drop in water vapor, triggering a downdraft.

[0060] In the radial velocity field at low elevation angles (≤3°), areas with values >17.2 m / s were marked, and the area of continuous large value areas was counted. When the area was >50 km², it was directly associated with the risk of strong winds on the ground.

[0061] Calculate the centroid height for the reflectivity REF ≥ 45dBZ area , the rate of change of center of mass height When the rate is >1km / 6min, it is judged to be a strong sinking trend.

[0062] Extracting dynamic parameters based on wind profiler radar data:

[0063] Vertical speed ( ): Get the vertical velocity of the wind profiler radar. Filter the turbulence noise by 5-point sliding average. A strong downdraft warning is triggered when the speed is ≤-2m / s.

[0064] 0km-3km wind shear ( ): Extract the horizontal wind speed components at the ground and 3km altitude ( , )and( , ), calculate the 0km-3km wind shear ,when When the speed is greater than 15m / s, it is conducive to the development of bow echo and the risk of strong winds is enhanced. is the u wind at 3km height, is the u wind at 0 km altitude (ground level), is the wind v at a height of 3 km, is the v wind at 0 km altitude.

[0065] Extract environmental parameters from sounding data:

[0066] Convective Available Potential Energy (CAPE): When an air parcel moves vertically in an unstable atmosphere (where buoyancy is greater than gravity), its vertical velocity increases continuously, meaning that its kinetic energy increases continuously. This increased kinetic energy is converted from a portion of the energy stored in the unstable atmosphere. Therefore, this portion of energy that can be converted into kinetic energy is called Convective Available Potential Energy (CAPE). The calculation formula is as follows:

[0067] ,in: It is due to the heat of the air mass. is the ambient virtual temperature, LFC is the free convection height, is the equilibrium height, R d represents the dry adiabatic rate, d is the sign of the differential, and ln(p) represents the logarithm of pressure.

[0068] When CAPE>1000J / kg, it indicates that the atmospheric stratification is extremely unstable, which is conducive to the formation of severe thunderstorms.

[0069] Zero degree layer altitude ( ): Locate the 0℃ layer height by linear interpolation of the temperature profile, When the altitude is >4km, the falling hail melts and releases latent heat, which significantly enhances the downdraft.

[0070] Extract and verify parameters using ground automatic station data:

[0071] Temperature and humidity mutation rate (ΔT / ΔRH): Calculate the temperature drop rate (ΔT≤-3°C) and humidity rise rate (ΔRH≥15%) within 10 minutes. When the standards are met, it indicates that the cold pool front has arrived and strong winds are about to occur on the ground.

[0072] Maximum wind speed (WS_max): The maximum instantaneous wind speed within 10 minutes is counted. When WS_max ≥ 17.2 m / s, it meets the thunderstorm and gale criteria of the China Meteorological Administration and serves as a ground truth verification indicator.

[0073] 3. Parameter fusion;

[0074] Multi-source parameter standardization: To address the heterogeneity of dual-polarization radar, wind profiler radar, sounding, and ground station data, the following preprocessing is performed:

[0075] Dimension normalization: Use the minimum-maximum normalization method to eliminate dimensional differences: in, and is the extreme value of the training set parameter, and the normalized data is mapped to the [0,1] interval.

[0076] Principal Component Analysis (PCA) Dimensionality Reduction: Project high-dimensional parameters into low-dimensional space through linear transformation, retaining principal components with cumulative contribution rates ≥ 95%: , where W is the eigenvector matrix and Y is the eigenvector after dimensionality reduction. This step eliminates redundant information between parameters (such as VILD and correlation), improve the efficiency of model training.

[0077] The vertical-dynamic fusion module uses dual-polarization radar and wind profiler radar data to collaboratively assess downdraft intensity. This module aims to accurately determine the triggering conditions and credibility of downdrafts by comprehensively considering key factors such as height variations in areas of strong reflectivity, vertical velocity, and vertically integrated liquid water density using multi-source radar data.

[0078] Calculation of the change in the centroid height of reflectivity: Tracking the change in the centroid height of the strong reflectivity area (≥45dBZ): ,in, is the reflectivity centroid height at the current moment, is the height of the reflectivity centroid at the previous moment, =6min, when When the speed is greater than 1km / 6min, it is judged to be a strong sinking trend.

[0079] Vertical speed collaborative judgment: combined with wind profiler radar vertical speed And the vertically integrated liquid water content density VILD, construct the trigger condition:

[0080] ,

[0081] ,

[0082] ;

[0083] in, It is a parameter determined by the collaborative judgment of the reflectivity core height change and vertical velocity, which can reflect the situation in the vertical direction. is the parameter converted from the vertically integrated liquid water content according to the threshold value, For combination and The two parameters are used to obtain the parameters reflecting the velocity and water vapor content in the vertical direction.

[0084] Fusion weight calculation: Quantify the dual-source data to obtain the downdraft-related weights:

[0085] ,

[0086] when When it is greater than 0.7, it is determined to be a high-confidence downdraft signal.

[0087] Microphysics-environment coupling module: Based on the dual-polarization radar microphysical characteristics and sounding environment parameters, the identification logic of hail melting enhancing downdraft is established.

[0088] Differential reflectivity column height threshold detection: When the centroid height of the differential reflectivity column (ZDR ≥ 1.5dB) meets the following conditions, a potential hail melt signal is triggered:

[0089] ,

[0090] in, It is a risk parameter derived from the combination of the differential reflectivity column height and the zero-degree layer, reflecting the state at the microphysical and environmental levels. is the differential reflectivity column height.

[0091] Downdraft enhancement index: when the sounding 0℃ layer height When the microphysical height characteristics and dynamic signals are combined, the enhancement index is calculated. .

[0092] Dynamic threshold adjustment module: Adaptively adjusts the warning threshold based on environmental energy and wind shear. The sudden drop in temperature and the sudden rise in humidity reflect the evaporative cooling of raindrops, forming a density flow (outflow boundary), which is an important triggering mechanism for thunderstorms and strong winds.

[0093] Vertical speed threshold correction: Based on the convective effective potential energy (CAPE) and 0km-3km wind shear (shear 0km-3km ) Dynamically adjust the vertical speed threshold ;

[0094] The high-energy environment combined with strong wind shear can easily maintain deep convection, and the threshold for triggering downdrafts is lowered.

[0095] Non-thunderstorm strong wind suppression: When the temperature and humidity mutation rate and wind speed do not meet the thunderstorm conditions, that is, the cold pool is not triggered, the non-severe weather suppression index is obtained as .

[0096] 4. SVM-CNN integrated learning model;

[0097] 1. Dataset Preparation: Preprocess the collected data, including data cleaning, normalization, data labeling, and parameter calculation. Data labeling involves labeling the data as "windy" and "non-windy" samples based on meteorological observation records and wind standards, with non-windy samples labeled as 0 and windy samples as 1.

[0098] 2. CNN feature extraction layer: includes network structure and input function. The network structure is as follows: the first layer is the input layer, which determines the input size according to the data dimension and inputs multiple meteorological parameters as tensors. The second layer is the convolution layer, with 4 convolution layers. The first layer has a convolution kernel size of 7×7, a step size of 2, and 16 convolution kernels, which are used to preliminarily extract features over a wide range. The second layer has a convolution kernel size of 5×5, a step size of 2, and 32 convolution kernels, which further refine the feature extraction. The third and fourth layers have a convolution kernel size of 3×3, a step size of 1, and 64 and 128 convolution kernels, respectively, to capture finer local features. This is followed by a pooling layer, with an average pooling layer added after every two convolution layers. The pooling window size is 2×2 to reduce the data dimension and the amount of computation. Then comes the fully connected layer, which flattens the pooled features and inputs them into the fully connected layer. Two fully connected layers are set up with 512 and 256 neurons respectively. Dropout (such as a probability of 0.5) is used to prevent overfitting. Finally, it comes to the output layer, which converts the feature information processed by the previous layers into the final output result of the model. The activation function used is the ReLU activation function.

[0099] 3. SVM classification layer: First, perform data dimensionality reduction, that is, use independent component analysis (ICA) to reduce the dimensionality of the feature vector output by the CNN fully connected layer, extract independent feature components, and remove high-order correlations between data. Then set the sum function and parameters, use the Gaussian sum function, and search for the optimal parameter combination through the Bayesian optimization algorithm to improve classification performance. Finally, construct a classification hyperplane, and based on the features after dimensionality reduction, solve the quadratic programming problem to determine the classification hyperplane and determine the optimal classification hyperplane. Use the Lagrange multiplier method to transform the original problem into a dual problem for solution. After obtaining the optimal solution, classify the test sample and output the probability of belonging to the positive class. .

[0100] 4. Multi-parameter integrated warning model: First, multiple key parameters are obtained from the previous three modules, such as the downdraft-related weights output by the vertical-dynamic fusion module. It combines factors such as the change in reflectivity core height and vertical velocity to reflect the vertical dynamic characteristics; the convection enhancement index , which can be calculated from the intensity change and energy accumulation of the convective system; non-severe weather suppression index , by comprehensively judging environmental factors such as temperature, humidity and specific meteorological conditions, it is used to exclude non-severe convective weather conditions; and vertically integrated liquid water content density , Convective Effective Potential Energy (CAPE), 0km-3km wind shear The CNN feature vectors obtained through ICA dimensionality reduction are then combined with the parameters output by the above three modules. To ensure that each parameter has a reasonable weight and influence in the model, these parameters are first normalized to the same magnitude range.

[0101] Construct a linear combination model to calculate the probability P of strong winds:

[0102] ,

[0103] in, is a logistic function that maps the linear combination result to the interval [0,1] as the output of the probability of strong wind. The weights of each parameter are determined based on the analysis of historical meteorological data and meteorological theory. is the probability that the output belongs to the positive class, is the normalized convective effective potential energy, is the normalized 0km-3km wind shear, is the normalized vertically integrated liquid water density.

[0104] Model Ensemble and Training: The dataset was split into training, validation, and test sets (6:2:2). A CNN was trained using the stochastic gradient descent (SGD) optimizer with a learning rate of 0.001 and a momentum of 0.9. Network parameters were adjusted based on validation set loss and metrics. After dimensionality reduction, the features output by the CNN on the validation set were used to train an SVM. The optimal parameters were determined based on validation set performance. The models were ensembled, and the test set was used to evaluate and fine-tune the parameters.

[0105] Evaluation indicators: accuracy, recall rate, F1 value, hit rate (TS score), and false alarm rate.

[0106] The real-time collected data is preprocessed, and multi-source parameters are calculated from the data, and then the multi-source parameters are normalized; the normalized multi-source features (ZDR column height, CAPE, , ΔT, ΔRH, etc.), calculate the fusion model parameters , , , input the parameters into the trained model for calculation, and output the probability of thunderstorm and gale P. If P ≥ 0.7, an early warning is issued, otherwise continuous monitoring is performed.

[0107] The real-time collected data includes:

[0108] Dual-polarization Doppler radar data: S-band dual-polarization radar basic data (reflectivity ZH, differential reflectivity ZDR, radial velocity V, etc.) are collected with a time resolution of 6 minutes. Preprocessing is performed, including electromagnetic interference removal, velocity deblurring, and gridding.

[0109] Wind profiler radar data: Collect wind profiler radar data. The selected data is hourly data from the same period as the dual-polarization radar data, namely ROBS data.

[0110] Ground automatic station data: minute-level data from the same period as the dual-polarization radar is selected to extract temperature, humidity, wind direction, wind speed, and maximum wind parameters.

[0111] Sounding data: The sounding data selected are those in a time period similar to the dual-polarization radar data.

[0112] The foregoing description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein and should not be construed as excluding other embodiments. Rather, the present invention is capable of various other combinations, modifications, and improvements, and is capable of modification within the scope of the concepts described herein, through the above teachings, or through techniques or knowledge in the relevant fields. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention are intended to be within the scope of the appended claims.

Claims

1. A thunderstorm and gale identification method based on multi-dimensional feature parameter fusion, characterized by: The identification method comprises: Step 1: The dual-polarization Doppler weather radar data is subjected to the neighborhood value method to remove electromagnetic interference, and then subjected to de-velocity blurring and gridding. Step 2: Obtain dual-polarization radar data, wind profiler radar data, sounding data, and ground automatic station data respectively and extract their respective characteristic parameters; Step 3: Normalize the characteristic parameters and obtain the corresponding parameters from the vertical-dynamic fusion module, the microphysics-environmental fusion module and the dynamic threshold adjustment module; Step 4: Build a multi-parameter integrated early warning model based on multiple fusion parameters, and train and verify the model; Step 5: Input the parameters into the trained model for calculation and output the probability P of thunderstorm and gale.

2. The thunderstorm and gale identification method based on multi-dimensional feature parameter fusion according to claim 1 is characterized by: The use of the neighborhood value method to remove electromagnetic interference specifically includes the following contents: A1. Use the 5×5 neighborhood threshold method to calculate whether the number of valid numerical points around the center point in the 5×5 neighborhood exceeds the set value. If not, the center point is marked as invalid. A2. Remove banner-shaped interference echoes, count the total number of bins in each radial direction and the number of continuous valid echo bins, locate the radial direction N with the maximum valid echo bin number, and calculate the radial echo continuity index C. If C > 20%, radial direction N is considered an interference radial direction. If C ≤ 20%, radial direction N is considered not an interference radial direction. A3. If radial N is an interference radial, calculate the absolute value ΔS of the difference between it and the adjacent radial effective echo library. When ΔS ≤ 30, determine the starting radial N1 and ending radial N2 of the interference band. The adjacent radial is the radial N ± 20. Perform edge effective echo detection in the area N1-1 to N2+1. If there is no effective echo, remove the interference echo in this area. A4. Repeat operations A2 and A3 until there is no interfering radial direction that meets the conditions.

3. The thunderstorm and gale identification method based on multi-dimensional feature parameter fusion according to claim 1 is characterized in that: The de-speed fuzzy processing specifically includes the following contents: Traverse the set elevation layer radials, screen the radials that meet the adjacent distance reservoir wind shear less than or equal to 0.8m / s as good radials, use the radial velocity zero area to locate the weak wind area, and give priority to the radials with a continuous non-fuzzy reservoir number greater than or equal to 40 and a velocity average less than or equal to 0.3 times the Nyquist velocity as the initial reference. If the initial screening fails, relax the condition to use the radials with a continuous non-fuzzy reservoir number greater than or equal to 35 and a velocity average less than or equal to 0.3 times the Nyquist velocity as the initial reference for iterative search; Select 3 adjacent radials to calculate the reference speed, process the adjacent radials in the clockwise or counterclockwise direction, compare the difference between the current radial speed and the reference speed, and when the difference exceeds the threshold, use V 真实 =V 观测 +2nV N Correct blur and mark processing status; Search for the starting point that meets the set conditions at both ends of the radial direction, and process each library inward from the starting point. Use the average of the speeds of the first three libraries that meet the set conditions as a reference. If the data is interrupted, search for the starting point again, and search the radial direction for the nearest three consecutive processed libraries for complete correction. The correction success rate in complex wind fields is improved by dynamically adjusting the wind shear threshold and the minimum number of continuous bins, the bidirectional traversal mechanism, and the marking priority mechanism.

4. The method for identifying thunderstorms and gale winds based on multi-dimensional feature parameter fusion according to claim 3, characterized in that: The marking processing status includes: 1 for high confidence, 2 for low confidence, and 3 for folding or missing; The setting conditions include: there are continuous non-missing reservoirs greater than or equal to 10 km on both sides and there are greater than or equal to 2 neighboring reservoirs marked as 1, and there are greater than or equal to 3 reservoirs marked as 1 in the three adjacent radial directions.

5. The thunderstorm and gale identification method based on multi-dimensional feature parameter fusion according to claim 1 is characterized in that: The second step specifically includes the following contents: The differential reflectivity column height and vertically integrated liquid water content density were extracted from the dual-polarization radar data. The vertical column heights with differential reflectivity columns ≥1.5dB above the 0°C layer were screened. When the vertical column height obtained by screening was >1.5km, it indicated a strong correlation with downdrafts. When the vertically integrated liquid water content density was >3g / m, it indicated a sudden drop in water vapor, triggering downdrafts. In the low-elevation radial velocity field, the area with a velocity >17.2m / s was marked, and the area of continuous large value areas was counted. When the area of continuous large value areas was >50km, the area of continuous large value areas was >17.2m / s. 2 Directly related to the ground high wind risk, where the low elevation angle is ≤3°; the centroid height is calculated for the reflectivity REF≥45dBZ area , the rate of change of center of mass height When the time is >1km / 6min, it is judged as a strong subsidence trend; Based on the vertical velocity and 0km-3km wind shear extracted from the wind profiler radar data, the vertical velocity of the wind profiler radar is obtained, and the turbulence noise is filtered through a 5-point sliding average. A strong downdraft warning is triggered when the wind speed is less than or equal to -2m / s; the horizontal wind speed components at the ground level of 0km and 3km are extracted to calculate the wind shear between 0km and 3km. ,when When the speed is greater than 15m / s, it is conducive to the development of bow echo and increases the risk of strong winds. is the vertical speed; The convective effective potential energy and 0℃ layer height are extracted from the sounding data by the formula Calculate the convective effective potential energy CAPE, where It is due to the heat of the air mass. is the ambient virtual temperature, LFC is the free convection height, is the equilibrium height, R d represents the dry adiabatic rate, d is the sign of the differential, ln(p) represents the logarithm of pressure, when CAPE>1000J / kg, it indicates that the atmospheric stratification is extremely unstable, which is conducive to the formation of strong thunderstorms; the 0℃ layer height is located by linear interpolation of the temperature profile, when When the distance is >4km, the melting of hail releases latent heat, which significantly enhances the downdraft. is the height of the 0℃ layer; The temperature and humidity mutation rate and maximum wind speed are extracted using ground automatic station data, and the temperature drop rate and humidity rise rate within 10 minutes are calculated. When they meet the standards, it means that the cold pool front has arrived and strong winds are about to occur on the ground. The maximum instantaneous wind speed within 10 minutes is counted. When the maximum wind speed is ≥17.2m / s, it meets the thunderstorm strong wind judgment standard and is used as the ground reality verification indicator.

6. The method for identifying thunderstorms and gale winds based on multi-dimensional feature parameter fusion according to claim 5, characterized in that: The specific implementation process of the vertical-power fusion module includes: Calculation of the change of the centroid height of the reflectivity: Track the change of the centroid height of the area with reflectivity ≥ 45dBZ. When the speed is greater than 1km / 6min, it is judged as a strong sinking trend; Vertical speed collaborative judgment: vertical speed combined with wind profiler radar And the vertically integrated liquid water density VILD, construct the trigger condition 、 and ， in, It is a parameter determined by the change in reflectivity core height and vertical velocity, reflecting the situation in the vertical direction. is the parameter converted from the vertically integrated liquid water content according to the threshold value, For combination and The two parameters are used to obtain the parameters reflecting the velocity and water vapor content in the vertical direction; Fusion weight calculation: Quantify the dual-source data to obtain the downdraft-related weights: , when the weight of the sinking airflow When it is greater than 0.7, it is determined to be a high-confidence downdraft signal.

7. The method for identifying thunderstorms and gale winds based on multi-dimensional feature parameter fusion according to claim 5, characterized in that: The specific implementation process of the microphysics-environment fusion module includes: Differential emissivity column height threshold detection: When the differential reflectivity column centroid height meets the following conditions When the hail melting potential signal is triggered, It is a risk parameter derived from the combination of the differential reflectivity column height and the zero-degree layer, reflecting the state at the microphysical and environmental levels. is the differential reflectivity column height; Downdraft enhancement index: when the sounding 0℃ layer height When the microphysical height characteristics and dynamic signals are combined, the convection enhancement index is calculated as , It is an indicator of convection enhancement.

8. The method for identifying thunderstorms and gale winds based on multi-dimensional feature parameter fusion according to claim 5, characterized in that: The specific implementation process of the dynamic threshold adjustment module includes: Vertical speed threshold correction: based on convective effective potential energy CAPE and 0km-3km wind shear 0km-3km Dynamically adjust vertical speed threshold ; Non-thunderstorm strong wind suppression: When the temperature and humidity mutation rate and wind speed do not meet the thunderstorm conditions, the non-strong weather suppression index is , where WS_max is the maximum wind speed, ΔT is the temperature drop rate, and ΔRH is the humidity rise rate. It is an indicator of non-severe weather suppression.

9. The method for identifying thunderstorms and gale winds based on multi-dimensional feature parameter fusion according to claim 5, characterized in that: The step 4 specifically includes the following contents: B1. Obtain the downdraft-related weights obtained by the vertical-power fusion module , convection enhancement index obtained by microphysics-environment fusion module , dynamic threshold adjustment fusion module to obtain non-severe weather suppression index , vertically integrated liquid water content density , Convective Effective Potential Energy (CAPE) and 0km-3km wind shear ; B2. Normalize the parameters obtained in B1 to the same magnitude range, combine the CNN feature vectors obtained by ICA dimensionality reduction with the parameters obtained in B1, and construct a linear combination model to calculate the probability of strong winds. Output as the probability of strong winds, where is the logistic function, which maps the linear combination result to the [0,1] interval. is the probability that the output belongs to the positive class, is the normalized convective effective potential energy, is the normalized 0km-3km wind shear, is the vertically integrated liquid water density after normalization; B3. Divide the dataset into training, validation, and test sets in proportion. Train the model on the training set using a stochastic gradient descent optimizer with a learning rate of 0.001 and a momentum of 0.

9. Adjust the model parameters based on the validation set loss and metrics. Train the model after reducing the dimension of the features output by the validation set. Determine the optimal parameters based on the validation set performance. Integrate the models and evaluate and adjust the parameters using the test set.

10. The method for identifying thunderstorms and gale winds based on multi-dimensional feature parameter fusion according to claim 9, characterized in that: The CNN feature vector of the ICA dimension reduction specifically includes the following: The collected data is preprocessed by data cleaning, normalization, data annotation and parameter calculation; Constructing the CNN feature extraction layer: including the network structure and input function, the network structure includes input layer, convolution layer, pooling layer, fully connected layer and output layer; the input layer determines the input size according to the data dimension and inputs multiple meteorological parameters as tensors. The convolution layer is used to capture features, and the pooling layer is used to reduce the data dimension and reduce the amount of calculation. The fully connected layer has two layers of neurons with 512 and 256 neurons respectively. Dropout is used to prevent overfitting. The output layer converts the feature information processed by the previous layers into the final output result; Construct the SVM classification layer: Use independent component analysis to reduce the dimension of the feature vector output by the fully connected layer in CNN, extract independent feature components, and remove high-order correlations between data; set the sum function and parameters, use the Gaussian sum function, and search for the optimal parameter combination through the Bayesian optimization algorithm to improve the classification performance. Finally, construct the classification hyperplane. Based on the features after dimensionality reduction, solve the quadratic programming problem to determine the classification hyperplane, determine the optimal classification hyperplane, and transform the original problem into a dual problem through the Lagrange multiplier method. After obtaining the optimal solution, classify the test samples and output the probability of belonging to the positive class. .

Images