A weather disaster early warning calling method and system based on a weather early warning machine

Abstract

The application discloses a meteorological disaster early warning method and system based on a meteorological early warning machine, and the system comprises a data acquisition module, which is used for constructing a database containing real-time meteorological data, terrain data, historical disaster data and operation data of the early warning machine; a multi-disaster type terrain adaptive analysis module, which is used for analyzing the comprehensive risk of waterlogging and gale in plains by using a synchronous superposition model through a double-branch fusion neural network, and analyzing the chain risk of landslides and lightning-gale in mountainous areas by using a time sequence chain model; an early warning response strategy generation module, which is used for generating differentiated strategies according to regional risks and terrain; and a terminal interaction module, which is used for pushing information and receiving feedback. Through multi-source meteorological data integration and analysis, the system dynamically generates protection guidelines for different geographical characteristics, such as high-risk areas in plains and high-risk areas in mountainous areas, and improves the timeliness and coverage of early warning information.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of early warning devices, and in particular to a meteorological disaster early warning method and system based on a meteorological early warning machine. BACKGROUND

[0002] The existing meteorological disaster early warning technology is mainly based on a preset fixed threshold triggering mechanism, that is, a critical value of rainfall, wind speed, etc. is set for a single disaster (such as heavy rain, lightning), and when the monitoring data reaches the threshold value, the meteorological early warning machine issues a warning information. Its principle is to rely on meteorological private network, Internet and other networks to interface with the provincial early warning information release system, receive standardized warning signals, and display them on the terminal through text, sound and light alarm and other ways. At the same time, a hierarchical management mode is adopted, and the provincial platform uniformly issues early warning rules, and the municipal and county level platforms are responsible for the push and audit of terminals in the jurisdiction, and the grassroots users receive the state through the terminal feedback.

[0003] The existing meteorological disaster early warning technology has many significant defects. In the risk assessment dimension, only the fixed threshold triggering mechanism of a single disaster is relied on, which cannot capture the composite risk of multiple disasters, such as the simultaneous occurrence of heavy rain and strong wind in plain areas, and the amplification effect of the two disasters is not considered, which may underestimate the actual harm. For mountainous areas, the chain evolution law of disasters is ignored, and it is difficult to predict the time sequence correlation risk of the secondary disasters caused by the early rainfall and the subsequent strong wind predicted by lightning, resulting in a lag in the early warning of secondary disasters.

[0004] Therefore, it is necessary to improve the meteorological disaster early warning method and system based on the meteorological early warning machine in the prior art to solve the above problems. SUMMARY

[0005] The present application overcomes the shortcomings of the prior art and provides a meteorological disaster early warning method and system based on a meteorological early warning machine, aiming to solve the problem that the single fixed threshold triggering mechanism and the unified early warning strategy in the prior art cannot adapt to the chain risk of multiple disasters.

[0006] To achieve the above purpose, the technical scheme adopted by the present application is as follows: a meteorological disaster early warning system based on a meteorological early warning machine, comprising:

[0007] A data acquisition module is used to acquire and preprocess real-time meteorological monitoring data, terrain data, historical disaster data and early warning machine operation data, and to construct an early warning system database;

[0008] A multi-disaster terrain adaptation analysis module is connected to the data acquisition module and is used to extract meteorological features and terrain features based on the preprocessed data, divide the region into plains and mountains according to the terrain type, and perform risk analysis through a double-branch fusion neural network model to output the risk level of each region and each disaster;

[0009] The double-branch fusion neural network comprises a plain synchronous superposition subnetwork and a mountainous area time-series chain subnetwork.

[0010] The plain synchronous superposition subnetwork fuses the waterlogging and gale disaster factors and outputs a waterlogging-gale comprehensive risk grade.

[0011] The early warning response strategy generation module is connected to the multi-disaster terrain adaptation analysis module and is configured to generate a differentiated targeted early warning response strategy according to the risk grades of the plain and the mountainous area and the terrain features.

[0012] The terminal interaction module is connected to the early warning response strategy generation module and is configured to push early warning information to a target object and receive feedback.

[0013] In a preferred embodiment of the present application, the plain synchronous superposition subnetwork quantifies the waterlogging-gale comprehensive risk grade through the following steps:

[0014] The feature extraction layer extracts waterlogging features and gale features through parallel convolution modules.

[0015] The synchronous correlation layer introduces a self-attention mechanism to calculate synchronous correlation weights of the two groups of features and generates a synchronous occurrence coefficient.

[0016] The risk output layer fuses the features through a full connection network, corrects the basic risk index by combining real-time features and a similarity mean of a similar case set, and obtains a plain multi-disaster synchronous superposition risk index.

[0017] The plain multi-disaster synchronous superposition risk index is quantified into a 0-5 grade waterlogging-gale comprehensive risk grade through a mapping rule.

[0018] In a preferred embodiment of the present application, the mountainous area time-series chain subnetwork quantifies the landslide risk grade and the lightning-gale chain risk grade through the following steps:

[0019] The feature extraction layer extracts waterlogging features and gale features through parallel convolution modules.

[0020] The time-series feature extraction layer adopts an LSTM network to process precipitation time-series data and introduces a time decay coefficient to output a mountainous area landslide basic risk index.

[0021] The time-series feature extraction layer adopts a GRU network to learn the time-series correlation of lightning and gale and outputs a chain conduction probability and a time delay factor.

[0022] The time-series feature extraction layer adopts a GRU network to learn the time-series correlation of lightning and gale and outputs a chain conduction probability and a time delay factor.

[0023] The risk output layer fuses real-time features and the average similarity of the similar case set, and respectively calculates a mountain landslide time series chain risk index and a mountain thunderstorm gale chain risk index.

[0024] The mountain landslide time series chain risk index and the mountain thunderstorm gale chain risk index are quantified into 0-5 grades of landslide risk levels and thunderstorm gale chain risk levels through mapping rules.

[0025] In a preferred embodiment of the present application, the differentiated targeted early warning strategy of the early warning calling strategy generation module comprises:

[0026] The plain area is divided into a core layer, a key layer and a general layer according to risk levels and spatial attributes, and the differentiated early warning content and the push frequency are matched;

[0027] The early warning object in the mountain area is divided according to the disaster evolution stage, and the early warning content is updated according to different stages.

[0028] In a preferred embodiment of the present application, the calling system database comprises:

[0029] A real-time meteorological data table stores meteorological monitoring data;

[0030] A terrain space table stores spatial data of terrain types, slopes and geological disaster hidden points, adopts a grid index structure to store terrain data, and differentiates and stores in association for plains and mountains;

[0031] A historical disaster case table stores the start and end time of disaster occurrence, the grid ID set of the affected area, the real-time monitoring sequence of the disaster-causing meteorological element, the terrain feature parameters and the disaster loss description; the cases are stored in classification according to disaster types and terrain types;

[0032] An early warning machine running state table stores terminal online states and user operation records.

[0033] In a preferred embodiment of the present application, the extraction of terrain features comprises: the GDAL library is used for spatial overlay analysis in the plain area to extract the low-lying area proportion, the distance from the river and the river network density; the buffer analysis is used to generate slope grid data in the mountain area to extract the slope, the geological disaster hidden point density and the elevation difference.

[0034] In a preferred embodiment of the present application, the construction steps of the similar case set comprise:

[0035] A feature-disaster association library classified according to disaster types is constructed, and the mapping relationship between the feature vectors of historical disaster cases and disaster results is stored;

[0036] The real-time extracted meteorological and terrain features are combined and converted into a multi-dimensional feature vector, and the Euclidean distance algorithm is used to calculate the similarity of the feature vector with the feature vector of the historical disaster case.

[0037] Screening a set of historical disaster cases with similarity exceeding a threshold, and counting disaster occurrence probability;

[0038] Assigning higher similarity calculation weight to recent cases, and adjusting distance calculation coefficient to strengthen the reflection of new disaster rules;

[0039] Inputting the real-time features and the average similarity of the similar case set into the plain synchronous superposition sub-network and the mountainous area time sequence chain sub-network, and correcting the comprehensive risk index and the chain risk index.

[0040] The application provides a meteorological disaster early warning method based on a meteorological early warning machine, comprising the steps of:

[0041] S1, collecting and preprocessing original data related to meteorological disasters to construct a warning system database; wherein the original data includes real-time meteorological monitoring data, terrain data, historical disaster data and early warning machine operation data;

[0042] S2, extracting meteorological features and terrain features from the preprocessed data, fusing historical disaster data, constructing a double-branch fusion neural network model for multi-disaster risk linkage analysis, and performing risk assessment for plains and mountainous areas respectively to output regional and disaster-specific risk levels;

[0043] S3, generating a targeted early warning strategy according to the risk levels and terrain features of different regions;

[0044] S4, pushing early warning information according to the characteristics and risk levels of the warning objects, and obtaining feedback.

[0045] In a preferred embodiment of the application, the preprocessing in step S1 includes data cleaning, data standardization and spatio-temporal alignment processing;

[0046] Data cleaning filters outliers based on physical thresholds and statistical rules, and uses linear interpolation or neighborhood mean method to complete short-term missing data;

[0047] Data standardization unifies data format and unit, and performs coordinate conversion on spatial data;

[0048] Spatio-temporal alignment matches real-time meteorological monitoring data and terrain data to spatial grids, assigns a unique ID to each spatio-temporal grid, and unifies the time granularity.

[0049] In a preferred embodiment of the application, the training steps of the double-branch fusion neural network model are:

[0050] Use historical data in the warning system database to construct a training set, which contains feature vector-risk level pairs of plain synchronous disaster cases and mountainous area chain disaster cases;

[0051] Using the mean square error loss function, the losses of the outputs of the plain synchronous superposition subnetwork and the mountainous area time sequence chain subnetwork are calculated respectively, and then weighted summation is performed;

[0052] An Adam optimizer is used to optimize the model parameters.

[0053] The present application solves the defects in the background art and has the following advantages:

[0054] (1) The multi-disaster terrain adaptive analysis module of the present application adopts a double-branch fusion neural network architecture, and synchronous superposition models and time sequence chain models are designed respectively for the terrain differences between plains and mountainous areas. The plain model integrates real-time comprehensive risk analysis of waterlogging and gale disaster factors, and the mountainous area model analyzes disaster chain evolution risk based on time decay factor and chain conduction probability. It can accurately capture the risk characteristics of multiple disasters in different terrains. The synchronous superposition of the plain considers the immediate synergistic effect of disasters, and the time sequence chain of the mountainous area focuses on the evolution law of disasters, which directly makes the output of regional and disaster risk level more in line with the actual disaster scene.

[0055] (2) The present application adopts a terrain-driven differentiated targeting strategy. In the plain area, the core layer, the key layer and the general layer are divided according to the risk level and the spatial attribute, and different warning contents and push frequencies are matched. In the mountainous area, monitoring prompts, evacuation instructions and equipment protection guidelines are pushed according to the disaster evolution stage, and a backup channel is enabled for signal blind areas. This strategy fully combines the risk characteristics and terrain features of different regions. The layered push in the plain ensures that high-risk areas receive more intensive and critical information, and the staged push in the mountainous area adapts to the disaster evolution rhythm. The backup channel makes up for the problem of insufficient signal coverage, achieving accurate push and comprehensive coverage of warning information.

[0056] (3) The present application introduces a dynamic construction mechanism for the training set and gives higher weight to recent cases, so that the model can continuously absorb the characteristics of new disaster cases and strengthen the adaptability to recent disaster patterns, enhancing the robustness of the system under the background of climate change. Compared with the limitations of traditional static models relying on historical full data training, the present application greatly improves the sustainable guarantee ability of warning accuracy and further reduces the risk of model failure due to changes in disaster patterns.

[0057] (4) The present application is aimed at the chain evolution characteristics of mountain hazards, constructs a time sequence chain risk model, introduces a time decay factor to quantify the sustained influence of early rainfall on landslides, and designs a chain conduction probability to associate the occurrence time sequence of lightning and gale. The problem of secondary disaster response lag in the prior art is effectively solved, and the present scheme can predict in advance the landslide risk that may be caused by rainfall accumulation and the gale risk indicated by lightning activity. The mountain hazard warning window period is advanced by several hours, which wins critical time for evacuating personnel in high-risk areas and protecting key facilities, and further reduces the losses caused by disaster chain conduction. BRIEF DESCRIPTION OF DRAWINGS

[0058] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0059] Figure 1 is a meteorological disaster early warning response system module diagram of the preferred embodiment of the present application;

[0060] Figure 2 is a meteorological disaster early warning response method flowchart of the preferred embodiment of the present application;

[0061] Figure 3 is a risk linkage analysis model architecture diagram of the preferred embodiment of the present application;

[0062] Figure 4 is an early warning strategy diagram of the preferred embodiment of the present application;

[0063] Figure 5 is a double-branch fusion neural network model structure diagram of the preferred embodiment of the present application. DETAILED DESCRIPTION

[0064] The technical solutions in the embodiments of the present application will be described clearly and completely in the following description with reference to the drawings of the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0065] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, but the present application can also be implemented in other ways different from those described herein, therefore, the scope of protection of the present application is not limited by the specific embodiments disclosed below.

[0066] The present application focuses on the core scene of multi-disaster linkage in meteorological disaster prevention in the region with different terrains of plains and mountains. The existing technology, in the early stage of popularization of meteorological warning machines, takes solving the basic warning information coverage of the grass-roots level as the primary goal, adopts a single fixed threshold triggering mechanism and a unified warning strategy, and is difficult to adapt to the chain risk formed by the superposition of multiple disasters. For example, heavy rain in mountainous areas may trigger landslides and form a compound disaster with short-term strong winds. At the same time, the existing warning strategy does not take into account the differences in terrain between plains and mountains, and it is difficult to meet the differentiated needs of large-scale rapid response in plain areas and lagging defense of secondary disasters in mountainous areas, resulting in insufficient warning accuracy and timeliness.

[0067] The concept of the present application is derived from the deep combination of meteorological disaster rules and terrain characteristics: based on the historical data accumulated by the operation of meteorological warning machines, it is found that multiple disasters in plain areas often occur simultaneously, and the risk has immediate superposition; mountain disasters show a time series chain feature; accordingly, the conventional idea of single threshold and unified strategy is broken, and a differentiated response mechanism for different terrains is proposed: a risk superposition model is used in plain areas to quantify the real-time comprehensive risk of multiple disasters, and a dynamic time series model is used in mountainous areas to predict the evolution trend of disaster chain, while optimizing the scene adaptability of warning information pushing. By integrating real-time meteorological monitoring data, historical disaster cases and terrain characteristics, accurate triggering and hierarchical pushing of multi-disaster linkage are realized, and the grass-roots disaster response capability is improved.

[0068] As shown in Figure 1 A meteorological disaster warning response system based on a meteorological warning machine, comprising modules:

[0069] A data acquisition module for acquiring and preprocessing raw data related to meteorological disasters, wherein the raw data is stored in the response system database, and the raw data includes real-time meteorological monitoring data, terrain data, historical disaster data and warning machine operation data;

[0070] A multi-disaster terrain adaptation analysis module connected to the data acquisition module for extracting meteorological features and terrain features based on preprocessed data, dividing the region into plains and mountains according to terrain types, and performing risk analysis through a double-branch fusion neural network model to output risk levels for different regions and different disaster types;

[0071] Among them, the plain branch adopts a synchronous superposition model, fuses waterlogging and wind disaster factors, and analyzes the real-time comprehensive risk level of multiple disasters; the mountain branch adopts a time series chain model, based on time decay factor and chain conduction probability, to analyze the evolution risk level of disaster chain;

[0072] A warning response strategy generation module connected to the multi-disaster terrain adaptation analysis module for generating differentiated targeted warning response strategies according to the risk levels and terrain characteristics of plains and mountains;

[0073] The terminal interaction module is connected with the early warning calling strategy generation module, and is configured to push early warning information to a target object and receive feedback.

[0074] In the data collection module, a calling system database is created, and a storage structure and data specification of the database are defined, wherein the calling system database comprises a real-time meteorological data table, a terrain space table, a historical disaster case table, and a warning machine running state table.

[0075] The calling system database is a relational database used for centralized storage and management of meteorological disaster early warning calling related data, is constructed based on a database management system, and supports high concurrency reading and writing and spatial data storage.

[0076] The real-time meteorological data table is a structured table used for storing minute / hour dynamic meteorological monitoring data, and comprises a monitoring point ID, a timestamp, rainfall, wind speed, wind direction, humidity, air temperature, radar echo intensity, satellite cloud data and the like.

[0077] The terrain space table is a table structure designed by extending a spatial database, and is used for storing terrain data comprising geometric information, such as plain / mountain division, boundary information, terrain slope, altitude, geological disaster hidden point, land use type, micro-terrain feature and the like.

[0078] The historical disaster case table is a table used for storing details of meteorological disaster events in the past 10 years, and comprises a disaster ID, occurrence time, affected area, disaster type, disaster-causing meteorological element extreme value, loss statistics, associated past calling record and equipment failure record.

[0079] The warning machine running state table is a table used for recording real-time states of meteorological warning machine terminals (township version / building version / enterprise version), and comprises a terminal ID, a timestamp, an online / offline state, signal strength, battery power, and basic user operation record.

[0080] For example, the following table is a table structure design of the real-time meteorological data table.

[0081]

[0082] The remaining tables are also designed in a similar manner, and will not be described in detail here.

[0083] The multi-disaster type terrain adaptation analysis module is constructed based on a distributed computing framework, and realizes quantitative output of multi-disaster type risk levels through collaborative operation of a feature extraction engine, an association rule miner and a risk calculation model.

[0084] The meteorological feature extraction uses a sliding window algorithm, OpenCV library analysis and DBSCAN clustering algorithm to obtain hourly rainfall intensity time series features, radar echo related features and ground lightning density. The terrain feature extraction uses GDAL library, spatial overlay analysis and buffer analysis to obtain mountainous grid slope, plain grid low-lying area proportion and river distance grid data. The feature vector construction associates meteorological and terrain features according to the grid UUID to form a 128-dimensional feature vector, which is stored in the Milvus feature vector database after normalization and compression to the [0, 1] interval.

[0085] The multi-disaster terrain adaptation analysis module needs to construct a feature-disaster correlation library, which is a structured database table set storing the correspondence between the combination of meteorological features and terrain features and historical disaster events, for mining the implicit rules of "feature combination → disaster occurrence";

[0086] The feature combination refers to a set composed of multiple meteorological features and terrain features, where the meteorological features include rainfall (magnitude / duration), wind speed, humidity, radar echo characteristics, and the terrain features include terrain type (plain / mountainous), slope, elevation, geological property and land use type.

[0087] The correlation library table structure design constructs a feature combination table and a disaster event correlation table. The feature combination table stores the unique identifier of the feature combination, i.e., combination ID, and the meteorological feature items and terrain feature items that constitute the combination, and is associated with the corresponding disaster event through the combination ID. The disaster event correlation table stores the combination ID, associated disaster type, qualitative description of disaster occurrence and terrain range where the correlation rule is effective.

[0088] In terms of feature combination classification and mapping rule establishment, sub-libraries are constructed according to disaster type classification. For different disaster types such as heavy rainfall, waterlogging, landslide and lightning, sub-libraries are established respectively, and each sub-library only stores feature combinations and mapping relationships related to the corresponding disaster type. At the same time, based on historical data statistics mapping relationship, by retrieving historical disaster cases, the occurrence frequency of corresponding disaster after the appearance of a specific feature combination is counted to form a mapping rule. The rule content covers feature combination composition, associated disaster type, correlation degree description and applicable terrain range.

[0089] In terms of dynamic maintenance mechanism of the correlation library, regular updates are implemented. Every quarter, according to newly occurred disaster cases, the mapping relationship between feature combination and disaster is supplemented. If the correlation degree of a certain feature combination and disaster in the new case is inconsistent with the record in the library, the correlation rule is updated. In addition, every year, professional teams in the fields of meteorology and geology evaluate the mapping relationship in the correlation library, eliminate the rules that are invalid due to environmental changes such as terrain reconstruction and land use type change, and add new correlation rules that conform to the current disaster rules.

[0090] The early warning call strategy generation module is constructed based on a rule engine and a spatial analysis service, and dynamically generates a targeted early warning strategy according to a risk level and a terrain feature, including specific configurations of a warning object, content, a channel, and a frequency.

[0091] In a plain area: Based on a risk level and a spatial query, a grid set of a core layer and a key layer is screened, population data and important site POIs in the grid are connected through spatial connection, a warning object list is generated, including a name, an address, and a contact method of a responsible person.

[0092] In a mountainous area: The early warning area is divided according to a disaster stage, a warning polygon with a range of 500 meters around a landslide hidden danger point is generated by buffer analysis, key objects such as forest management stations and communication base stations are screened through attribute query, and are stored in a warning object relationship table.

[0093] The push strategy configurator automatically selects a push channel based on a signal coverage state of a warning object: a warning machine terminal, a mobile phone short message, and a WeChat group robot are enabled in a 4G / 5G signal coverage area; an emergency broadcast and a gong and bell team dispatching system are enabled in a signal blind area.

[0094] The terminal interaction module realizes bidirectional communication between a weather warning machine terminal and a system platform through hardware interface adaptation and software protocol encapsulation, supports early warning information pushing, state monitoring, and user feedback collection.

[0095] An exemplary method:

[0096] As shown in Figure 2 A weather disaster early warning call method based on a weather warning machine, including the steps of:

[0097] S1, collecting and preprocessing original data related to weather disasters, and constructing a call system database; wherein the original data includes real-time weather monitoring data, terrain data, historical disaster data, and warning machine operation data;

[0098] S2, extracting weather features and terrain features from the preprocessed data, fusing historical disaster data, constructing a double-branch fusion neural network model for multi-disaster risk linkage analysis, and performing risk assessment for plains and mountainous areas respectively, and outputting a risk level of each region and each disaster;

[0099] S3, generating a targeted early warning call strategy according to the risk level and the terrain feature of different regions;

[0100] S4, pushing early warning call information according to the characteristics of the warning object and the risk level, and obtaining feedback;

[0101] In step S1, real-time meteorological monitoring data, terrain data, historical disaster data and early warning machine operation data are acquired through multiple channels, preprocessed through cleaning, standardization and other operations, and stored in a database to form a full-dimensional basic data resource covering meteorological elements, geographical space, historical rules and equipment status, thereby providing a unified and reliable data input for the meteorological disaster early warning and response system.

[0102] The original data refers to various initial information related to meteorological disaster early warning analysis, including real-time meteorological monitoring data for reflecting the current weather conditions, terrain data for representing geographical spatial characteristics, historical disaster data for recording past disaster patterns, and early warning machine operation data for reflecting the status of early warning equipment. The response system database refers to a structured, safe and controllable data collection formed by storing the preprocessed data according to the specifications, thereby providing reliable data support for subsequent meteorological disaster risk analysis.

[0103] The original data acquisition includes:

[0104] The real-time meteorological monitoring data is collected by the township automatic weather station, including rainfall, wind speed, wind direction, humidity and temperature. The radar echo intensity is obtained through the radar station, and the satellite cloud image remote sensing data is received through the meteorological satellite.

[0105] The terrain data is obtained by accessing the regional GIS map database and extracting spatial attribute data of plains and mountainous areas, including boundary information, terrain slope, elevation and geological disaster hidden point. On this basis, high-resolution remote sensing image data is integrated for dynamically updating micro-terrain features such as valleys and cliffs.

[0106] The GIS map database is a database that stores geographical spatial information, including terrain features, administrative divisions and geological disaster hidden points, thereby providing a basis for obtaining terrain data.

[0107] The historical disaster data is obtained from the Tianxing platform, including disaster types, occurrence time, impact range and loss degree. The historical operation logs of the early warning machine are also obtained, including past early warning and response records and equipment failure records.

[0108] The Tianxing platform is a meteorological big data cloud platform that centrally stores and manages provincial meteorological observation, forecast, early warning and historical disaster data, thereby serving as the core source of historical disaster data.

[0109] The early warning machine operation data is uploaded by the township version, building version and enterprise version early warning machine terminals in real time, including online / offline status, signal strength, battery level and operation records of grassroots users.

[0110] The data preprocessing includes data cleaning, data standardization and spatio-temporal alignment.

[0111] Data cleaning filters unreasonable data and removes outliers based on physical thresholds and statistical rules, including negative values in rainfall monitoring and data exceeding physical limits in wind speed; linear interpolation or neighborhood mean method is used to complete short-term missing data to ensure data continuity; image data such as radar echoes and satellite cloud images are denoised; spatial interpolation is performed on long-term missing data in mountainous radar blind areas;

[0112] Data standardization unifies data format and units, performs coordinate conversion on spatial data, and realizes spatial matching of meteorological data and terrain data; structured coding is performed on text-based feedback data;

[0113] Temporal and spatial alignment matches real-time meteorological monitoring data and terrain data to spatial grids with a range of 5km×5km-10km×10km, maps discrete monitoring point data to grid center points through the nearest neighbor method, performs grid cutting on cross-grid cells to ensure unique terrain properties within each grid; unifies the time granularity (minute) of real-time data and historical data to ensure consistent temporal and spatial properties of data in subsequent analysis and avoid analysis errors caused by temporal and spatial misalignment; assigns a unique ID to each spatio-temporal grid to realize the correlation of meteorological data, terrain data, and historical disaster cases within the same grid;

[0114] The preprocessed data is stored in the response system database to form a multi-table associated, safe and controllable database.

[0115] As shown in Figure 3 Step S2, meteorological features and terrain features are extracted from the preprocessed data, historical disaster data is fused, a multi-disaster risk linkage analysis model is constructed, and a risk level is output;

[0116] Among them, the meteorological feature refers to the core index closely related to the occurrence of meteorological disasters extracted from the preprocessed real-time meteorological monitoring data; the terrain feature refers to the geographical attribute that has a significant impact on the formation and development of disasters extracted from the terrain data; the historical disaster data fusion refers to the correlation analysis of historical disaster cases and real-time extracted features to mine the internal laws of features and disaster occurrence; the multi-disaster risk linkage analysis model refers to a mathematical model that can quantify the risk of multi-disaster superposition or chain occurrence based on the extracted features and historical laws; the risk level is the output of the model, which is used to represent the classification results of the possibility and impact degree of disaster occurrence, and contains temporal and spatial attributes.

[0117] Multi-dimensional feature extraction is used to extract core indicators related to disasters;

[0118] For different disaster types and terrain differences, key features are accurately extracted to provide input variables for model construction:

[0119] Weather feature extraction:

[0120] Heavy rainfall related: Extract hourly rain intensity, 3-hour accumulated rainfall, 6-hour accumulated rainfall, radar echo intensity, radar echo top height; among them, radar echo intensity ≥ 45dBZ is a strong rainfall signal, and radar echo top height reflects the convective development height;

[0121] Lightning related: Extract ground flash frequency, ground flash density, thunderstorm moving speed;

[0122] Gale related: Extract 10-minute average wind speed, maximum wind speed, wind speed duration;

[0123] Feature processing: Calculate the average rain intensity of the past 1 hour for the extracted weather features, to enhance the timeliness of the features.

[0124] Extracting weather features related to heavy rainfall, lightning, and gale is the core of accurately capturing unique disaster factors of different disasters and achieving fine-grained early warning of multiple disaster superposition or chain risk. Heavy rainfall is the core inducement of waterlogging, landslide and other disasters, and its hourly rain intensity and accumulated rainfall directly reflect the intensity and duration of rainfall. Radar echo features can predict the development trend of rainfall. Ground flash data of lightning can quantify the risk of lightning strikes and avoid fires or equipment damage caused by lightning strikes. Wind speed features of gale can assess its potential damage to buildings and outdoor facilities. Heavy rainfall often accompanies lightning and gale to form a compound disaster, and extracting a single feature will lead to one-sided risk assessment.

[0125] Terrain feature extraction:

[0126] Plain area: Extract low-lying area proportion, river network density, distance to river; among them, the low-lying area proportion is the ratio of the low-lying area to the total area of a grid; river network density reflects drainage capacity; the closer the distance to the river, the higher the risk of waterlogging;

[0127] Mountainous area: Extract slope, slope direction, geological disaster hazard point density, elevation difference; among them, slope ≥ 25° is a high-risk landslide, and the sun / shade slope affects soil moisture content, and the elevation difference reflects the degree of terrain undulation;

[0128] General features: Extract population density and number of important places in the grid to assess the impact of disasters.

[0129] Historical disaster data fusion and rule mining are used to establish the association rules between features and disasters. By matching the real-time extracted feature combination with the feature combination of historical disaster cases, the risk probability is directly mapped, providing an objective association basis for the model, which includes:

[0130] Building a feature-disaster correlation library, historical disaster data fusion and rule mining By building a feature-disaster correlation library classified by disaster type, containing complete feature combination and disaster result, the real-time feature combination is converted into a vector and the historical disaster case vector is matched for similarity, similar cases are screened to calculate the probability of disaster occurrence to determine the current risk and refer to the impact degree, and the case library is updated regularly, the recent case weight is adjusted, and controversial cases are handled to establish the correlation rules between features and disasters.

[0131] Classification by disaster type: historical disaster cases are classified by heavy rainfall, landslides, and lightning disasters, so that each disaster type corresponds to specific meteorological and topographic features, achieving precise binding of features and disaster types.

[0132] Quantifying the correlation strength, the probability of disaster occurrence under a specific combination of meteorological and topographic features in historical data is calculated to determine the impact of different feature combinations on disaster occurrence.

[0133] Normalize the feature values of historical disaster cases, convert the hourly rain intensity to a percentage relative to the historical maximum, eliminate the calculation bias caused by different units, and ensure the comparability of real-time features and historical disaster cases.

[0134] Convert the real-time extracted meteorological and topographic feature combination into a multi-dimensional feature vector, and convert the historical disaster cases into feature vectors of the same dimension; use the Euclidean distance algorithm to calculate the similarity between the real-time feature vector and the historical disaster case feature vector, the smaller the distance, the higher the similarity, set the similarity threshold or the number of selected cases, and filter out the case set similar to the current scene;

[0135] Statistical probability of disaster occurrence in similar case set, if more than 70% of the cases have occurred corresponding disaster, it is determined that the risk of disaster occurrence under the current feature combination is high; at the same time, record the disaster impact degree of similar cases as the reference basis for the risk level of the current scene, where the disaster impact degree includes the depth and duration of waterlogging;

[0136] In addition, new disaster cases occurring every quarter are added to the case library, the case must contain complete feature combination and disaster result to ensure the timeliness of the case library;

[0137] Give recent cases a higher similarity calculation weight, multiply a certain proportion of the coefficient in the distance calculation, so that the emerging disaster rules can be reflected in the matching results faster, where the emerging disaster rules include the changes in disaster-causing features caused by climate change;

[0138] For cases with similar feature combinations but significantly different disaster results, automatically mark them as "controversial cases" and manually calibrate them with expert experience to avoid abnormal values interfering with the accuracy of matching.

[0139] A multi-disaster risk linkage analysis model is constructed: a differentiated design of plain and mountain area models; according to the differences in disaster characteristics between plains and mountains, a double-branch fusion neural network model is designed, and risk models are constructed respectively, and through the differentiated structure of the plain branch and the mountain branch, the end-to-end quantification of multi-disaster composite risk is realized. The model input is meteorological, topographical and historical feature vectors, and the output is the risk level (0-5 levels) of each region and each disaster.

[0140] Multi-disaster risk linkage analysis aims at the differences in terrain and disaster characteristics between plains and mountains, and constructs a double-track analysis mechanism for different terrains. Through the differentiated design of the plain synchronous superposition model and the mountain time series chain model, the precise quantification of multi-disaster composite risk is realized. The model input is preprocessed meteorological, topographical and historical data, and the output is the risk level of each region and each disaster, which provides the basis for targeted early warning strategies.

[0141] As shown in Figure 5 The overall architecture logic of the double-branch fusion neural network model is as follows:

[0142] After the input features are classified by the terrain classifier, they are automatically routed to the corresponding branch network:

[0143] The features of the plain area are input into the plain synchronous superposition subnetwork, and the comprehensive risk level of waterlogging and gale is output.

[0144] The features of the mountain area are input into the mountain time series chain subnetwork, and the risk level of landslide and the chain risk level of lightning and gale are output.

[0145] Among them, the plain area focuses on the synchronous superposition risk of waterlogging and gale, and the synergistic amplification effect of multiple disasters occurring at the same time is reflected through the real-time superposition coefficient; the mountain area emphasizes the time series risk of landslide and lightning, and introduces a time decay factor to quantify the conduction effect of the previous disaster on the subsequent disaster, so that the model calculation logic is deeply adapted to the disaster evolution rule dominated by terrain.

[0146] The construction steps of the plain synchronous superposition subnetwork include:

[0147] The multi-disaster synchronous superposition risk model is used in the plain area, which considers the synergistic effect of multi-disaster superposition by quantifying the real-time comprehensive risk of waterlogging and gale occurring at the same time in the plain area; waterlogging and gale often occur at the same time with strong convective weather, and the risk has immediate superposition. By integrating the disaster-causing factors and synchronous occurrence probability of waterlogging and gale, the comprehensive risk level is quantified.

[0148] Input layer:

[0149] 64-dimensional feature vector, including:

[0150] Meteorological features: hourly rainfall, 3-hour / 6-hour accumulated rainfall, 10-minute average wind speed, maximum wind speed, wind speed duration, radar echo intensity (enhanced feature weight when ≥ 50 dBZ);

[0151] Topographic features: low-lying area proportion, distance from river, river network density, population density;

[0152] Historical correlation features: average similarity of real-time features and similar case set, i.e., plain synchronous disaster matching degree ;

[0153] Feature extraction layer:

[0154] Parallel convolution module, through 2 1D convolution kernels, respectively extracts waterlogging features (fuses hourly rainfall, 3-hour / 6-hour accumulated rainfall, low-lying area proportion, distance from river, and river network density) and gale features (10-minute average wind speed, maximum wind speed, wind speed duration, and population density), outputs 2 groups of 32-dimensional feature maps, corresponding to the implicit features of waterlogging risk index and gale risk index;

[0155] Synchronous correlation layer: introduces self-attention mechanism to calculate the synchronous correlation weight of the two groups of features, simulates the synchronous occurrence coefficient ; Statistics the proportion of waterlogging and gale occurring simultaneously in historical data to generate the synchronous occurrence coefficient;

[0156] Risk output layer:

[0157] Fully connected network fuses the above features, maps to 0-5 levels through Sigmoid activation function combined with linear scaling, and outputs the plain comprehensive risk level:

[0158] Retrieves similar cases in the historical disaster case table with the current real-time meteorological features and topographic features, calculates the average similarity of the real-time feature vector and the similar case set, and modifies the basic risk index as a weight;

[0159] Finally, the comprehensive risk index is obtained by the following formula: , wherein, is the plain multi-disaster synchronous superposition risk index, the value range is 0-5 and is a continuous value, the larger the value, the higher the comprehensive risk of waterlogging and gale occurring simultaneously; are the plain waterlogging risk index and the plain gale risk index, respectively; the plain waterlogging risk index is calculated based on hourly rainfall, 3-hour accumulated rainfall, 6-hour accumulated rainfall, low-lying area proportion, and distance from river parameters; the plain gale risk index is calculated based on 10-minute average wind speed, maximum wind speed, wind speed duration, and population density parameters; is the synchronous occurrence coefficient, the value range is 1.0-1.5, calculated from the probability of waterlogging and gale occurring simultaneously under real-time meteorological conditions; The plain synchronous disaster matching degree is the average similarity of the real-time feature vector and the similar case set. The higher the similarity, the closer to 1;

[0160] Further, the risk indicators are quantified into 0-5 integer risk levels by mapping rules;

[0161] Level 0 (no risk): <0.5;

[0162] Level 1 (low risk): 0.5≤ <1.5;

[0163] Level 2 (low-medium risk): 1.5≤ <2.5;

[0164] Level 3 (medium risk): 2.5≤ <3.5;

[0165] Level 4 (high risk): 3.5≤ <4.5;

[0166] Level 5 (extremely high risk): ≥4.5.

[0167] The final output is the comprehensive risk level of waterlogging and strong wind.

[0168] Trigger synchronous superposition calculation when any of the following conditions are met:

[0169] Radar echo intensity ≥ 50 dBZ (strong convective signal) and at the same time monitoring maximum wind speed ≥ 10 m / s;

[0170] In the last 1 hour, the waterlogging risk index is greater than or equal to 2 and the strong wind risk index is greater than or equal to 1.

[0171] In the conditions that trigger synchronous superposition calculation, when the matching degree of real-time features and historical synchronous cases is ≥ 0.7, even if the radar echo intensity standard is not met, the superposition calculation is still triggered.

[0172] Further, the plain synchronous superposition feature adaptation adds a strong convective synchronous monitoring index to determine the potential of synchronous occurrence of waterlogging and strong wind through the combination of radar echo top height and vertical wind shear, providing real-time calculation basis for the synchronous occurrence coefficient For grids with a synchronous event proportion ≥ 30% in the historical disaster case library, automatically upgrade the baseline value, so as to ensure that the superposition effect of high-risk areas is not underestimated.

[0173] The construction steps of the mountainous area time sequence chain sub-network include:

[0174] The mountainous area adopts a multi-disaster time sequence chain risk model. For the evolution risk of disaster chain, the chain evolution risk of "precipitation-landslide" and "thunder and lightning-gale" in mountainous areas is predicted, considering the time sequence correlation and cumulative effect of disasters. Mountain disasters show time sequence chain characteristics. Precipitation in the early stage is easy to trigger landslides in the later stage, and lightning activity often predicts subsequent gales. The model introduces time decay factor and chain conduction probability. The matching degree of historical chain cases is used as a correction term for time decay factor and conduction probability to quantify the evolution risk of disaster chain.

[0175] Input layer: 64-dimensional feature vector, including:

[0176] Landslide-related features: current hourly rain intensity, past 6 / 12 / 24-hour cumulative rainfall, slope, geological disaster hidden point density, radar echo intensity, real-time precipitation characteristics, and similarity average of similar case set;

[0177] Thunder and lightning-gale-related features: ground flash frequency, ground flash density, current wind speed, elevation difference, real-time lightning characteristics, and similarity average of similar case set;

[0178] Time sequence feature extraction layer:

[0179] For landslide risk path: LSTM network is used to process time sequence data of precipitation, including 6 / 12 / 24-hour cumulative rainfall. Time decay coefficient is introduced through the gating mechanism to simulate the cumulative effect of precipitation, and the basic risk index of mountain landslide is output. The time decay coefficient is set based on the time interval between precipitation and landslide in historical landslide cases to quantify the sustained impact of precipitation.

[0180] For thunder and lightning-gale risk path: GRU network is used to learn the time sequence correlation of thunder and lightning and gale. The time sequence attention mechanism is used to capture the rules of gale 1 hour after lightning, and the chain conduction probability and time delay factor are output.

[0181] Based on the extracted time sequence features, the landslide risk level and the thunder and lightning-gale chain risk level are finally output through two groups of fully connected layers. The calculation logic includes:

[0182] Mountain landslide time sequence chain risk index: wherein, is the similarity average of real-time precipitation characteristics and historical early-stage precipitation-late-stage landslide chain cases, is the mountain landslide time sequence chain risk index, with a numerical range of 0-5 and a continuous numerical value, representing the cumulative impact of early-stage precipitation on current landslide risk; is the basic risk index of mountain landslide, calculated based on slope, geological disaster hidden point density, and radar echo intensity parameters; is the time decay coefficient, representing the decay rate of the impact of early-stage precipitation over time; The radar echo intensity quantization value is for a time period t, t is the time period, and the first 6 hours, 12 hours, and 24 hours are selected as the key time periods to reflect the cumulative effect of precipitation of different lengths.

[0183] Mountain lightning-gale chain risk index:

[0184] The frequency of gale within 1 hour after lightning in historical disaster cases is taken as the chain conduction probability, and the time delay factor (based on the time sequence law of lightning and gale occurrence) is calculated to obtain the mountain lightning-gale chain risk index: , wherein, is the mountain lightning-gale chain matching degree, which is calculated based on the similarity between the real-time lightning characteristics and the historical lightning-gale chain cases; is the mountain lightning-gale chain risk index, which is a continuous value ranging from 0 to 5; is the mountain lightning risk index, which is calculated based on parameters such as ground flash frequency and ground flash density; is the chain conduction probability, which is determined by the frequency of gale within 1 hour after lightning in historical data; is the time delay factor, which is 1.0 for 0-1 hour after lightning, 0.5 for 1-3 hours after lightning, and 0 for more than 3 hours after lightning, reflecting the decay law of time sequence correlation.

[0185] The output mountain landslide time sequence chain risk index and mountain lightning-gale chain risk index are both continuous values, which are mapped to 0-5 through the same mapping rule to form landslide risk levels and lightning-gale chain risk levels, respectively.

[0186] Dynamic adaptation mechanism:

[0187] Further, the mountain time sequence chain feature adaptation is achieved by establishing a pre-impact factor library to store the time interval distribution of chain cases in the mountain area in the past 10 years, providing data support for the dynamic adjustment of the time decay coefficient and the time delay factor , and when the soil moisture content is ≥80% (saturated state), it is automatically increased to 0.8 to shorten the decay period and strengthen the chain impact of pre-precipitation.

[0188] For the double-branch fusion neural network model, the historical data of the response system database are used, including synchronous disaster cases in the plain and chain disaster cases in the mountain area in the past 10 years, and the training set is constructed according to the feature vector-risk level pair;

[0189] The loss function adopts mean square error (MSE) loss, and the loss is calculated for the double-branch output and then weighted and summed, wherein the sample weights of the plain and the mountain area are distributed according to the regional disaster occurrence frequency;

[0190] The optimizer uses the Adam optimizer, and gives a weight of 1.2 to the cases in the past two years to enhance the adaptability of the model to the recent disaster patterns.

[0191] The model output is:

[0192] The output form is:

[0193] Plain area: single value (0-5 level), representing the comprehensive risk level of waterlogging and strong wind;

[0194] Mountainous area: double value (0-5 level), representing landslide risk level and lightning-strong wind chain risk level respectively.

[0195] In this embodiment, through the differential design of the plain synchronous superposition subnetwork and the mountainous time sequence chain subnetwork, the risk calculation logic is accurately matched with the disaster characteristics dominated by the terrain: the plain captures the synergistic effect of the simultaneous superposition of multiple disasters, and the mountainous area quantifies the time sequence correlation of disasters through time decay factor and chain conduction probability, so that the risk level output by the double-branch fusion neural network model not only contains the intensity of a single disaster, but also reflects the multi-disaster evolution law driven by the terrain, providing more practical quantitative basis for targeted early warning response strategies.

[0196] As shown in Figure 4 , in step S3, a targeted early warning response strategy is generated based on the risk level and terrain characteristics. The core is to match different early warning objects, content and frequency according to the disaster risk differences and evolution laws of plains and mountains, to ensure that early warning information accurately reaches high-risk areas and populations, and to improve response efficiency.

[0197] Targeted early warning response scheme for plain areas:

[0198] According to the multi-disaster synchronous superposition risk index of the plain, population density and important site distribution, the early warning objects are divided into core layer, key layer and general layer. Different layers are designed with immediate risk avoidance instructions, prevention measures guidelines and risk prompt differentiated early warning content, and the early warning frequency is dynamically adjusted according to the risk level.

[0199] The core layer is the area with a risk level of ≥3.5 (high risk) and a low-lying area proportion of ≥50%, covering low-lying residential areas, underground parking lots and urban waterlogging points; the key layer is the area with a risk level of 2.5≤ <3.5 (medium-high risk) and a distance from the river of ≤500 meters, covering farmland along the river and industrial parks; the general layer is the area with a risk level of <2.5 (medium-low risk), mainly ordinary residential areas.

[0200] The core layer early warning content needs to include immediate risk avoidance instructions, clearly indicating the risk avoidance location, evacuation route, and prohibited matters; the focus layer early warning content focuses on prevention measures guidance, including: reinforcing outdoor billboards, suspending high-altitude operations, checking drainage facilities; the general layer early warning content is mainly risk prompt, reminding to pay attention to subsequent early warning information and close doors and windows.

[0201] When ≥ 3.5 and the strong convection synchronous monitoring index is continuously triggered, the early warning is pushed every 15 minutes; when 2.5 ≤ < 3.5, it is pushed every 30 minutes; when < 2.5, it is pushed every 1 hour. If the risk level decreases by ≥ 30%, the frequency is automatically reduced to the next level.

[0202] Targeted early warning calling scheme for mountainous areas:

[0203] According to the mountain landslide time series chain risk index and the lightning-gale chain risk index , combined with the density of geological disaster hidden points, the early warning objects are divided according to the disaster evolution stage, the early warning content is updated according to the stage time sequence, and multi-channel cooperative pushing is adopted according to the signal coverage difference to ensure information coverage.

[0204] Landslide risk dominant stage ≥ 3.5, the early warning objects are preferentially villages, mountain road construction teams within 500 meters of the landslide hidden point; lightning-gale chain stage ≥ 3.5, the early warning objects are extended to forest area management and protection stations and high mountain communication base stations.

[0205] In the early stage of landslide risk (pre-precipitation accumulation stage), the early warning content is mainly monitoring prompt, requiring grassroots monitoring personnel to intensify patrol of valleys and cliffs; in the middle stage of landslide risk ≥ 3.5 and soil moisture content ≥ 80%, evacuation preparation instructions are pushed, clearly indicating the evacuation assembly point and the list of items to carry; in the lightning-gale chain stage, the early warning content focuses on equipment protection guidance, including closing forest power, suspending outdoor work, and reinforcing base station antennas.

[0206] For signal coverage differences in mountainous areas, multi-channel cooperative pushing is adopted: for areas with 4G / 5G signal, graphic and text early warning is pushed through early warning machine terminal, mobile phone short message, and village-level WeChat group; for signal weak areas, township emergency broadcasting, gong and drum team, and grid personnel door-to-door notification are linked to ensure the "last kilometer" coverage of early warning information.

[0207] In step S4, the push of the early warning call information and the acquisition of the feedback are key links for realizing the early warning closed-loop management. The information is ensured to reach the target through multi-channel pushing, and the response effect is monitored through real-time feedback, thereby providing a basis for dynamically adjusting the early warning strategy.

[0208] Based on the regional characteristics and risk levels of the early warning objects, the differentiated pushing channels are dispatched: high-risk areas in plains ≥3.5, the early warning machine terminal (sound and light alarm + text display) in the town version, the early warning machine (elevator screen scrolling play) in the building version, the mobile phone pop-up window pushing, and the synchronous linkage of the urban emergency broadcast system are preferentially used; high-risk areas in mountains ≥3.5, the village-level early warning machine, the emergency broadcast vehicle, the unmanned aerial vehicle air broadcast, and the supplementary paper early warning sheet sent by the grid personnel are mainly used.

[0209] The first round of pushing of the high-risk early warning information needs to be completed within 5 minutes after the strategy is generated, and the second round of pushing of the high-risk early warning information needs to be completed within 10 minutes (for the objects that are not confirmed to receive). The pushing of the medium and low-risk early warning information needs to be completed within 15 minutes. The channel delivery rate is monitored in real time during the pushing process. When the channel delivery rate is less than 70%, the backup channel is automatically switched (for example, when the early warning machine terminal is offline, the short message is immediately re-sent).

[0210] According to the ideal embodiments of the present application, through the above description, relevant personnel can make various changes and modifications without deviating from the technical idea of the present application. The technical scope of the present application is not limited to the content of the specification, and must be determined according to the scope of the claims.

Claims

1. A meteorological disaster early warning and response system based on a meteorological early warning machine, characterized in that, The system comprises a module: a data acquisition module for acquiring and preprocessing real-time meteorological monitoring data, terrain data, historical disaster data and early warning machine operation data, and constructing a response system database; a multi-disaster type terrain adaptation analysis module connected to the data acquisition module, for extracting meteorological features and terrain features based on the preprocessed data, dividing the region into plains and mountainous areas in combination with terrain types, performing risk analysis through a double-branch fusion neural network model, and outputting regional and disaster-specific risk levels; the double-branch fusion neural network comprises a plain synchronous superposition subnetwork and a mountainous area time series chain subnetwork; wherein the plain synchronous superposition subnetwork fuses waterlogging and wind disaster factors to output a waterlogging-wind comprehensive risk level; and the mountainous area time series chain subnetwork outputs a landslide risk level and a lightning-wind chain risk level based on a time decay factor and a chain conduction probability; an early warning response strategy generation module connected to the multi-disaster type terrain adaptation analysis module, for generating a differentiated targeted early warning response strategy according to the risk levels and terrain features of plains and mountainous areas; a terminal interaction module connected to the early warning response strategy generation module, for pushing early warning information to target objects and receiving feedback.

2. The weather disaster early warning and response system based on the weather early warning machine according to claim 1, characterized in that: The plain synchronous superposition subnetwork quantifies the waterlogging-wind comprehensive risk level through the following steps: inputting a feature vector containing meteorological features, terrain features and historical correlation features; extracting waterlogging features and wind features by a parallel convolution module in a feature extraction layer; introducing a self-attention mechanism in a synchronous correlation layer to calculate the synchronous correlation weight of the two sets of features and generate a synchronous occurrence coefficient; fusing features by a full connection network in a risk output layer, combining the real-time features with the similarity mean of the similar case set to correct the base risk index, and obtaining a plain multi-disaster synchronous superposition risk index; quantifying the plain multi-disaster synchronous superposition risk index into a 0-5 level waterlogging-wind comprehensive risk level through a mapping rule.

3. The weather disaster early warning and response system based on the weather early warning machine according to claim 1, characterized in that: The mountainous area time series chain subnetwork quantifies the landslide risk level and the lightning-wind chain risk level through the following steps: inputting a feature vector containing landslide-related features, lightning and wind-related features and time series features; using an LSTM network to process the time series data of precipitation and introducing a time decay coefficient in a time series feature extraction layer to output a mountainous area landslide base risk index; using a GRU network to learn the time series correlation of lightning and wind and output a chain conduction probability and a time delay factor in a time series feature extraction layer; fusing real-time features with the similarity mean of the similar case set in a risk output layer to calculate a mountainous area landslide time series chain risk index and a mountainous area lightning and wind chain risk index, respectively; quantifying the mountainous area landslide time series chain risk index and the mountainous area lightning and wind chain risk index into a 0-5 level landslide risk level and lightning-wind chain risk level, respectively, through a mapping rule.

4. The weather disaster early warning and response system based on the weather early warning machine according to claim 1, characterized in that: The differentiated targeted early warning response strategy of the early warning response strategy generation module comprises: dividing the plains into a core layer, a key layer and a general layer according to risk levels and spatial attributes, matching differentiated early warning content and push frequency; dividing the mountainous areas into early warning objects according to disaster evolution stages and updating early warning content according to different stages.

5. The weather disaster early warning and response system based on the weather early warning machine according to claim 1, characterized in that: The response system database comprises: a real-time meteorological data table for storing meteorological monitoring data; A terrain space table stores spatial data of terrain types, slopes, and geological disaster hidden points, and adopts a grid index structure to store terrain data, and differentiates and stores in association for plains and mountainous areas; A historical disaster case table stores data including: start and end times of disasters, grid ID sets of affected areas, real-time monitoring sequences of disaster-causing meteorological elements, terrain characteristic parameters, and disaster loss descriptions; cases are stored in classification according to disaster types and terrain types; A warning machine operation state table stores terminal online states and user operation records.

6. The weather disaster early warning and response system based on the weather early warning machine according to claim 1, characterized in that: Extraction of terrain characteristics includes: spatial overlay analysis of GDAL library for plain areas to extract low-lying area proportion, distance from river, and river network density; and buffer analysis for mountainous areas to generate slope grid data to extract slope, geological disaster hidden point density, and elevation difference.

7. The weather disaster early warning and response system based on the weather early warning machine according to claim 2, characterized in that: The construction steps of a similar case set include: A feature-disaster association library classified according to disaster types is constructed to store the mapping relationship between feature vectors of historical disaster cases and disaster results; Real-time extracted meteorological and terrain features are combined to convert into a multi-dimensional feature vector, and the similarity of the feature vector to historical disaster case feature vectors is calculated using the Euclidean distance algorithm; A set of historical disaster cases with a similarity exceeding a threshold is screened, and the disaster occurrence probability is calculated; A higher similarity calculation weight is given to recent cases, and the reflection of new disaster rules is adjusted and strengthened through distance calculation coefficients; The average similarity of real-time features to the similar case set is input into a plain synchronous superposition sub-network and a mountainous area time series chain sub-network to correct the comprehensive risk index and the chain risk index.

8. A method for meteorological disaster early warning and response based on a meteorological early warning machine, a meteorological disaster early warning and response system based on the meteorological early warning machine according to any one of claims 1-7, characterized in that, The steps include: S1, collect and pre-process original data related to meteorological disasters, and construct a warning system database; wherein the original data includes: real-time meteorological monitoring data, terrain data, historical disaster data, and warning machine operation data; S2, extract meteorological features and terrain features from the pre-processed data, fuse historical disaster data, construct a double-branch fusion neural network model for multi-disaster risk linkage analysis, and perform risk assessment for plains and mountainous areas respectively to output risk levels of different regions and different disasters; S3, generate a targeted warning and response strategy according to the risk levels of different regions and terrain characteristics; S4, push warning and response information according to the characteristics and risk levels of the warning objects, and obtain feedback.

9. The weather disaster early warning and response method based on the weather early warning machine according to claim 8, characterized in that: The pre-processing in step S1 includes data cleaning, data standardization, and spatio-temporal alignment processing; Data cleaning filters abnormal values based on physical thresholds and statistical rules, and uses linear interpolation or neighborhood mean method to complete short-time missing data; Data standardization unifies data formats and units, and performs coordinate conversion on spatial data; Spatio-temporal alignment matches real-time meteorological monitoring data and terrain data to spatial grids, assigns a unique ID to each spatio-temporal grid, and unifies the time granularity.

10. The weather disaster early warning response method based on the weather early warning machine of claim 8, characterized in that: The training steps of the double-branch fusion neural network model are: Use historical data in the warning system database to construct a training set containing feature vector-risk level pairs of plain synchronous disaster cases and mountainous area chain disaster cases; Use a mean square error loss function to calculate the loss of the outputs of the plain synchronous superposition sub-network and the mountainous area time series chain sub-network respectively, and then weighted sum; Use the Adam optimizer to optimize the model parameters.

Images