Method and system for dynamically generating rainstorm disaster alarm based on grid real condition

Abstract

The invention discloses a method and a system for dynamically generating a rainstorm disaster alarm based on a grid actual condition. The method comprises the following steps: collecting site observation rainfall data, grid actual rainfall data, radar data, geographical environment data and rainstorm disaster alarm events in a preset time period; performing deviation analysis and correction on the grid real-time rainfall data by using the site observation rainfall data, and calculating a grid rainfall storm disaster alarm threshold according to the corrected grid real-time rainfall data and the site observation rainfall storm disaster alarm threshold; constructing a training data set, and performing training optimization on the rainstorm disaster alarm model by using the training data set; and carrying out red and orange rainstorm disaster alarm on villages, towns and cities based on the trained and optimized rainstorm disaster alarm model. According to the method, the defects of observation station network layout can be made up, effective monitoring of the rainstorm disaster in a station-free area is realized, and the accuracy, the spatial resolution and the disaster prevention applicability of rainstorm disaster risk early warning are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of meteorological early warning, and more particularly relates to a rainstorm disaster warning dynamic generation method and system based on grid live. BACKGROUND

[0002] Rainstorm has the characteristics of strong burst and great destructive power. Rainstorm disaster warning is one of the important means to prevent rainstorm meteorological disasters. In recent years, meteorological workers have carried out relevant research on the actual rainstorm early warning signal release. The commonly used method at present is to obtain the frequency and rainfall corresponding to different durations and different percentile thresholds by using the percentile threshold method, and to determine or revise the rainstorm disaster warning threshold through extreme precipitation frequency statistical analysis, by using the precipitation observation data of ground meteorological stations (such as the rainfall 1, 3, 6, 12 hours before the disaster occurs) in the past 30 years when the disaster occurred.

[0003] The traditional static index based on a single site and a single element (such as process rainfall) has been difficult to adapt to the business needs of precise early warning. Due to the layout of the station network of site observation, there are blind spots in the monitoring of rainstorms in areas not covered, which restricts the accuracy and timeliness of rainstorm disaster warning. SUMMARY

[0004] In view of the above defects or improvement needs of the prior art, the present application provides a rainstorm disaster warning dynamic generation method and system based on grid live, which can significantly improve the accuracy, spatial resolution and disaster prevention applicability of rainstorm disaster risk early warning.

[0005] To achieve the above-mentioned purpose, according to one aspect of the present application, a rainstorm disaster warning dynamic generation method based on grid live is provided, comprising the steps of: Collecting site observation precipitation data, grid live precipitation data, radar data, geographical environment data and rainstorm disaster warning events in a preset time period; Correcting the grid live precipitation data by using the site observation precipitation data, and calculating the grid precipitation rainstorm disaster warning threshold according to the corrected grid live precipitation data and the site observation precipitation rainstorm disaster warning threshold; Constructing a training data set, each training sample including the grid live precipitation data, radar data, geographical environment data and rainstorm disaster warning event label of each training sample, the rainstorm disaster warning event label of each training sample being determined according to the grid precipitation rainstorm disaster warning threshold, rainstorm disaster warning event and radar data, and training and optimizing the rainstorm disaster warning model based on convolutional neural network by using the training data set; Inputting the grid live precipitation data, radar data and geographical environment data of the sample to be predicted into the rainstorm disaster warning model trained and optimized, and obtaining a rainstorm disaster warning event early warning value.

[0006] Further, the step of correcting the grid live precipitation data using the station observed precipitation data comprises steps of: sorting the station observed precipitation data and the grid live precipitation data, dividing the precipitation into multiple intervals, and calculating the cumulative probability distribution functions of the station observed precipitation and the grid live precipitation respectively; determining multiple quantile interval, and calculating the transfer function of the station observed precipitation and the grid live precipitation in each quantile interval, so that the cumulative probability distribution function of the grid live precipitation after being corrected by the transfer function is the same as the cumulative probability distribution function of the station observed precipitation.

[0007] Further, the step of calculating the grid precipitation rainstorm disaster warning threshold according to the corrected grid live precipitation data and the station observed precipitation rainstorm disaster warning threshold comprises steps of: splitting the station precipitation data as the dependent variable into multiple quantile points, estimating by using the weighted least absolute deviation method, establishing the regression model of the station observed precipitation and the grid live precipitation, and obtaining the grid precipitation rainstorm disaster warning threshold corresponding to the station observed precipitation reaching the rainstorm disaster warning threshold.

[0008] Further, the grid precipitation rainstorm disaster warning threshold comprises rainstorm disaster warning thresholds of different levels in different regions of cities and towns, and the towns include mountainous areas and plain areas.

[0009] Further, the radar data of each training sample comprises echo intensity, echo top height and vertical integrated liquid water content.

[0010] Further, the geographical environment data of each training sample comprises a category variable representing the city or town area to which the grid belongs, and the category variable of the town area to which the grid belongs is calculated according to the elevation and slope of the grid.

[0011] Further, the rainstorm disaster warning event label of each training sample is determined according to the grid precipitation rainstorm disaster warning threshold, the rainstorm disaster warning event and the radar data, comprising steps of: pre-determining the number of categories of the rainstorm disaster warning event label, and different categories correspond to different levels of rainstorm; pre-determining the conditions required to be met by the grid precipitation rainstorm disaster warning threshold, the rainstorm disaster warning event and the radar data corresponding to the rainstorm disaster warning event label of each category, and determining the rainstorm disaster warning event label of each training sample according to the conditions.

[0012] Further, the loss function of the rainstorm disaster warning model is calculated by using any one of the following methods: Method 1, ; wherein,L represents a loss function, C represents the class of the rainstorm disaster warning event label, represents the weight of the i-th class sample, represents the true rainstorm disaster warning event label of the i-th class sample, represents the class probability of the predicted rainstorm disaster warning event label, and the weight size is inversely proportional to the class sample size; Method 2, wherein, represents the Focal Loss loss value, and N represents the total number of samples, C represents the class of the rainstorm disaster warning event label; represents a class weight balancing factor, which is inversely proportional to the frequency of the class c; represents the estimated probability that the model predicts that the i-th sample belongs to the class c, and the value range is between [0, 1]; γ represents a focusing parameter, which is an adjustable hyperparameter greater than or equal to 0; represents a sign function, for the i-th sample, if its true class is c, then =1 otherwise =0.

[0013] Further, the training and optimization of the rainstorm disaster warning model based on the convolutional neural network using the training data set comprises the steps of: initially training the rainstorm disaster warning model using the training data set, generating a rainstorm disaster warning event early warning value of the rainstorm disaster risk area using the rainstorm disaster warning model after initial training, and calculating the difference between the rainstorm disaster warning event early warning value of the rainstorm disaster risk area and the true value of the rainstorm disaster warning event of the rainstorm disaster risk area; adjusting the calculation method of the loss function, and / or the weight factor of the loss function, and / or oversampling the samples of the rainstorm disaster risk area according to the difference between the rainstorm disaster warning event early warning value and the true value of the rainstorm disaster warning event.

[0014] According to another aspect of the present application, a rainstorm disaster warning dynamic generation system based on grid live is provided, comprising: a data acquisition module for acquiring site observation precipitation data, grid live precipitation data, radar data, geographic environment data and rainstorm disaster warning events in a preset time period; a grid precipitation rainstorm disaster warning threshold calculation module for correcting the grid live precipitation data using the site observation precipitation data, and calculating the grid precipitation rainstorm disaster warning threshold according to the corrected grid live precipitation data and the site observation precipitation rainstorm disaster warning threshold; ​The training module is configured to build a training data set, each training sample including grid live precipitation data, radar data, geographical environment data and a storm disaster warning event label of each training sample, the storm disaster warning event label of each training sample being determined according to a grid precipitation storm disaster warning threshold, a storm disaster warning event and radar data, and the storm disaster warning model based on the convolutional neural network is trained and optimized by using the training data set. The prediction module is configured to input the grid live precipitation data, radar data and geographical environment data of a to-be-predicted sample into the storm disaster warning model after training and optimization to obtain a storm disaster warning event early warning value.

[0015] Overall, compared with the prior art, the above technical solutions conceived by the present application fuse site observation, radar remote sensing, geographical environment information, high spatiotemporal resolution grid products and other multi-source meteorological observation data, and realize scientific, fine and dynamic dynamic storm disaster warning generation through modeling and machine learning objective quantitative methods, break through the limitations of site observation, effectively identify storm processes in areas without site observation, and significantly improve the accuracy, spatial resolution and disaster prevention applicability of storm disaster risk warning. Specifically, the present application has the following advantages: (1) A quantile mapping / regression-depth learning fusion modeling framework is constructed. The present application introduces quantile regression and quantile mapping methods to analyze and correct the bias of high spatiotemporal resolution grid live precipitation products, providing a high-quality data basis for model construction. The present application breaks through the limitations of traditional statistical or threshold methods, fuses multi-source data to construct a deep learning model, which can automatically and efficiently learn the complex nonlinear relationship between multi-dimensional features and storm disasters, significantly enhancing the expression ability of the model.

[0016] (2) A seamless spatial coverage grid-based storm disaster warning index is proposed. Traditional storm disaster warning indexes are highly dependent on ground site observation, resulting in monitoring blind spots and uncertainty in storm disaster warning in areas without site observation. The present application expands site observation to grid-based storm disaster warning covering the entire domain based on high-precision grid live products, and differentiates modeling according to different terrain characteristics such as mountains, plains and cities, improving the scientificity and practicality of the index.

[0017] (3) A targeted testing and model optimization mechanism for disaster risk is established. Traditional model testing focuses on overall accuracy, and the present application constructs a testing and optimization closed loop with disaster risk points as the core. On the basis of conventional independent / non-independent sample testing, further applicability testing is carried out in disaster risk point areas, mountains, plains, cities and other typical underlying surfaces, and the model is optimized according to the results to ensure the highest reliability in key areas and realize the leap from "high model accuracy" to "strong disaster prevention effectiveness". BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 is a working principle schematic diagram of a grid live-based rainstorm disaster warning dynamic generation method of an embodiment of the present application. DETAILED DESCRIPTION

[0019] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0020] In the description of the embodiments of the present application, the term "a plurality of" means two or more, unless otherwise specified.

[0021] In the embodiments of the present application, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, device, product or equipment comprising a series of steps or modules does not have to be limited to the clearly listed steps or modules, but can include other steps or modules that are not clearly listed or inherent to these processes, methods, products or equipment.

[0022] The naming or numbering of the steps appearing in the embodiments of the present application does not mean that the steps in the method process must be performed in the time / logical order indicated by the naming or numbering. The named or numbered process steps can change the execution order according to the technical purpose to be achieved, as long as the same or similar technical effects can be achieved.

[0023] Reference to "embodiments" herein means that the specific features, structures or properties described in conjunction with the embodiments can be included in at least one embodiment of the present application. The phrase appears at various places in the specification does not necessarily refer to the same embodiment, nor is it independent or alternative to other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0024] The embodiments of the present application propose a grid live-based rainstorm disaster warning dynamic generation method, the working principle of which is as shown in Figure 1As shown, using the precipitation observation data of the research area minute-level ground station in the past three years, the 1km grid real-time precipitation product, the radar data, the meteorological disaster risk point zoning and the geological disaster information, firstly, the quantile regression and quantile mapping method are used to quantitatively analyze the systematic deviation of the grid real-time precipitation and the station observation precipitation, combined with the rainstorm disaster warning index based on the station observation, the grid real-time rainstorm disaster warning index for different terrain regions and different rainstorm grades is proposed; secondly, based on the geographical environment and the radar data, the characteristic factors related to the rainstorm are screened out, and the rainstorm disaster warning model based on the grid real-time precipitation is established by using the deep learning method; finally, the back calculation based on the grid real-time rainstorm disaster warning is completed by using the observation data in the past three years, the accuracy of the rainstorm disaster warning based on the grid real-time is evaluated by using the independent sample and non-independent sample test method, and the applicability test is carried out in the disaster risk point region, the mountainous area, the plain and the city, and the rainstorm disaster warning model is further optimized according to the test result.

[0025] The grid real-time rainstorm disaster warning dynamic generation method based on the embodiment of the application comprises steps (1)-(4).

[0026] (1) Collecting the station observation precipitation data, the grid real-time precipitation data, the radar data, the geographical environment data and the rainstorm disaster warning event in a preset time period.

[0027] In one embodiment, the preset time period can be set as the past three years, the related data in the research region in the past three years is collected and processed, and the rainstorm disaster event is screened out, the observation data quality control of the meteorological station and the hydrological station is carried out, and the nearest neighbor method is used to perform the space-time matching processing on the station observation and the grid real-time product.

[0028] Table 1 Data list ; (2) Correcting the grid real-time precipitation data by using the station observation precipitation data, and calculating the grid precipitation rainstorm disaster warning threshold according to the station observation rainstorm disaster warning threshold and the corrected grid real-time precipitation data.

[0029] The quantile mapping method can be used to correct the grid real-time precipitation data by using the station observation precipitation data, comprising steps (i)-(iii).

[0030] i) Calculate the cumulative distribution function The grid precipitation and the observation precipitation are sorted in ascending order, the precipitation is divided into multiple intervals, and the cumulative probability distribution functions of the station observation precipitation and the grid real-time precipitation are calculated by using the precipitation sequence, and the following calculation formula can be used: ; ; where x is the precipitation amount; pdf(x) is the frequency distribution of precipitation amount in the interval [a, b]; cdf(x) is the cumulative probability distribution of precipitation amount less than a certain precipitation threshold x p when x p is the maximum precipitation amount, the cumulative probability is 1.

[0031] ii) Constructing the transfer function To better capture extreme precipitation, the conventional values and extreme values are segmented by a certain precipitation probability threshold, and the error correction is performed on each segment of data. Through parameter adjustment, combined with the rainstorm disaster warning standard, the optimal threshold quantile of the cumulative precipitation probability is determined. Different transfer functions are established in different quantile segment intervals, so that after the grid real-time precipitation is corrected by the transfer function, the cumulative distribution function of the observed precipitation value is the same as the cumulative distribution function of the grid real-time precipitation: ; ; where and respectively represent the precipitation amount of the i-th quantile and the j-th quantile; TF(x) is the transfer function; and respectively represent the cumulative probability distribution of the observed precipitation and the grid precipitation.

[0032] iii) Correcting the grid real-time precipitation data using the transfer function , where the grid real-time precipitation value, is the equivalent station observed precipitation value, that is, the corrected grid real-time precipitation value.

[0033] Through the above quantile mapping method, the inherent systematic bias between the station observation and the grid real-time, which is irrelevant to the weather process, is eliminated, so that any grid precipitation value is input, and a corrected grid precipitation data distribution consistent with the station observed precipitation climate distribution is output , which can be explained as "the station precipitation equivalent to the grid precipitation in climate statistics". For example, if the grid real-time 50mm precipitation is the 95% quantile in its historical sequence, and the 95% quantile of the station history is 60mm, then after correction by the quantile mapping method, this grid real-time precipitation is corrected to 60mm.

[0034] Then, a quantile regression model is established to calculate the grid rainstorm disaster warning threshold according to the corrected grid live rainfall data and the site observed rainfall rainstorm disaster warning threshold. Quantile regression can be used to study the regression relationship at multiple quantile points. On the basis of correcting systematic bias by quantile mapping, the site rainfall data is split into multiple quantile points as the dependent variable, and the weighted least absolute deviation method is used to estimate and establish the regression model of site rainfall and grid rainfall to obtain the corresponding value of grid rainfall when the site observation reaches the rainstorm disaster warning standard. Specifically, steps (i)~(iii) are included: i) Construct a linear quantile regression: ; where a set of high quantiles τ = [0.75, 0.80, 0.85, 0.90, 0.95, 0.99] is selected, representing from heavy rainfall to extreme rainfall; is the τ quantile of the site rainfall y given the grid rainfall ; β0(τ) and β1(τ) are the parameters to be estimated depending on the quantile τ.

[0035] ii) Construct a quantile regression model based on the weighted least absolute deviation method: For a given quantile τ, minimize the following objective function: to obtain the estimates of parameters β0(τ) and β1(τ) and , where ρ_τ(u) is the test function, defined as: ; is an indicator function that takes the value 1 when the residual is negative and 0 otherwise.

[0036] iii) Inference of the grid rainfall rainstorm disaster warning threshold: If the warning standard is the site observed rainfall , for each fitted quantile regression model, solve the equation to obtain a series of grid rainfall rainstorm disaster warning thresholds ; When the grid rainfall reaches , the site observed rainfall has a probability of no more than .

[0037] The above quantile mapping and quantile regression method quantitatively reveals the differences between site observed rainfall and grid rainfall from multiple dimensions, and calculates the grid rainfall rainstorm disaster warning threshold according to the site rainfall rainstorm disaster warning threshold.

[0038] Furthermore, the grid-based rainstorm disaster warning thresholds include different levels and regions for rainstorm disaster warnings in cities and towns, with towns including mountainous and plain areas. In one embodiment, as shown in Table 2, rainstorm disaster warning thresholds based on grid conditions can be set by combining station-based rainstorm disaster warning issuance standards and comprehensively considering the vulnerability of the underlying surface and disaster-bearing bodies, through regional division and precipitation intensity classification, as indicators for various types of rainstorm disaster warnings.

[0039] Table 2. Rainstorm Disaster Warning Indicators Based on Grid Data ; (3) Construct a training dataset. Each training sample includes grid-based real-time precipitation data, radar data, geographic environment data, and rainstorm disaster warning event labels. The rainstorm disaster warning event labels for each training sample are determined based on the grid precipitation rainstorm disaster warning threshold, rainstorm disaster warning events, and radar data. The training dataset is used to train and optimize the rainstorm disaster warning model based on convolutional neural networks.

[0040] In one embodiment, a convolutional neural network (CNN) is used to establish a rainstorm disaster warning model for townships and cities, using three types of events—orange and red rainstorm disaster warning events and no rainstorm disaster warning events—as the output targets of the disaster warning model.

[0041] i) Input data Each training sample includes grid-based real-time precipitation data, radar data, geographic environmental data, and rainstorm disaster warning event labels.

[0042] Each training sample's grid-based real-time precipitation data includes real-time precipitation over a 1km grid, comprising precipitation data for the past 12 time intervals over 10 minutes.

[0043] The radar data for each training sample includes echo intensity, echo top height, and vertical cumulative liquid water content. Based on the relationship between radar and heavy rainfall, echo intensity, echo top height, and vertical cumulative liquid water content are selected as radar characteristic factors.

[0044] Precipitation is closely related to topography, especially the geographical location of high-altitude areas, which has a significant impact on precipitation. Considering the influence of complex topography on precipitation, altitude, slope, and aspect are selected as geographical environment modeling factors.

[0045] The geographic environment data for each training sample includes categorical variables representing the city and township region to which the grid belongs. The categorical variables for the township region to which the grid belongs are calculated based on the grid's elevation and slope. Using ArcGIS software, the 90-meter DEM geographic environment data is resampled to 1 km using bilinear interpolation, and the elevation of each grid is extracted. ) and slope ( Centered on the current grid point, the local computation range (e.g., a 3x3 grid) is defined using `window_size`, and the terrain roughness of a 1km grid is calculated. ): ; Among them: local standard deviation of elevation It reflects the degree of local topographic relief; mean slope This reflects the average steepness of the area.

[0046] If the terrain roughness exceeds the threshold, it is a mountainous area; otherwise, it is a plain.

[0047] ii) Marking of rainstorm disaster warning event tags The rainstorm disaster warning event labels are generated based on historical data and using a method of "alarm threshold + comprehensive judgment", which constitutes the "truth value" required for supervised learning.

[0048] Specifically, the rainstorm pre-disaster warning event label for each training sample is determined based on gridded precipitation rainstorm pre-disaster warning thresholds, rainstorm pre-disaster warning events, and radar data, including the following steps: Predetermine the number of categories for rainstorm disaster warning event tags, with different categories corresponding to different levels of rainstorms; The threshold for grid precipitation rainstorm warnings, the conditions that rainstorm warning events and radar data must meet are predetermined for each category of rainstorm warning event label, and the rainstorm warning event label for each training sample is determined based on these conditions.

[0049] In one embodiment, based on the real-time rainstorm warning index (see Table 2), factors such as radar and weather conditions are introduced, and orange rainstorm, red rainstorm warning events and no rainstorm warning events are marked using echo intensity, echo top height, and vertical cumulative liquid water content.

[0050] Orange rainstorm emergency warning event: When the real-time precipitation in the 1, 3, 6, and 12-hour grid reaches the rainfall standard corresponding to the "orange" level in Table 2, and there are strong echoes, high / low echo top heights, and high / medium vertical cumulative liquid water content.

[0051] Red Rainstorm Disaster Alert Event: When the real-time rainfall in the 1, 3, 6, and 12-hour grid reaches the rainfall standard corresponding to the "Red" level in Table 2, and there are strong echoes, high echo top heights, and high vertical cumulative liquid water content.

[0052] No rainstorm disaster warning event: When the actual rainfall in the grid does not reach any "orange" or "red" warning standard.

[0053] iii) Training of rainstorm disaster warning models By comparing and analyzing the algorithm's accuracy, running efficiency, and resource consumption, the model parameters were adjusted to construct a rainstorm disaster warning model based on deep learning, enabling intelligent generation of rainstorm disaster warnings based on real-time precipitation data from a 1km grid every 10 minutes.

[0054] Since the rainstorm disaster warning event is a typical class imbalance problem (the vast majority of samples are "no rainstorm disaster warning", while the very few are "red rainstorm disaster warning"), directly using the standard cross-entropy loss will cause the model to be biased towards the majority class. Therefore, the loss function of the rainstorm disaster warning model in this invention is calculated using any of the following methods: Method 1, ; in, L Represents the loss function. C This indicates the category of the rainstorm disaster warning event label. This represents the weight of the i-th class of samples. This represents the label of the real rainstorm disaster warning event for the i-th type of sample. The label represents the category probability of the predicted rainstorm disaster warning event. The weight is inversely proportional to the sample size of the category, which forces the model to pay more "attention" to rare rainstorm events during training.

[0055] Method 2, ; in, This represents the Focal Loss value, where N represents the total number of samples. C Indicates the category of the rainstorm disaster warning event label; This represents the category weight balancing factor, a manually set hyperparameter that is typically inversely proportional to the frequency of category c (e.g., the frequency of a red rainstorm disaster warning). Maximum, no rainstorm disaster warning. Minimum), used to resolve imbalances between categories; γ represents the estimated probability that sample i belongs to class c, and its value ranges from [0, 1]. γ represents the focusing parameter, which is an adjustable hyperparameter greater than or equal to 0. It is used to address the imbalance between easy and difficult samples. Let represent a sign function, that is, for sample i, if its true class is c, then =1 otherwise =0.

[0056] Focal Loss, building upon weighted cross-entropy, introduces a dynamic scaling factor. By reducing the weight of "easily classifiable" samples (i.e., samples for which the model is already confident in issuing rainstorm warnings), it allows the model to focus more on learning "difficult-to-classifiable" samples near a threshold during training. This perfectly aligns with the need for precise capture of "critical states" in early warning operations.

[0057] iv) Testing, evaluation, and model optimization of rainstorm disaster warnings Using observational data from the past three years, we completed a back-calculation of rainstorm disaster warnings based on grid-based real-time data. The accuracy of the rainstorm disaster warnings based on grid-based real-time data was verified from the dimensions of non-independent samples, independent samples, rainstorm disaster events, rainstorm disaster risk areas, and different terrain areas. Based on the verification results, we further optimized the rainstorm disaster warning model.

[0058] There are two training optimization phases.

[0059] The first step is initial model training and back-calculation, which involves using the training dataset to initially train the rainstorm disaster warning model, using the initially trained rainstorm disaster warning model to generate rainstorm disaster warning event warning values ​​for rainstorm disaster risk areas, and calculating the difference between the rainstorm disaster warning event warning values ​​and the actual rainstorm disaster warning event values ​​for rainstorm disaster risk areas.

[0060] Using all or most of the observation data from the past three years, the model is trained for the first time to obtain an initial model. This initial model is then used to perform a retrospective test (i.e., "backtesting") on the data from those three years, generating corresponding rainstorm disaster warnings. The rainstorm disaster warnings are then compared with the actual situation (i.e., "true values," including station-observed precipitation, grid-based actual precipitation, and disaster information). A comprehensive diagnosis of the model's performance and deficiencies is then conducted. For example: Does the model generally miss warnings or issue false alarms? How does it perform in complex terrain such as mountainous areas? How well does it detect red and orange rainstorm disaster warning events? The second step is model tuning based on the test results. This involves adjusting the calculation method of the loss function, and / or the weighting factor of the loss function, and / or oversampling the samples in the rainstorm disaster risk area based on the difference between the warning value and the actual value of the rainstorm disaster warning event in the rainstorm disaster risk area.

[0061] After determining the optimal model structure, loss function, and training strategy, the optimized new model is trained in a final, complete, one-time operation using all data from the past three years (including previous training and validation datasets). For example, the loss function is adjusted: if serious underreporting of red rainstorm disaster alerts is found, its weight can be increased in the loss function; the training strategy is adjusted by oversampling samples from rainstorm disaster risk areas, etc., to determine the final business model.

[0062] ① Inspection and evaluation Non-independent sample test: Using precipitation observed at meteorological stations as the non-independent sample test standard, the accuracy of rainstorm disaster warnings is evaluated.

[0063] Independent sample validation: Precipitation data from meteorological and hydrological stations not integrated into the grid were used as independent sample validation criteria to assess the accuracy of rainstorm disaster warnings.

[0064] Testing of rainstorm disaster events: Based on disaster information, select typical rainstorm disaster risk areas and test the applicability of rainstorm pre-disaster warning indicators under disaster events.

[0065] Test of rainstorm disaster risk points: Based on the zoning of rainstorm disaster risk points, test the applicability of rainstorm disaster warning indicators to areas with different levels of rainstorm disaster risk.

[0066] Testing in different terrain regions: Applicability testing was conducted in different terrain regions such as mountainous areas, plains, and cities.

[0067] ② Model optimization Based on the verification results of the multidimensional evaluation, the rainstorm disaster warning model was further optimized by adjusting the alarm index thresholds, feature factors and model parameters of the grid real-time situation.

[0068] (4) Input the grid-based real-time precipitation data, radar data and geographic environment data of the sample to be predicted into the trained and optimized rainstorm disaster warning model to obtain the rainstorm disaster warning event warning value.

[0069] Through experimental verification, this invention can realize the dynamic and automatic generation of real-time rainstorm disaster warnings down to the township level, with an accuracy rate of over 90% and a generation time of less than 5 minutes, providing support for rainstorm disaster warnings and short-term early warnings of geological disaster hazard points in townships (streets).

[0070] An embodiment of the present invention provides a dynamic generation system for rainstorm disaster warnings based on grid-based real-time data, comprising: The data acquisition module is used to collect station-observed precipitation data, grid-based real-time precipitation data, radar data, geographic environment data, and rainstorm disaster warning events within a preset time period; The grid precipitation storm warning threshold calculation module is used to correct the grid real-time precipitation data using station-observed precipitation data, and calculate the grid precipitation storm warning threshold based on the corrected grid real-time precipitation data and the station-observed precipitation storm warning threshold. The training module is used to construct a training dataset. Each training sample includes grid-based real-time precipitation data, radar data, geographic environment data, and rainstorm disaster event labels. The rainstorm disaster event labels for each training sample are determined based on the grid-based precipitation rainstorm disaster warning threshold, rainstorm disaster warning events, and radar data. The training dataset is used to train and optimize the rainstorm disaster warning model based on the convolutional neural network. The prediction module is used to input the grid-based real-time precipitation data, radar data, and geographic environment data of the sample to be predicted into the trained and optimized rainstorm disaster warning model to obtain the rainstorm disaster warning value.

[0071] The dynamic generation system for rainstorm disaster warnings in this embodiment of the invention has the same working principle and technical effect as the above-described dynamic generation method for rainstorm disaster warnings, and will not be described again here.

[0072] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for dynamically generating rainstorm disaster warnings based on grid-based real-time data, characterized in that, Including the following steps: Collect station-observed precipitation data, grid-based real-time precipitation data, radar data, geographic environment data, and rainstorm disaster warning events within a preset time period; The deviation of the grid-based real-time precipitation data is corrected using the station-observed precipitation data. The grid-based real-time precipitation data and the station-observed precipitation rainstorm warning threshold are then used to calculate the grid-based rainstorm rainstorm warning threshold. A training dataset is constructed, in which each training sample includes grid-based real-time precipitation data, radar data, geographic environment data, and rainstorm disaster warning event labels. The rainstorm disaster warning event labels for each training sample are determined based on the grid-based precipitation rainstorm disaster warning threshold, rainstorm disaster warning events, and radar data. The training dataset is used to train and optimize the rainstorm disaster warning model based on the convolutional neural network. The real-time precipitation data, radar data, and geographic environment data of the sample to be predicted are input into the trained and optimized rainstorm disaster warning model to obtain the rainstorm disaster warning event warning value.

2. The method for dynamically generating rainstorm disaster warnings based on grid-based real-time data as described in claim 1, characterized in that, The correction of grid-based real-time precipitation data using station-observed precipitation data includes the following steps: The station-observed precipitation data and grid-based real-time precipitation data are sorted and the precipitation is divided into multiple intervals. The cumulative probability distribution functions of the station-observed precipitation and grid-based real-time precipitation are calculated respectively. Multiple quantile intervals are determined, and the transfer function between station-observed precipitation and grid-based actual precipitation is calculated for each quantile interval. This ensures that after correcting the grid-based actual precipitation using the transfer function, the cumulative probability distribution function of the grid-based actual precipitation is the same as that of the station-observed precipitation.

3. The method for dynamically generating rainstorm disaster warnings based on grid-based real-time data as described in claim 1, characterized in that, The calculation of the grid-based precipitation storm warning threshold based on the corrected grid-based real-time precipitation data and the station-observed precipitation storm warning threshold includes the following steps: The station precipitation data is split into multiple quantile points as the dependent variable, and the weighted least absolute deviation method is used for estimation. A regression model of station observed precipitation and grid actual precipitation is established to obtain the grid precipitation rainstorm disaster warning threshold when the station observed precipitation reaches the rainstorm disaster warning threshold.

4. The method for dynamically generating rainstorm disaster warnings based on grid-based real-time data as described in claim 1, characterized in that, The grid-based rainstorm disaster warning thresholds include different levels of rainstorm disaster warning thresholds for different areas of cities and towns, with towns including mountainous and plain areas.

5. The method for dynamically generating rainstorm disaster warnings based on grid-based real-time data as described in claim 1, characterized in that, The radar data for each training sample includes echo intensity, echo top height, and vertical cumulative liquid water content.

6. The method for dynamically generating rainstorm disaster warnings based on grid-based real-time data as described in claim 4, characterized in that, The geographic environment data for each training sample includes category variables representing the city or township region to which the grid belongs. The category variables for the township region to which the grid belongs are calculated based on the grid's elevation and slope.

7. The method for dynamically generating rainstorm disaster warnings based on grid-based real-time data as described in claim 4, characterized in that, The rainstorm emergency warning event label for each training sample is determined based on gridded precipitation rainstorm emergency warning thresholds, rainstorm emergency warning events, and radar data, including the following steps: Predetermine the number of categories for rainstorm disaster warning event tags, with different categories corresponding to different levels of rainstorms; The threshold for grid precipitation rainstorm warnings, the conditions that rainstorm warning events and radar data must meet are predetermined for each category of rainstorm warning event label, and the rainstorm warning event label for each training sample is determined based on these conditions.

8. The method for dynamically generating rainstorm disaster warnings based on grid-based real-time data as described in claim 1, characterized in that, The loss function of the rainstorm disaster warning model is calculated using any of the following methods: Method 1, ; in, L Represents the loss function. C This indicates the category of the rainstorm disaster warning event label. This represents the weight of the i-th class of samples. This represents the label of the real rainstorm disaster warning event for the i-th type of sample. The label represents the category probability of the predicted rainstorm disaster warning event, and the weight is inversely proportional to the sample size of the category; Method 2, ; in, This represents the Focal Loss value, where N represents the total number of samples. C Indicates the category of the rainstorm disaster warning event label; This represents the category weight balancing factor, which is inversely proportional to the frequency of category c. γ represents the estimated probability that sample i belongs to class c, and its value ranges from [0, 1]; γ represents the focusing parameter, which is an adjustable hyperparameter that is greater than or equal to 0. Let represent a sign function, where for sample i, if its true class is c, then =1 otherwise =0.

9. The method for dynamically generating rainstorm disaster warnings based on grid-based real-time data as described in claim 1, characterized in that, The process of training and optimizing the rainstorm disaster warning model based on the convolutional neural network using the training dataset includes the following steps: The rainstorm disaster warning model is initially trained using the training dataset. The rainstorm disaster warning model after initial training is used to generate rainstorm disaster warning event warning values ​​for rainstorm disaster risk areas. The difference between the rainstorm disaster warning event warning values ​​and the actual rainstorm disaster warning event values ​​for rainstorm disaster risk areas is calculated. Adjust the calculation method of the loss function and / or the weighting factor of the loss function based on the difference between the warning value and the actual value of the rainstorm disaster warning event, and / or oversample the samples in the rainstorm disaster risk area.

10. A dynamic generation system for rainstorm disaster warnings based on grid-based real-time data, characterized in that, include: The data acquisition module is used to collect station-observed precipitation data, grid-based real-time precipitation data, radar data, geographic environment data, and rainstorm disaster warning events within a preset time period; The grid precipitation storm warning threshold calculation module is used to correct the deviation of the grid real precipitation data using the station observation precipitation data, and to calculate the grid precipitation storm warning threshold based on the corrected grid real precipitation data and the station observation precipitation storm warning threshold. The training module is used to construct a training dataset. Each training sample includes grid-based real-time precipitation data, radar data, geographic environment data, and rainstorm disaster warning event labels. The rainstorm disaster warning event labels for each training sample are determined based on the grid-based precipitation rainstorm disaster warning threshold, rainstorm disaster warning events, and radar data. The training dataset is used to train and optimize the rainstorm disaster warning model based on the convolutional neural network. The prediction module is used to input the grid-based real-time precipitation data, radar data, and geographic environment data of the sample to be predicted into the trained and optimized rainstorm disaster warning model to obtain the rainstorm disaster warning event warning value.

Images