Hail prediction method and device, medium and program product

Abstract

The invention discloses a hail prediction method and device, a medium and a program product, and relates to the technical field of meteorology, and the method comprises the steps: obtaining a neural network model; acquiring historical lightning characteristic variable data and historical hail characteristic variable data; dividing the historical lightning characteristic variable data and the historical hail characteristic variable data into a training set and a verification set of corresponding characteristic variables; training the neural network model through the training set and the verification set of the corresponding feature variables, and taking the trained neural network model as a target neural network model; acquiring current lightning characteristic variable data; determining a dynamic weighting loss function according to the target neural network model; and predicting hail weather according to the current lightning characteristic variable data and the dynamic weighted loss function to obtain hail prediction parameters. According to the technology provided by the invention, misjudgment caused by forecasting the hail by independently using the radar reflectivity factor is reduced, and the hail forecasting precision is improved by indicating the factor through the lightning characteristic data.

Description

Technical Field

[0001] This application relates to the field of meteorological technologies, and particularly to a hail prediction method, device, medium and program product. Background Art

[0002] Current hail forecasts mainly rely on meteorological radars and numerical weather prediction models (ECMWF, European Centre for Medium-Range Weather Forecasts), and their main prediction methods are based on precipitation intensity, convective parameters (such as wind shear), and radar echoes (such as maximum reflectivity). However, these methods have the disadvantages of poor timeliness, high false alarm rate, and inaccurate convective intensity estimation.

[0003] Since the formation time of hail is usually short, while the time for a radar to complete a volume scan is long, the information obtained solely by traditional radar methods is not fast enough; relying only on radar echoes for hail identification, and the most commonly used in the business is non-polarized radar, which cannot provide information on the shape of particles in the cloud and is prone to misjudging heavy precipitation as hail, resulting in a high false alarm rate; the formation of hail depends on convective intensity, however, existing radar echo and convective parameter estimation methods are difficult to accurately reflect the energy distribution and airflow structure in the cloud. Summary of the Invention

[0004] This application provides a hail prediction method, device, medium and program product. The method includes obtaining a neural network model; obtaining historical lightning feature variable data and historical hail feature variable data; dividing the historical lightning feature variable data and historical hail feature variable data into training sets and validation sets of corresponding feature variables; training the neural network model through the training sets and validation sets of corresponding feature variables, and using the trained neural network model as the target neural network model; obtaining current lightning feature variable data; determining a dynamic weighted loss function according to the target neural network model; predicting hail weather according to the current lightning feature variable data and the dynamic weighted loss function to obtain hail prediction parameters, where the hail prediction parameters include: hail occurrence probability parameter, hail prediction occurrence time parameter, hail prediction diameter parameter, and hail prediction location parameter. The technology of this application reduces misjudgments caused by predicting hail solely based on radar reflectivity factor, and improves the prediction accuracy of hail through lightning feature data indication factors.

[0005] This application provides a hail prediction method, and the method includes: Obtaining a neural network model; Obtaining historical lightning feature variable data and historical hail feature variable data; Dividing the historical lightning feature variable data and historical hail feature variable data into training sets and validation sets of corresponding feature variables; Train a neural network model using the training set and validation set of corresponding feature variables, and use the trained neural network model as the target neural network model; Obtain the current lightning feature variable data; Determine a dynamic weighted loss function according to the target neural network model; Predict the hail weather based on the current lightning feature variable data and the dynamic weighted loss function to obtain hail prediction parameters, where the hail prediction parameters include: hail occurrence probability parameter, hail prediction occurrence time parameter, hail prediction diameter parameter, and hail prediction location parameter.

[0006] This application also provides a hail prediction device, including: An acquisition module, configured to acquire a neural network model; acquire historical lightning feature variable data and historical hail feature variable data; acquire current lightning feature variable data; A division module, configured to divide the historical lightning feature variable data and historical hail feature variable data into a training set and a validation set of corresponding feature variables; A training module, configured to train the neural network model using the training set and validation set of corresponding feature variables, and use the trained neural network model as the target neural network model; A determination module, configured to determine a dynamic weighted loss function according to the target neural network model; A prediction module, configured to predict the hail weather based on the current lightning feature variable data and the dynamic weighted loss function to obtain hail prediction parameters, where the hail prediction parameters include: hail occurrence probability parameter, hail prediction occurrence time parameter, hail prediction diameter parameter, and hail prediction location parameter.

[0007] This application also provides an electronic device, including: a memory for storing a computer program; a processor for implementing the steps of the hail prediction method when executing the computer program, and the method includes: Acquire a neural network model; Acquire historical lightning feature variable data and historical hail feature variable data; Divide the historical lightning feature variable data and historical hail feature variable data into a training set and a validation set of corresponding feature variables; Train the neural network model using the training set and validation set of corresponding feature variables, and use the trained neural network model as the target neural network model; Acquire the current lightning feature variable data; Determine a dynamic weighted loss function according to the target neural network model; Predict the hail weather based on the current lightning feature variable data and the dynamic weighted loss function to obtain hail prediction parameters, where the hail prediction parameters include: hail occurrence probability parameter, hail prediction occurrence time parameter, hail prediction diameter parameter, and hail prediction location parameter.

[0008] This application also provides a computer-readable storage medium storing a computer program, where the computer program, when executed by a processor, implements the steps of the hail prediction method, and the method includes: Obtain a neural network model; Obtain historical lightning feature variable data and historical hail feature variable data; Divide the historical lightning feature variable data and historical hail feature variable data into training sets and validation sets for corresponding feature variables; Train the neural network model with the training sets and validation sets for corresponding feature variables, and use the trained neural network model as the target neural network model; Obtain the current lightning feature variable data; Determine the dynamic weighted loss function according to the target neural network model; Predict the hail weather based on the current lightning feature variable data and the dynamic weighted loss function to obtain hail prediction parameters, where the hail prediction parameters include: hail occurrence probability parameter, hail prediction occurrence time parameter, hail prediction diameter parameter, and hail prediction location parameter.

[0009] This application also provides a computer program product including a computer program, where the computer program, when executed by a processor, implements the steps of the hail prediction method, and the method includes: Obtain a neural network model; Obtain historical lightning feature variable data and historical hail feature variable data; Divide the historical lightning feature variable data and historical hail feature variable data into training sets and validation sets for corresponding feature variables; Train the neural network model with the training sets and validation sets for corresponding feature variables, and use the trained neural network model as the target neural network model; Obtain the current lightning feature variable data; Determine the dynamic weighted loss function according to the target neural network model; Predict the hail weather based on the current lightning feature variable data and the dynamic weighted loss function to obtain hail prediction parameters, where the hail prediction parameters include: hail occurrence probability parameter, hail prediction occurrence time parameter, hail prediction diameter parameter, and hail prediction location parameter.

[0010] Through this application, since the method includes obtaining a neural network model; obtaining historical lightning feature variable data and historical hail feature variable data; dividing the historical lightning feature variable data and historical hail feature variable data into training sets and validation sets of corresponding feature variables; training the neural network model through the training sets and validation sets of corresponding feature variables, and using the trained neural network model as the target neural network model; obtaining current lightning feature variable data; determining a dynamic weighted loss function according to the target neural network model; and predicting hail weather based on the current lightning feature variable data and the dynamic weighted loss function to obtain hail prediction parameters. Therefore, the technology of this application reduces the misjudgment caused by predicting hail solely based on radar reflectivity factor, and improves the prediction accuracy of hail through the lightning feature data indication factor.

[0011] The technical solution of this application can effectively improve the prediction accuracy of hail size and falling area, establish an analysis model of hail formation mechanism based on lightning activities, and better understand the growth process of hail in thunderstorms; and can develop an automated hail warning system, integrate the prediction model into the meteorological forecasting system, and improve the practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the embodiments of this application, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of this application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0013] Figure 1 It is the first flowchart of the hail prediction method provided by the embodiment of this application; Figure 2 It is the second flowchart of the hail prediction method provided by the embodiment of this application; Figure 3 It is the specific flowchart of the hail prediction method provided by the embodiment of this application; Figure 4 It is the structural diagram of the hail prediction device provided by the embodiment of this application; Figure 5 It is an exemplary system that can be used to implement the various embodiments described in this application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0014] The following will clearly and completely describe the technical solutions in the embodiments of this application with reference to the drawings in the embodiments of this application. Obviously, the described embodiments are only some of the embodiments of this application, rather than all of them. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the protection scope of this application.

[0015] It should be noted that in the description of this application, the terms "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, such that a process, method, article or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or device. The terms "first", "second", etc. in this application are used to distinguish similar objects and not to describe a specific order or sequence.

[0016] To enable those skilled in the art of this technology to better understand the solution of this application, the following further details this application in conjunction with the accompanying drawings and specific embodiments.

[0017] In combination with the specific application environment architecture or specific hardware architecture on which the execution of the hail prediction method depends, the specific application environment architecture or specific hardware architecture is described herein.

[0018] Hail is a severe weather phenomenon that can easily cause agricultural losses, traffic disruptions, and property damage.

[0019] In recent years, research has shown that there is a close relationship between lightning characteristics (polarity, occurrence height, frequency, etc.) and hail formation; in a severe convective system, a sharp increase in the number of positive lightning strikes, an increase in the average height of lightning, and a drastic change in lightning frequency may all be precursors to hail formation. Therefore, by combining lightning data with neural network technology here, it is expected to break through the limitations of traditional radar forecasting and improve the accuracy and lead time of hail warnings.

[0020] In addition, lightning data can provide information about the charge distribution in clouds, which is crucial for understanding the microphysical mechanism of hail growth; for example, in a supercell storm, an increase in positive lightning may mean enhanced ice crystal growth in the upper and middle layers of the thunderstorm, leading to the formation of larger hailstones. Therefore, using artificial intelligence in combination with lightning data for forecasting can effectively improve the prediction ability of hail.

[0021] The prior art "Hail Prediction Method and System Based on the Jump Characteristics of Spaceborne Lightning Observation Data" is based on generating a change rate from lightning frequency, determining whether the lightning jump condition is met, and then using it to predict hail.

[0022] Disadvantages: The algorithm is single, the data information density is low, the method is unreliable, and the latest neural network-related technologies are not used.

[0023] An embodiment of this application provides a hail prediction method, as Figure 1 shown, the method includes: Obtain a neural network model; Obtain historical lightning feature variable data and historical hail feature variable data; Divide the historical lightning feature variable data and historical hail feature variable data into training sets and validation sets for the corresponding feature variables; Train a neural network model using the training sets and validation sets of the corresponding feature variables, and use the trained neural network model as the target neural network model; Obtain the current lightning feature variable data; Determine a dynamic weighted loss function according to the target neural network model; Predict hailstorm weather based on the current lightning feature variable data and the dynamic weighted loss function to obtain hail prediction parameters, where the hail prediction parameters include: hail occurrence probability parameter, hail predicted occurrence time parameter, hail predicted diameter parameter, and hail predicted position parameter.

[0024] It can be understood that this application proposes a hail prediction method based on lightning feature data (positive and negative polarities, three-dimensional coordinates, frequencies), using a neural network and combining physical constraint modeling. This method can achieve: 1. Based on the development and change characteristics of multi-dimensional information of lightning activities, a neural network model is trained to obtain a hail prediction weight model, improving the hail prediction accuracy; 2. On this basis, further combine the polarization radar reflectivity to correct the model weight; 3. Combine wind field data parameters to optimize the hail movement trajectory prediction.

[0025] Based on the above method, the forecast accuracy of hail size and falling area can be effectively improved, and an analysis model of hail formation mechanism based on lightning activities can be established to better understand the growth process of hail in thunderstorms; and an automated hail warning system can be developed, integrating the prediction model into the meteorological forecast system to improve the practical application value.

[0026] An embodiment of this application provides a hail prediction method, as Figure 2 shown, the method includes: This application proposes a hail warning method based on lightning features and neural networks, which combines lightning observations, radar data, and wind field parameters, and uses a neural network to predict the probability, size, and falling area of hail occurrence.

[0027] Step S01, obtain a neural network model; Obtain the historical lightning feature variable data and historical hail feature variable data; Divide the historical lightning feature variable data and historical hail feature variable data into training sets and validation sets for the corresponding feature variables; Here, data collection includes a lightning detection system (such as LMA / BLNET), meteorological radar data, wind profiler radar data, reanalysis data (such as ERA5 / WRF forecast output), and hail record data (such as meteorological station information and user-provided hail information data collected through mobile devices. The data requires a sample set of 50 hail days and at least 150 non-hail days in thunderstorm weather).

[0028] Determine the spatial grid: For example, a horizontally uniform longitude-latitude grid and a vertically uniform grid can be adopted to unify the multi-source data into grid data.

[0029] Step S02: Convert the historical lightning feature variable data and historical hail feature variable data into tensor data recognizable by the neural network model.

[0030] Feature extraction: Convert the multi-source original collected data into tensor data that can be input into the neural network.

[0031] Step S03: Train the neural network model with the training set and validation set corresponding to the feature variables, and use the trained neural network model as the target neural network model.

[0032] Specifically, the division ratio: Usually, the dataset corresponding to the feature variables can be divided in the ratio of 70% training set, 15% validation set, and 15% test set.

[0033] Extract the lightning feature variables as input X and the hail feature variables as output Y from the divided dataset.

[0034] Construct and train a neural network model based on, for example, Transformer, and select an appropriate loss function according to the actual situation to train the model.

[0035] Step S031: Clean the training set and validation set corresponding to the feature variables; Train the neural network model with the cleaned training set corresponding to the feature variables, and initialize the loss function and optimizer of the neural network model; Verify the neural network model with the cleaned validation set corresponding to the feature variables, calculate the loss value and determination coefficient of the neural network model; Adjust and optimize the loss value and determination coefficient of the neural network model according to the training results.

[0036] Specifically, organize the collected data into the input and output required for neural network model training, including extracting lightning feature variables from lightning data, including: Lightning-related feature variables: ① Lightning time ( T l): Absolute timestamp (or normalized time); ② Polarity ( E l ): Positive / negative lightning (0 / 1 encoding); ③ Three-dimensional coordinates ( L3 l ): Latitude and longitude + altitude, or rasterized into a 3D spatial distribution map; ④ Minute-level frequency ( F l ): Can be counted as the number of lightning strikes within a grid per unit time; Output hail characteristic variables: including: ① Whether hail occurs ( TF h ): 0 / 1 binary classification; ② Hail diameter ( D h ): Can be used as a regression target; ③ Hail landing point ( L2 h ): Used for subsequent fall area prediction; ④ Time ( T h ): Can be used as the time window of the target event.

[0037] Hail prediction: Divide historical data into training sets and validation sets according to certain rules, and input them into a neural network model (such as based on Transformer) for training. The input is all lightning characteristic variable data, and the output is all hail characteristic variable data; combined with a targeted loss function, complete the prediction of all hail characteristic variables.

[0038] Step S04, obtain the current lightning characteristic variable data; Determine the dynamic weighted loss function according to the target neural network model.

[0039] Step S041, the dynamic weighted loss function is L total : ; Among them, L TF is the hail binary classification loss function, L T is the hail time prediction loss function, L LOC is the hail position prediction loss function, L D is the hail diameter prediction loss function; is the time weighting coefficient, is the position weighting coefficient, is the diameter weighting coefficient; 、 、 、 are the loss weight coefficients.

[0040] Specifically, a. Input data: All lightning feature variables: b. Neural network model: Transformer c. Output data: Includes 4 prediction targets, namely different task types: ① Whether hail exists: Classification; ② Hail time, probability, diameter, and location regression; d. Loss function design: Predicting all hail variables (occurrence, time, landing location, diameter) simultaneously means that the neural network model needs to perform classification (whether hail occurs) + regression (time, location coordinates, diameter) tasks simultaneously; however, a general weighted loss function may cause interference among these 4 tasks, resulting in non - convergence. And since there is a causal relationship among the variables predicted in this task, an optimized cascaded prediction function module is designed here. Design a dynamic weighted loss function to ensure that the loss of each task is weighted according to the completion of the previous task and the weights are dynamically adjusted according to the prediction quality of each task (such as through prediction error or confidence measure).

[0041] Specific formulas are as follows: For multi - task joint training, the total loss function is as follows: ; ; Step S0411, Hail binary classification loss function ; Wherein, is the predicted probability of hail occurrence, α is the hail category weight, and γ is an adjustable parameter.

[0042] Step S0412, Hail time prediction loss function ; Wherein, is the predicted time of hail occurrence, is the actual time of hail occurrence.

[0043] Step S0413, Hail diameter prediction loss function ; Wherein, is the predicted hail diameter, is the actual hail diameter.

[0044] Step S0414, Hail location prediction loss function ; Wherein, is the predicted hail location coordinates, is the actual hail location coordinates.

[0045] Step S0415, time weighting coefficient ; Position weighting coefficient ; Diameter weighting coefficient ; wherein, , , are adjustment factors, TF is the time frequency prediction value, Tpredict is the time prediction accuracy value, LOCpredict is the position prediction accuracy value.

[0046] Specifically, among them, the four loss functions are respectively: is the hail binary classification loss function. When the number of hail event samples is small, Focal Loss is used to improve the recognition ability of rare events (hail occurrence): ; is the predicted probability, α is the class weight, and γ is an adjustable parameter.

[0047] is the hail time prediction loss function (time regression) to predict the most likely time point of hail occurrence (unit: minute). The robust regression loss function Smooth L1 can be adopted to balance stability and the ability to resist outliers: ; is the predicted time, is the true time; is the hail diameter prediction loss function (regression). After adopting logarithmic scale regression, Smooth L1 loss is also adopted: ; is the predicted hail diameter, is the true hail diameter; is the hail position prediction loss function, adopting direct regression of coordinate points (MSE): ; is the predicted position coordinate, is the true hail coordinate.

[0048] Weighting coefficient , , can be defined in the following form. Here, , , is a regulatory factor used to control the rate of change of the weight of each task; for each task, the better the prediction effect of the previous task, the larger the weighting coefficient , , of the subsequent task will be, thus enhancing its contribution to the total loss.

[0049] ; ; ; Step S05: Predict the hail weather based on the current lightning feature variable data and the dynamic weighted loss function to obtain hail prediction parameters. Among them, the hail prediction parameters include: hail occurrence probability parameter, hail prediction occurrence time parameter, hail prediction diameter parameter, and hail prediction location parameter.

[0050] Step S06: Optimize the hail prediction location parameter according to the wind field data and the current lightning feature variable data; Obtain the wind field data, the current lightning feature variable data, and the hail prediction location parameter; Calculate the target horizontal wind speed based on the wind field data; Obtain a machine learning algorithm, where the machine learning algorithm includes a linear regression algorithm; Calculate the hail movement trajectory according to the machine learning algorithm, the current lightning feature variable data, and the hail prediction location parameter, and adjust the machine learning algorithm parameters according to the calculation results; Retrain the adjusted machine learning algorithm according to the target horizontal wind speed; Optimize the hail movement trajectory according to the trained machine learning algorithm to generate a list of areas affected by hail.

[0051] Specifically, obtain the 500 hPa wind field data: Obtain the wind speed and wind direction data at the 500 hPa height level, ensure that these data cover the corresponding geographical area, and have a high enough time resolution; Match the 500 hPa wind field data with other existing meteorological data (such as radar reflectivity, lightning activity, ground temperature, humidity, etc.) in terms of time and space for unified analysis; Based on the 500 hPa wind field data, calculate key indicators that may affect the hail movement path, such as vertical wind shear (wind speed difference between different height levels), horizontal wind speed, wind direction change, etc.; Observe the change trend of the wind field over time and space, and identify factors that may cause changes in the hail movement direction; Set initial conditions based on the currently observed hail position and size, which include but are not limited to the starting position, velocity of the hail, as well as the wind speed and direction of the surrounding environment; Use appropriate physical models or machine learning algorithms, combined with 500 hPa wind field data and other relevant meteorological data, to simulate the movement trajectory of hail in the atmosphere; Consider using particle tracking technology to simulate the movement paths of single or multiple hail particles; Continuously adjust and optimize the physical model parameters according to the actual observation results to improve the prediction accuracy; Add the key indicators extracted from the above 500 hPa wind field data as new feature variables to the existing feature set; Retrain the prediction model using the updated feature set to ensure that the newly added features do contribute to improving the model performance, and the improvement effect of the model can be evaluated by methods such as cross-validation; Use the trained model to predict the hail path in the future for a period of time and generate a list of areas that may be affected; Combine the predicted hail fall area and intensity information to evaluate the potential risk levels of each area; Timely release accurate hail warning information to the public to guide residents to take necessary protective measures.

[0052] Through the above steps, the 500 hPa wind field data can be effectively incorporated into the hail prediction system, thereby more accurately predicting the hail fall area and reducing disaster losses.

[0053] Step S061, calculate the predicted wind direction and speed through the Kalman filter wind speed budget formula and wind field data; Perform filtering processing on the actual measured wind direction and speed and the predicted wind direction and speed according to the Kalman filtering method to obtain the target horizontal wind speed; Calculate the predicted wind direction and speed through the Kalman filter wind speed budget formula and wind field data, including: Obtain the dynamic model of the wind speed F k , control input matrix B k , control vector , time k , at k the best wind speed estimate at -1 time ; Through the formula: , calculate at time k , based on k the predicted wind direction and speed at -1 time ; Obtain the best estimate error covariance at k -1 time , the process noise covariance matrix Q k ; Through the formula: , calculate at time k , at k -1 the predicted error covariance matrix at the moment .

[0054] At the same time, the predicted wind direction and wind speed are slightly adjusted through the predicted error covariance matrix.

[0055] Specifically, use the Kalman filtering method to filter the actual measured true wind direction and wind speed, the predicted wind direction and wind speed, etc., to obtain an optimal estimate to improve the prediction effect; The Kalman filter wind speed budget formula includes predicting the current wind speed and estimating the error, including wind speed prediction and predicted error covariance: ; ; where is the predicted wind speed at time k (based on k -1 moment data); B k is the control input matrix; F k is the dynamic model of wind speed, Q k is the process noise covariance; And the update step is to use the observation to correct the prediction, which includes three parts: Kalman gain, wind speed correction, and error covariance update: ; ; ; where, is the actual observed wind speed at time k ; H k is the observation model (usually H k = 1, directly observe the wind speed); R k is the observation noise covariance; K k is the Kalman gain, which determines the importance of the new observation value relative to the predicted value.

[0056] Step S07, obtain polarization radar data, where the polarization radar data includes reflectivity factor, differential reflectivity, specific differential phase, and correlation coefficient; Obtain the microphysical characteristic data of hail particles and the mapping relationship between the polarimetric radar data and the microphysical characteristic data of hail particles according to the polarimetric radar data; Optimize the predicted hail diameter parameter according to the microphysical characteristic data of hail particles and the mapping relationship between the polarimetric radar data and the microphysical characteristic data of hail particles; Obtain the microphysical characteristic data of hail particles according to the polarimetric radar data, including: Determine the hail presence value according to the correlation coefficient and the reflectivity factor; Determine the diameter of hail according to the relationship between the differential reflectivity and the hail size and the responsiveness of the specific differential phase; Determine the density and shape of hail according to the change curves of the differential reflectivity and the correlation coefficient; Obtain the mapping relationship between the polarimetric radar data and the microphysical characteristic data of hail particles according to the polarimetric radar data, including: Create a convolutional neural network model; Train the convolutional neural network model with historical polarimetric radar data and the microphysical characteristic data of historical hail particles to obtain the mapping relationship between the polarimetric radar data and the microphysical characteristic data of hail particles; Improve the convolutional neural network model according to the optimization result of the predicted hail diameter parameter.

[0057] Specifically, the polarimetric radar data can provide characteristic quantities of the microphysical characteristics (such as size, density, shape) of hail particles. Combining the existing state predictors, integrating the actual and the predicted, give the optimal budget and improve the forecasting effect; The polarimetric radar obtains detailed information of the target by transmitting horizontally and vertically polarized electromagnetic waves and receiving echo signals, and can provide information about the precipitation particle types (such as raindrops, snowflakes, hail, etc.) and their microphysical properties; Collect observation data using a dual-polarization weather radar in the S-band or C-band, including the reflectivity factor (Z), differential reflectivity (Zdr), specific differential phase (Kdp), correlation coefficient (ρhv), etc.; Remove noise and non-meteorological echoes, and correct the errors caused by factors such as ground clutter and sidelobe effects; Develop a special algorithm based on polarization parameters to distinguish hail from other precipitation types. For example, use a low ρhv value combined with a high Z value to indicate the presence of hail; Utilize the relationship between Zdr and the hail size, and the response characteristics of Kdp in the case of large particles to estimate the diameter of hail; Infer the density and shape of hail particles by comprehensively considering the change patterns of Zdr and ρhv; for example, hailstones with a higher sphericity usually exhibit lower Zdr values; Take the above-mentioned microphysical features extracted from polarimetric radar data as input variables and add them to the existing prediction model; Select a suitable model architecture (such as random forest, support vector machine, convolutional neural network, etc.) and train it with historical data sets, aiming to establish the mapping relationship between polarimetric radar features and hail diameter and intensity level; Adopt the cross-validation method to evaluate the model performance and adjust the hyperparameters to optimize the prediction accuracy; Apply the trained model to the real-time polarimetric radar data stream to dynamically update the hail fall area forecast and intensity level prediction; Here, in addition to polarimetric radar data, information from multiple sources such as satellite remote sensing and ground observation stations can be integrated to further improve the prediction accuracy; According to the feedback of actual hail events, continuously adjust and improve the network model to ensure its adaptability and reliability; Through the above methods, the unique advantages of polarimetric radar products can be effectively utilized to deeply explore the microphysical characteristics of hail particles, thereby significantly improving the prediction ability of hail diameter and intensity level; this not only helps to enhance the effectiveness of the hail warning system, but also provides more accurate services for industries such as agriculture and aviation.

[0058] Step S08, automatically generate hail weather warning information based on the hail prediction result and send it to the meteorological center.

[0059] Here, as Figure 3 shown, this application collects data: obtain hail-related data from various sources, including but not limited to radar reflectivity, lightning activity, and meteorological fields (such as wind speed, temperature, humidity, etc.); Ensure that the time and space resolutions of all data sources are consistent for subsequent processing; Data cleaning: Remove invalid or abnormal data points and fill in missing values; Standardization / Normalization: Convert data of different magnitudes to the same magnitude to improve the efficiency and accuracy of model training; Time series alignment: Ensure that all feature variables are aligned in time for subsequent analysis; Extraction of radar reflectivity characteristic variables: Calculate the statistical characteristics of radar reflectivity factor (Z), such as maximum value, average value, standard deviation, etc.; Extract the reflectivity distribution characteristics of specific altitude layers; Lightning characteristic variable extraction: Count the number of lightning strikes per unit time; Analyze the relationship between lightning positions and radar echoes; Meteorological field characteristic variable extraction: Extract the spatio-temporal variation characteristics of meteorological elements such as wind speed, wind direction, temperature, humidity, etc.; Calculate parameters closely related to hail formation, such as vertical wind shear, CAPE (Convective Available Potential Energy), etc.; Model selection and design: Select neural network models suitable for processing time series data, such as Transformer, LSTM, etc.; Combine the above-extracted characteristic variables into the input vector of the model; Use historical data to train the model and optimize the model parameters to enable accurate prediction of hail characteristic variables; According to different stages of hail formation, design a cascaded prediction model to gradually refine the prediction results; Use the prediction results of the previous stage as the input of the next stage to improve the prediction accuracy; Based on the output of the neural network model, predict characteristic variables such as the probability, intensity, and position of hail occurrence; Set reasonable thresholds to convert continuous prediction results into binary or multi-class hail warning information; Conduct a rationality check on the prediction results to avoid false alarms or missed alarms; Release hail warning information in a timely manner through various channels (such as text messages, websites, APPs, etc.); Collect user feedback on the warning information for continuous optimization and improvement of the model.

[0060] The above steps, from data collection to the final release of warning information, constitute a complete hail prediction system.

[0061] The technical solution of this application reduces the misjudgment caused by predicting hail solely based on radar reflectivity factor, uses lightning data indication factors to improve the hail recognition rate, and can also construct a richer lightning-hail physical model; Since both are the results of the actions of different ice-phase particles, the maximum hail diameter can be inferred based on the lightning-hail relationship; Combine the wind field to calculate the hail trajectory and improve the accuracy of the predicted fall area.

[0062] The core innovation of this technical solution lies in the combination of lightning observation data and neural network model prediction, providing an efficient and accurate hail warning method.

[0063] In addition, the method also includes: Optimize the hail time prediction loss function (time regression), including: Use an asymmetric loss function to improve the sensitivity to the direction of prediction error; Asymmetric loss function Loss = α·error, if error>0 (prediction is too late); β·error, if error≤0 (prediction is too early); Where: α: Control the penalty strength for late prediction; β: Controls the penalty strength for early prediction bias; It is possible to set α > β to emphasize "it is better to report early than late"; Use a learning rate scheduler to dynamically adjust α and β; Set α and β as learnable parameters; Introduce time uncertainty modeling through probability output + Gaussian NLL; Modify the neural network model structure so that it outputs two values: mean μ and standard deviation σ; Define the Gaussian NLL Loss function; ; where: y is the true value, μ is the predicted mean, is the predicted standard deviation; Perform forward propagation, calculate the loss, and backpropagate the gradient for the Gaussian NLL Loss function; Evaluate the uncertainty of the neural network model based on the standard deviation; Introduce time series modeling to enhance the time modeling ability; Process long sequence data through LSTM (Long Short-Term Memory) model and GRU (Gated Recurrent Unit) model to alleviate the vanishing gradient problem in traditional RNNs; Use causal convolutions and dilated convolutions to capture the long-term dependencies of the model without losing the time order; Add metrics such as MAE / RMSE / R² to better evaluate the performance of the hail time prediction loss function.

[0064] It can be understood that here, the accuracy and stability of the neural network model for hail time prediction can be improved by increasing the sensitivity to the direction of time error (such as punishing late reports more severely than early reports); introducing uncertainty modeling (probability output); combining time series modeling to enhance temporal consistency; and introducing an attention mechanism to capture key time points.

[0065] Without departing from the technical solution of this application, several improvements and optimizations can be made to the hail prediction method provided by the embodiments of this application, and these improvements and optimizations should also be regarded as the protection scope of this application.

[0066] This application can also provide and optimize the construction of a hail-lightning correlation model, and use the model to explain the research on physical mechanisms; combined with atmospheric models, it can be extended to other disaster predictions: severe convective rainstorms, thunderstorm gales, etc., expanding the scope of application of the technical solutions of this application; it can be extended to other fields outside meteorology, such as agricultural insurance, aviation safety, etc.

[0067] The beneficial effects brought by the technical solutions provided in the embodiments of this application are: The technology of this application reduces the misjudgment caused by predicting hail solely based on radar reflectivity factors. By using lightning feature data as an indicator factor, the prediction accuracy of hail is improved.

[0068] The technical solutions of this application can effectively improve the prediction accuracy of hail size and falling area, and establish an analysis model of hail formation mechanism based on lightning activities to better understand the growth process of hail in thunderstorms; and an automated hail warning system can be developed, integrating the prediction model into the meteorological forecasting system to improve the practical application value.

[0069] Through the description of the above embodiments, those skilled in the art can clearly understand that the method according to the above embodiments can be implemented by means of software plus a necessary general hardware platform. Of course, it can also be implemented by hardware, but in many cases the former is a better implementation method.

[0070] The embodiments of this application also provide a hail prediction device, as Figure 4 shown. The device includes: an acquisition module, a division module, a conversion module, a training module, a determination module, a prediction module, and an optimization module.

[0071] In this embodiment, the acquisition module is used to acquire a neural network model; acquire historical lightning feature variable data and historical hail feature variable data; acquire current lightning feature variable data; The division module is used to divide the historical lightning feature variable data and historical hail feature variable data into a training set and a validation set of corresponding feature variables; The training module is used to train the neural network model with the training set and validation set of corresponding feature variables, and use the trained neural network model as the target neural network model; The determination module is used to determine a dynamic weighted loss function according to the target neural network model; The prediction module is used to predict hail weather based on the current lightning feature variable data and the dynamic weighted loss function to obtain hail prediction parameters, where the hail prediction parameters include: hail occurrence probability parameter, hail prediction occurrence time parameter, hail prediction diameter parameter, and hail prediction location parameter.

[0072] In this embodiment, the conversion module is used to convert historical lightning feature variable data and historical hail feature variable data into tensor data recognizable by the neural network model.

[0073] In one embodiment, the training module is used to clean the training set and validation set of the corresponding feature variables; Train the neural network model with the training set of the corresponding feature variables after cleaning, and initialize the loss function and optimizer of the neural network model; Verify the neural network model with the validation set of the corresponding feature variables after cleaning, calculate the loss value and determination coefficient of the neural network model; Adjust and optimize the loss value and determination coefficient of the neural network model according to the training results.

[0074] In one embodiment, the determination module is used to dynamically weight the loss function as L total : ; Wherein, L TF is the hail binary classification loss function, L T is the hail time prediction loss function, L LOC is the hail position prediction loss function, L D is the hail diameter prediction loss function; is the time weighting coefficient, is the position weighting coefficient, is the diameter weighting coefficient; 、 、 、 are the loss weight coefficients.

[0075] In one embodiment, the determination module is used for the hail binary classification loss function ; Wherein, is the predicted probability of hail occurrence, α is the hail category weight, and γ is an adjustable parameter; The hail time prediction loss function ; Wherein, is the predicted hail occurrence time, is the true hail occurrence time; The hail diameter prediction loss function ; Wherein, is the predicted hail diameter, is the true hail diameter; Hail position prediction loss function ; Among them, is the predicted hail position coordinate, is the true hail position coordinate.

[0076] In one embodiment, the determination module is used for the time weighting coefficient ; Position weighting coefficient ; Diameter weighting coefficient ; Among them, , , are adjustment factors, TF is the time frequency prediction value, Tpredict is the time prediction accuracy value, LOCpredict is the position prediction accuracy value.

[0077] In one embodiment, the optimization module is used to obtain wind field data, current lightning feature variable data, and hail prediction position parameters; Calculate the target horizontal wind speed according to the wind field data; Obtain a machine learning algorithm, where the machine learning algorithm includes a linear regression algorithm; Calculate the hail movement trajectory according to the machine learning algorithm, the current lightning feature variable data, and the hail prediction position parameters, and adjust the machine learning algorithm parameters according to the calculation results; Retrain the adjusted machine learning algorithm according to the target horizontal wind speed; Optimize the hail movement trajectory according to the trained machine learning algorithm to generate a list of areas affected by hail.

[0078] In one embodiment, the optimization module is used to calculate the predicted wind direction and wind speed through the Kalman filter wind speed budget formula and the wind field data; Perform filtering processing on the actual measured wind direction and wind speed and the predicted wind direction and wind speed according to the Kalman filtering method to obtain the target horizontal wind speed; Calculating the predicted wind direction and wind speed through the Kalman filter wind speed budget formula and the wind field data includes: Obtain the dynamic model of the wind speed F k , control input matrix B k , control vector , time k , at k The best wind speed estimate at -1 time ; Through the formula: , calculate at time k , based on k the wind direction and speed predicted at time -1 ; Obtain the best estimate error covariance at k time -1 , the process noise covariance matrix Q k ; Through the formula: , calculate at time k , the predicted error covariance matrix at k time -1 .

[0079] In one embodiment, an optimization module is configured to obtain polarimetric radar data, where the polarimetric radar data includes a reflectivity factor, a differential reflectivity, a specific differential phase, and a correlation coefficient; Obtain the microphysical characteristic data of hail particles and the mapping relationship between the polarimetric radar data and the microphysical characteristic data of hail particles based on the polarimetric radar data; Optimize the predicted hail diameter parameter according to the microphysical characteristic data of hail particles and the mapping relationship between the polarimetric radar data and the microphysical characteristic data of hail particles; Obtain the microphysical characteristic data of hail particles based on the polarimetric radar data, including: Determine the hail presence value according to the correlation coefficient and the reflectivity factor; Determine the diameter of hail according to the relationship between the differential reflectivity and the hail size and the responsiveness of the specific differential phase; Determine the density and shape of hail according to the change curves of the differential reflectivity and the correlation coefficient; Obtain the mapping relationship between the polarimetric radar data and the microphysical characteristic data of hail particles based on the polarimetric radar data, including: Create a convolutional neural network model; Train the convolutional neural network model with historical polarimetric radar data and the microphysical characteristic data of historical hail particles to obtain the mapping relationship between the polarimetric radar data and the microphysical characteristic data of hail particles; Improve the convolutional neural network model according to the optimization result of the predicted hail diameter parameter.

[0080] Specifically, the present application can improve the accuracy of hail diameter prediction based on a hail prediction model featuring lightning characteristics (polarity + height + frequency), combined with radar data and convective parameters; and a method for calculating the hail trajectory by combining wind field data to accurately predict the hail fall area.

[0081] The beneficial effects brought by the technical solutions provided in the embodiments of the present application are as follows: The technology of the present application reduces the misjudgment caused by predicting hail solely based on radar reflectivity factor, and improves the prediction accuracy of hail through the lightning feature data indication factor.

[0082] The technical solution of the present application can effectively improve the prediction accuracy of hail size and falling area, establish an analysis model for the hail formation mechanism based on lightning activities, and better understand the growth process of hail in thunderstorms; and can develop an automated hail warning system, integrate the prediction model into the meteorological forecasting system, and improve the practical application value.

[0083] For the description of the features in the corresponding embodiments of the hail prediction device, reference can be made to the relevant descriptions in the corresponding embodiments of the hail prediction method, which will not be elaborated here one by one.

[0084] The embodiments of the present application further provide an electronic device, including a memory and a processor. A computer program is stored in the memory, and the processor is configured to run the computer program to execute the steps in the embodiments of the hail prediction method. The method includes: Obtain a neural network model; Obtain historical lightning feature variable data and historical hail feature variable data; Divide the historical lightning feature variable data and historical hail feature variable data into training sets and validation sets of corresponding feature variables; Train the neural network model through the training sets and validation sets of corresponding feature variables, and use the trained neural network model as the target neural network model; Obtain current lightning feature variable data; Determine a dynamic weighted loss function according to the target neural network model; Predict hail weather according to the current lightning feature variable data and the dynamic weighted loss function to obtain hail prediction parameters, where the hail prediction parameters include: hail occurrence probability parameter, hail prediction occurrence time parameter, hail prediction diameter parameter, and hail prediction position parameter.

[0085] As Figure 5 shown, the embodiments of the present application further provide a computer-readable storage medium, in which a computer program is stored. The computer program is configured to execute the steps in the embodiments of the hail prediction method when running. The method includes: Obtain a neural network model; Obtain historical lightning feature variable data and historical hail feature variable data; Divide the historical lightning feature variable data and historical hail feature variable data into training sets and validation sets of corresponding feature variables; Train a neural network model using the training set and validation set of corresponding feature variables, and use the trained neural network model as the target neural network model; Obtain the current lightning feature variable data; Determine a dynamic weighted loss function according to the target neural network model; Predict the hail weather based on the current lightning feature variable data and the dynamic weighted loss function to obtain hail prediction parameters, where the hail prediction parameters include: hail occurrence probability parameter, hail prediction occurrence time parameter, hail prediction diameter parameter, and hail prediction location parameter.

[0086] In an exemplary embodiment, the above computer-readable storage medium may include, but is not limited to: USB flash drive, read-only memory (ROM for short), random access memory (RAM for short), mobile hard disk, magnetic disk, or optical disc, etc., various media that can store computer programs.

[0087] The embodiment of the present application also provides a computer program product. The above computer program product includes a computer program. When the computer program is executed by a processor, it implements the steps in the embodiment of the hail prediction method. The method includes: Obtain a neural network model; Obtain historical lightning feature variable data and historical hail feature variable data; Divide the historical lightning feature variable data and historical hail feature variable data into a training set and a validation set of corresponding feature variables; Train a neural network model using the training set and validation set of corresponding feature variables, and use the trained neural network model as the target neural network model; Obtain the current lightning feature variable data; Determine a dynamic weighted loss function according to the target neural network model; Predict the hail weather based on the current lightning feature variable data and the dynamic weighted loss function to obtain hail prediction parameters, where the hail prediction parameters include: hail occurrence probability parameter, hail prediction occurrence time parameter, hail prediction diameter parameter, and hail prediction location parameter.

[0088] The embodiment of the present application also provides another computer program product, including a non-volatile computer-readable storage medium. The non-volatile computer-readable storage medium stores a computer program. When the computer program is executed by a processor, it implements the steps in the embodiment of the hail prediction method. The method includes: Obtain a neural network model; Obtain historical lightning feature variable data and historical hail feature variable data; Divide the historical lightning feature variable data and historical hail feature variable data into training sets and validation sets for the corresponding feature variables; Train the neural network model with the training sets and validation sets of the corresponding feature variables, and use the trained neural network model as the target neural network model; Obtain the current lightning feature variable data; Determine the dynamic weighted loss function according to the target neural network model; Predict the hail weather based on the current lightning feature variable data and the dynamic weighted loss function to obtain hail prediction parameters, where the hail prediction parameters include: hail occurrence probability parameter, hail predicted occurrence time parameter, hail predicted diameter parameter, and hail predicted position parameter.

[0089] The technology of this application reduces the misjudgment caused by predicting hail solely based on the radar reflectivity factor, and improves the prediction accuracy of hail through the lightning feature data indication factor.

[0090] The technical solution of this application can effectively improve the prediction accuracy of the hail size and the falling area, establish an analysis model of the hail formation mechanism based on lightning activities, and better understand the growth process of hail in thunderstorms; and an automated hail warning system can be developed, integrating the prediction model into the meteorological forecasting system to improve the practical application value.

[0091] Professionals can further realize that, in combination with the units and algorithm steps of each example described in the embodiments disclosed in this article, they can be implemented by electronic hardware, computer software, or a combination of the two. To clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described according to functions in the above description. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but this implementation should not be considered to exceed the scope of this application.

[0092] The above has introduced in detail a hail prediction method, device, medium, and program product provided by this application. Specific examples are used in this article to elaborate on the principle and implementation manner of this application. The description of the above embodiments is only used to help understand the method and its core idea of this application. It should be noted that for those of ordinary skill in the art, without departing from the principle of this application, several improvements and modifications can be made to this application, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A hail prediction method, characterized in that, The method includes: Obtain a neural network model; Obtain historical lightning feature variable data and historical hail feature variable data; Divide the historical lightning feature variable data and historical hail feature variable data into training sets and validation sets for the corresponding feature variables; Train the neural network model with the training sets and validation sets of the corresponding feature variables, and use the trained neural network model as the target neural network model; Obtain current lightning feature variable data; Determine a dynamic weighted loss function according to the target neural network model; Predict hail weather based on the current lightning feature variable data and the dynamic weighted loss function to obtain hail prediction parameters, where the hail prediction parameters include: hail occurrence probability parameter, hail predicted occurrence time parameter, hail predicted diameter parameter, and hail predicted location parameter.

2. The hail prediction method according to claim 1, wherein Before training the neural network model with the training sets and validation sets of the corresponding feature variables, it includes: Convert the historical lightning feature variable data and historical hail feature variable data into tensor data recognizable by the neural network model.

3. The hail prediction method according to claim 1, characterized in that The training of the neural network model with the training sets and validation sets of the corresponding feature variables includes: Clean the training sets and validation sets of the corresponding feature variables; Train the neural network model with the cleaned training sets of the corresponding feature variables, and initialize the loss function and optimizer of the neural network model; Verify the neural network model with the cleaned validation sets of the corresponding feature variables, calculate the loss value and determination coefficient of the neural network model; Adjust and optimize the loss value and determination coefficient of the neural network model according to the training results.

4. The hail prediction method according to claim 1, wherein The determination of the dynamic weighted loss function according to the target neural network model includes: The dynamic weighted loss function is L total : ； Among them, L TF is the hail binary classification loss function, L T is the hail time prediction loss function, L LOC is the hail position prediction loss function, L D is the hail diameter prediction loss function; is the time weighting coefficient, is the position weighting coefficient, is the diameter weighting coefficient; , , , are the loss weight coefficients.

5. The hail prediction method according to claim 4, wherein The hail binary classification loss function ; Among them, is the predicted probability of hail occurrence, α is the hail category weight, and γ is an adjustable parameter.

6. The hail prediction method according to claim 4, wherein The hail time prediction loss function ; Among them, is the predicted time of hail occurrence, is the actual time of hail occurrence.

7. The hail prediction method according to claim 4, wherein The hail diameter prediction loss function ; Among them, is the predicted hail diameter, is the actual hail diameter.

8. The hail prediction method according to claim 4, wherein The hail position prediction loss function ; Among them, is the coordinate of the predicted hail position, is the coordinate of the actual hail position.

9. The hail prediction method according to claim 4, characterized in that The time weighting coefficient ; The weighting coefficient ; The diameter weighting coefficient ; Among them, , , are adjustment factors, TF is the time-frequency prediction value, Tpredict is the time prediction accuracy value, and LOCpredict is the location prediction accuracy value.

10. The hail prediction method according to claim 1, characterized in that After predicting hail weather based on the current lightning feature variable data and the dynamic weighted loss function to obtain hail prediction parameters, it includes: Obtain wind field data, current lightning feature variable data, and hail predicted location parameter; Calculate the target horizontal wind speed according to the wind field data; Obtain a machine learning algorithm, where the machine learning algorithm includes a linear regression algorithm; Calculate the hail movement trajectory according to the machine learning algorithm, current lightning feature variable data, and hail predicted location parameter, and adjust the parameters of the machine learning algorithm according to the calculation results; Retrain the adjusted machine learning algorithm according to the target horizontal wind speed; Optimize the hail movement trajectory according to the trained machine learning algorithm to generate a list of areas affected by hail.

11. The hail prediction method according to claim 10, characterized in that, The calculation of the target horizontal wind speed according to the wind field data includes: Calculate the predicted wind direction and wind speed through the Kalman filter wind speed budget formula and wind field data; Perform filtering processing on the actual measured wind direction and wind speed and the predicted wind direction and wind speed according to the Kalman filtering method to obtain the target horizontal wind speed; The calculation of the predicted wind direction and wind speed through the Kalman filter wind speed budget formula and wind field data includes: Dynamic model F for obtaining wind speed k , control input matrix B k , control vector , at time k, the optimal wind speed estimate at time k-1 ; Through the formula: , calculate the wind direction and speed predicted at time k based on the wind direction and speed predicted at time k - 1 ; Obtain the optimal estimation error covariance at time k-1 , the process noise covariance matrix Q k ; Through the formula: , calculate the predicted error covariance matrix at time k for the time k-1.

12. The hail prediction method according to claim 1, characterized in that After predicting the hail weather based on the current lightning feature variable data and the dynamic weighted loss function to obtain the hail prediction parameters, the method further includes: Obtaining polarization radar data, where the polarization radar data includes reflectivity factor, differential reflectivity, specific differential phase, and correlation coefficient; Obtaining the microphysical characteristic data of hail particles and the mapping relationship between the polarization radar data and the microphysical characteristic data of hail particles based on the polarization radar data; Optimizing the hail prediction diameter parameter according to the microphysical characteristic data of hail particles and the mapping relationship between the polarization radar data and the microphysical characteristic data of hail particles; The obtaining of the microphysical characteristic data of hail particles based on the polarization radar data includes: Determining the hail presence value according to the correlation coefficient and the reflectivity factor; Determining the diameter of hail according to the relationship between the differential reflectivity and the hail size and the responsiveness of the specific differential phase; Determining the density and shape of hail according to the change curves of the differential reflectivity and the correlation coefficient; The obtaining of the mapping relationship between the polarization radar data and the microphysical characteristic data of hail particles based on the polarization radar data includes: Creating a convolutional neural network model; Training the convolutional neural network model with historical polarization radar data and the microphysical characteristic data of historical hail particles to obtain the mapping relationship between the polarization radar data and the microphysical characteristic data of hail particles; Perfecting the convolutional neural network model according to the optimization result of the hail prediction diameter parameter.

13. An electronic device, characterized in that, Including: A memory for storing a computer program; A processor for implementing the steps of the hail prediction method according to any one of claims 1 to 12 when executing the computer program.

14. A computer-readable storage medium, characterized in that, A computer program is stored in the computer-readable storage medium, where the computer program, when executed by a processor, implements the steps of the hail prediction method according to any one of claims 1 to 12.

15. A computer program product, comprising a computer program, characterized in that, The computer program, when executed by a processor, implements the steps of the hail prediction method according to any one of claims 1 to 12.

Images