Intelligent on-duty early warning method and system fusing meteorological business data

Abstract

The application relates to the technical field of meteorological early warning, and discloses an intelligent on-duty early warning method and system fusing meteorological business data, which comprises a historical prior database, a plurality of historical prediction types and a plurality of historical characteristic data, and each historical prediction type corresponds to a historical characteristic data; a collection unit collects meteorological data, radar data and satellite data of a preset area to form characteristic data; a judgment unit collects the characteristic data and compares the characteristic data with the historical prior database to determine a predicted meteorological type; when a strong convective burst early warning occurs, a processing unit performs data fusion on the characteristic data based on deep learning; when a slow risk early warning occurs, the processing unit performs data fusion on the characteristic data based on Kalman filtering, and determines an early warning grade according to a fusion result. The application automatically selects the deep learning and the Kalman filtering method, and improves the accuracy and real-time performance of a meteorological early warning system.

Description

Technical Field

[0001] This invention relates to the field of meteorological early warning technology, and more specifically, to an intelligent duty early warning method and system that integrates meteorological operational data. Background Technology

[0002] With the dramatic increase in meteorological data volume, modern meteorological early warning systems need to process data from multiple diverse sources, such as weather stations, radar, and satellites. This data not only involves real-time changes in meteorological parameters but also includes rich spatial and temporal information. However, in traditional meteorological early warning systems, the integration and efficient analysis of meteorological data still face multiple challenges.

[0003] Currently, meteorological early warning systems generally rely on manual judgment or traditional mathematical models for data processing, especially for warnings of severe convective weather and slowly changing weather. Traditional methods process and analyze meteorological data through rule-based pattern recognition. While this can provide some early warning effectiveness in simple weather conditions, its limitations are becoming increasingly apparent as meteorological models become more complex. Meteorological data comes from a wide range of sources, including surface meteorological data, radar data, and satellite data. These data have different resolutions, timeliness, and reliability. Accurately applying multi-source data has become a major challenge for meteorological early warning systems.

[0004] Therefore, it is necessary to design an intelligent duty and early warning method and system that integrates meteorological operational data to solve the problems existing in the current technology. Summary of the Invention

[0005] In view of this, the present invention proposes an intelligent duty and early warning method and system that integrates meteorological operational data, aiming to solve the problems of current reliance on manual judgment leading to difficulties in the application of meteorological data and low forecast accuracy.

[0006] In one aspect, this invention proposes an intelligent duty and early warning system that integrates meteorological operational data, comprising:

[0007] The historical prior database includes several historical forecast types and several historical feature data, with each historical forecast type corresponding to a historical feature data, which includes historical meteorological data, historical radar data, and historical satellite data.

[0008] The acquisition unit is used to acquire meteorological data, radar data, and satellite data of a preset area, and to process the meteorological data, radar data, and satellite data based on the mutual information method to form feature data.

[0009] The judgment unit is configured to collect the feature data and compare the feature data with the historical prior database, and determine the forecast weather type based on the comparison result. The forecast weather type includes severe convective weather sudden warning and slow risk warning.

[0010] The processing unit is configured to perform data fusion on the feature data based on deep learning when the forecast weather type is a severe convective weather sudden warning, and determine the sudden warning level based on the fusion result; and to perform data fusion on the feature data based on Kalman filtering when the forecast weather type is a slow risk warning, and determine the slow risk warning level based on the fusion result.

[0011] Furthermore, when the acquisition unit processes the meteorological data, radar data, and satellite data based on the mutual information method to form feature data, it includes:

[0012] The meteorological data includes temperature, humidity, air pressure, wind speed, and precipitation.

[0013] The radar data includes radar echo maps;

[0014] The satellite data includes cloud images, precipitation maps, and temperature distribution maps;

[0015] The meteorological data, radar data, and satellite data are preprocessed, including outlier removal and interpolation to fill in missing data, and spatial resolution alignment is performed on the meteorological data, radar data, and satellite data.

[0016] Convert the preprocessed meteorological data, radar data, and satellite data into discrete form;

[0017] The mutual information between meteorological data, radar data, and satellite data is calculated separately, and the data with the highest mutual information value among the same meteorological type data is retained to form the feature data.

[0018] Furthermore, when the judgment unit determines the forecast weather type based on the comparison results, it includes:

[0019] The neighborhood radius is determined based on the radar echo range, and MinPts is initialized to 4.

[0020] Based on DBSCAN clustering, the feature data is clustered with the historical feature data in the historical prior database to determine the cluster in which the feature data belongs. If there is a historical forecast type corresponding to the historical feature data in the cluster in which the feature data belongs, which is a historical severe convective weather sudden warning, the forecast meteorological type of the feature data is determined to be the severe convective weather sudden warning. If the historical forecast type corresponding to the historical feature data in the cluster in which the feature data belongs is a historical slow risk warning, the forecast meteorological type of the feature data is determined to be the slow risk warning.

[0021] Furthermore, when the forecast weather type is a severe convective weather emergency warning, the processing unit performs data fusion on the feature data based on deep learning, including:

[0022] The processing unit inputs the feature data into the deep learning model;

[0023] Convolutional neural network layer: used to extract spatial features from radar and satellite data, including precipitation intensity, cloud morphology, and radar echoes;

[0024] Long Short-Term Memory (LSTM) network layer: used to process the temporal features in meteorological data and capture the temporal trends of temperature, humidity, and wind speed;

[0025] Fusion layer: The outputs of the CNN layer and the LSTM layer are fused by weighted averaging to generate comprehensive features;

[0026] Fully connected layer: Classifies the fused integrated features and outputs a risk coefficient for sudden strong convection warning;

[0027] Output layer: Generates the probability value and risk coefficient of the final severe convective weather emergency warning.

[0028] Furthermore, when the forecast weather type is a slow risk warning, the processing unit performs data fusion on the feature data based on Kalman filtering, including:

[0029] The processing unit predicts the current weather conditions based on the historical prior database and the system dynamics model.

[0030] By comparing the feature data with the prediction results, adjusting the prediction results according to the Kalman gain, and updating the weather state estimate;

[0031] Use meteorological models to describe the changes in meteorological data and update the current meteorological status.

[0032] Based on the covariance of prediction error and observation error, the Kalman gain is calculated, and the weighted average of the feature data and the prediction data is used as the final meteorological state estimate.

[0033] A slow risk warning coefficient is generated based on the output of the Kalman filter.

[0034] Furthermore, when the processing unit determines the sudden warning level or the slow risk warning level based on the fusion result, it includes:

[0035] The processing unit compares the strong convection sudden warning risk coefficient with the strong convection sudden warning risk threshold, or compares the slow risk warning coefficient with the slow risk warning threshold, and determines the sudden warning level or the slow risk warning level based on the comparison result.

[0036] The level of the emergency warning is directly proportional to the difference between the risk coefficient of the emergency warning for severe convective weather and the risk threshold for the emergency warning for severe convective weather.

[0037] The slow risk warning level is directly proportional to the difference between the slow risk warning coefficient and the slow risk warning threshold.

[0038] Furthermore, it also includes:

[0039] The adjustment unit is configured to, after determining the emergency warning level, collect the location information of the preset area and determine whether to adjust the emergency warning level based on the location information;

[0040] When the location information is in a mountainous area, the adjustment unit determines to adjust the emergency warning level; when the location information is not in a mountainous area, the adjustment unit determines not to adjust the emergency warning level.

[0041] Furthermore, when the adjustment unit determines to adjust the emergency warning level, it includes:

[0042] The adjustment unit obtains the average altitude of the preset area, compares the average altitude with a first altitude and a second altitude respectively, and adjusts the emergency warning level according to the comparison results; the first altitude is less than the second altitude;

[0043] When the average altitude is less than or equal to the first altitude, the adjustment unit raises the emergency warning level by one level; when the average altitude is greater than the first altitude and less than or equal to the second altitude, the adjustment unit raises the emergency warning level by two levels; when the average altitude is greater than the second altitude, the adjustment unit raises the emergency warning level by three levels.

[0044] Furthermore, it also includes:

[0045] The intelligent question-answering module is used to support forecasters to interact via natural language and quickly query the historical prior database;

[0046] The feedback module provides suggestions based on forecasters' query results, generates decision support reports, and automatically updates decision reference content based on historical prior data and the current warning status.

[0047] Compared with existing technologies, the advantages of this invention are as follows: by introducing a historical prior database, multi-source data fusion, and a combination of deep learning and Kalman filtering, the accuracy and real-time performance of the meteorological early warning system are improved. The historical prior database integrates meteorological, radar, and satellite data from many years into a unified data source, facilitating effective comparison and analysis when facing complex meteorological events. The acquisition unit accurately extracts relevant features from meteorological, radar, and satellite data using mutual information, reducing data redundancy and improving data quality. The judgment unit intelligently identifies the meteorological type based on the comparison results, ensuring that the early warning system adopts different data fusion strategies according to severe convective weather or slowly changing risk conditions. The deep learning model focuses on sudden warnings of severe convective weather, enabling rapid response and complex pattern recognition; Kalman filtering effectively handles slowly changing weather, accurately predicting long-term trends and subtle changes. Through this technological combination, efficient and accurate early warnings for complex and variable meteorological phenomena are achieved, improving the intelligence and practicality of meteorological early warning systems.

[0048] On the other hand, this application also provides an intelligent duty and early warning method that integrates meteorological operational data, applied to the aforementioned intelligent duty and early warning system that integrates meteorological operational data, including:

[0049] Collect historical prior data, which includes several historical forecast types and several historical feature data, and each historical forecast type corresponds to a historical feature data, which includes historical meteorological data, historical radar data and historical satellite data;

[0050] Meteorological data, radar data, and satellite data of a preset area are collected, and feature data are formed by processing the meteorological data, radar data, and satellite data based on the mutual information method.

[0051] The feature data is collected and compared with the historical prior database. The forecast weather type is determined based on the comparison result. The forecast weather type includes severe convective weather sudden warning and slow risk warning.

[0052] When the forecast weather type is a severe convective weather sudden warning, the feature data is fused based on deep learning, and the sudden warning level is determined according to the fusion result; when the forecast weather type is a slow risk warning, the feature data is fused based on Kalman filtering, and the slow risk warning level is determined according to the fusion result.

[0053] It is understandable that the aforementioned intelligent duty and early warning methods and systems that integrate meteorological operational data have the same beneficial effects, and will not be elaborated upon here. Attached Figure Description

[0054] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0055] Figure 1 A functional block diagram of an intelligent duty and early warning system that integrates meteorological operational data, provided in an embodiment of the present invention;

[0056] Figure 2 This is an application diagram of the intelligent duty and early warning system that integrates meteorological business data provided in an embodiment of the present invention;

[0057] Figure 3 A schematic diagram illustrating the application of the intelligent question-and-answer module in the intelligent duty and early warning system that integrates meteorological operational data, as provided in an embodiment of the present invention.

[0058] Figure 4 A flowchart of an intelligent duty and early warning method that integrates meteorological operational data, provided in an embodiment of the present invention. Detailed Implementation

[0059] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0060] Traditional meteorological early warning systems face the challenge of processing multi-source heterogeneous data. Differences in spatiotemporal resolution and modal mismatches exist between surface meteorological data, radar data, and satellite data, leading to low feature extraction efficiency. Furthermore, the lack of a dynamic correlation analysis mechanism for historical prior data makes accurate classification between sudden severe convective weather warnings and slow-moving risk warnings impossible. During the fusion of time-series data such as temperature and humidity collected by meteorological stations with spatial data such as radar echo maps and satellite cloud images, traditional rule engines struggle to eliminate cross-modal data redundancy, resulting in an explosion of feature data dimensions. The comparison methods between historical feature databases and real-time data rely on static thresholds, which cannot adapt to the non-linear changes in meteorological elements, leading to an increased misclassification rate for warning types.

[0061] For example, in a severe convective weather monitoring scenario, a regional meteorological center needs to simultaneously process minute-level temperature and humidity time series uploaded by ground meteorological stations, reflectivity factor matrices updated by radar every 6 minutes, and multispectral image data acquired by geostationary meteorological satellites every 15 minutes. The ground data sampling frequency is 1Hz, the radar data spatial resolution is 1km×1km, and the satellite imagery includes 0.5km grid data for both visible and infrared channels. During data preprocessing, missing values ​​need to be interpolated and filled in; however, traditional methods use fixed-window linear interpolation, resulting in data reconstruction errors exceeding 40% in areas where radar beam obstruction is caused by mountainous terrain. During feature construction, manually set meteorological element association rules cannot effectively quantify the nonlinear relationship between radar echo intensity and precipitation. The curse of dimensionality occurs during DBSCAN clustering due to the lack of dimensionality reduction in the feature dimensions, causing the similarity calculation between real-time data and the historical feature database to take longer than the early warning response time window.

[0062] If the above problems are not resolved, spatiotemporal alignment errors of multi-source data will lead to distortion of the spatial distribution of feature data, affecting the tracking accuracy of thunderstorm cells in severe convective weather; inconsistency in matching historical data with real-time features will reduce the reliability of warning type discrimination, and there may be a risk of misjudging short-term heavy precipitation as continuous precipitation in sudden rainstorm scenarios; defects in cross-modal data fusion will weaken the system's ability to capture small- and medium-scale weather processes, resulting in a severe convective weather warning lead time of less than 30 minutes, which cannot meet the requirements of emergency response.

[0063] Therefore, in some embodiments of this application, see reference. Figure 1-2As shown, this application first considers how to effectively integrate multi-source heterogeneous data and achieve dynamic correlation between historical data and real-time features. Traditional methods use fixed interpolation windows to process radar and satellite data, leading to a significant increase in reconstruction errors in complex terrain areas. Furthermore, manually set correlation rules cannot capture the nonlinear relationships across modal data. To address this, this application establishes a priori database containing historical forecast types and multi-dimensional feature data. It quantifies the statistical dependencies between meteorological, radar, and satellite data using mutual information, and selects the feature combinations with the highest mutual information values ​​to eliminate redundancy. For the matching problem between real-time data and the historical database, a density clustering algorithm is used to adaptively partition the data space, dynamically determining the current meteorological model based on the historical warning types within the cluster. In the data fusion stage, a deep learning model is designed for the sudden characteristics of severe convective weather, using convolutional networks to extract spatial features and combining them with temporal analysis. For slowly changing processes, Kalman filtering is introduced to achieve dynamic correction of state estimation, thereby adapting to the data processing needs of different warning scenarios.

[0064] An intelligent duty and early warning system integrating meteorological operational data includes: a historical prior database, comprising several historical forecast types and several historical feature data, with each historical forecast type corresponding to a historical feature data, including historical meteorological data, historical radar data, and historical satellite data; a data acquisition unit, used to acquire meteorological data, radar data, and satellite data for a preset area, and process the meteorological data, radar data, and satellite data based on mutual information to form feature data; a judgment unit, configured to acquire feature data and compare the feature data with the historical prior database, and determine the forecast meteorological type based on the comparison result, including severe convective weather sudden warning and slow risk warning; and a processing unit, configured to perform data fusion on the feature data based on deep learning when the forecast meteorological type is a severe convective weather sudden warning, and determine the sudden warning level based on the fusion result; and to perform data fusion on the feature data based on Kalman filtering when the forecast meteorological type is a slow risk warning, and determine the slow risk warning level based on the fusion result.

[0065] The historical prior database refers to a structured dataset storing information related to historical meteorological events, including historical forecast types and their corresponding multi-dimensional data such as meteorological, radar, and satellite data. By establishing a mapping relationship between historical events and data features, it provides a benchmark for real-time data comparison. The mutual information method uses mutual information from information theory to measure the statistical dependence between different data sources. By calculating the mutual information values ​​between meteorological, radar, and satellite data, it selects the feature combinations with the highest correlation to eliminate redundant information and enhance data representation capabilities. The comparison of feature data with the historical prior database involves matching the real-time collected meteorological features with historical data for similarity. Clustering algorithms are used to identify the category pattern of the current data, thereby associating it with known historical warning types and achieving rapid classification decisions. Deep learning data fusion uses convolutional neural networks and long short-term memory networks to extract spatial and temporal features respectively, and generates comprehensive features through weighted fusion. This is suitable for modeling complex nonlinear relationships in sudden severe convective weather warnings. Kalman filter data fusion, based on system dynamics models and observational data, achieves state estimation through recursive prediction and error correction. This is suitable for dynamic tracking and smoothing of time-series data in slow-risk warnings.

[0066] This application solves the problem that traditional early warning systems are unable to effectively integrate multi-source heterogeneous data and adapt to different weather types by using a mutual information screening and intelligent classification processing mechanism for multi-source meteorological data, combined with a differentiated fusion strategy of deep learning and Kalman filtering. It significantly improves the accuracy and timeliness of early warnings for severe convective weather and slowly changing risks.

[0067] In this application, the historical prior database stores historical forecast types and historical feature data, with each historical forecast type corresponding to a set of historical feature data. The acquisition unit collects meteorological data, radar data, and satellite data for a preset area, and processes these data using mutual information to form feature data. The judgment unit compares the feature data with the historical prior database to determine whether the forecast meteorological type is a severe convective weather sudden warning or a slow-risk warning. The processing unit employs different data fusion methods based on the forecast type: for severe convective weather sudden warnings, deep learning is used for data fusion to determine the sudden warning level; for slow-risk warnings, Kalman filtering is used for data fusion to determine the slow-risk warning level.

[0068] The historical prior database provides a reference for current forecasts by storing historical data. The acquisition unit uses mutual information to process multi-source data, effectively extracting key features. The judgment unit makes a preliminary judgment on the forecast type through comparison. The processing unit employs two methods—deep learning and Kalman filtering—for different forecast types, suitable for the characteristics of sudden and gradual weather processes, respectively. Deep learning can capture complex nonlinear features and is suitable for handling sudden weather events such as severe convection; Kalman filtering can smooth time-series data and is suitable for slowly changing weather processes. Through this classification and processing approach, the accuracy of early warnings for different types of weather can be improved in a targeted manner.

[0069] As a preferred embodiment, the specific implementation of this application's solution is as follows: The historical prior database contains historical forecast types and corresponding historical feature data from the past 5 years. The acquisition unit collects meteorological station data, weather radar data, and meteorological satellite data within a preset area every 10 minutes. Meteorological station data includes temperature, humidity, air pressure, wind speed, and precipitation; radar data is the reflectivity factor in the radar echo image; satellite data includes visible light and infrared channel images. The collected raw data undergoes outlier removal and missing value imputation preprocessing. Then, the mutual information method is used to calculate the mutual information between different data sources, and the feature combination with the highest mutual information value is selected to constitute feature data. The judgment unit uses the DBSCAN clustering algorithm to cluster the feature data with historical data, and determines the current forecast type based on the historical forecast type of the cluster to which the feature data belongs. For cases determined to be severe convective weather sudden warnings, the processing unit uses a deep learning model containing convolutional layers and long short-term memory (LSTM) network layers for data fusion. The convolutional layers extract spatial features, and the LSM network layers capture temporal features. The fused features are processed through a fully connected layer to obtain the severe convective weather sudden warning risk coefficient. For cases deemed as slow risk warnings, the processing unit uses Kalman filtering for data fusion, predicts the current meteorological state based on historical data and system dynamics models, and then refines the prediction by combining it with observational data to finally obtain a slow risk warning coefficient. The final warning level is determined by comparing the warning coefficient with a preset threshold.

[0070] Through the above scheme, this application achieves effective integration and dynamic correlation analysis of multi-source heterogeneous meteorological data. The application of mutual information method solves the problem of low feature extraction efficiency and can effectively eliminate data redundancy. The introduction of DBSCAN clustering algorithm overcomes the limitations of traditional static threshold methods and improves the accuracy of warning type discrimination. The classification application of deep learning model and Kalman filter has optimized the processing of sudden severe convective weather warnings and slow risk warnings, respectively, improving the system's warning capability for different types of weather processes. This intelligent warning method that integrates multi-source data significantly improves the overall performance of the meteorological warning system, especially the accuracy and timeliness of warnings under complex weather conditions.

[0071] In some of the solutions described above in this application, during the process of integrating meteorological data, radar data, and satellite data into feature data by the acquisition unit, the multi-source data suffers from resolution differences, data loss, and interference from redundant information, resulting in reduced processing efficiency and insufficient reliability.

[0072] This application further proposes meteorological data including temperature, humidity, air pressure, wind speed, and precipitation; radar data including radar echo maps; and satellite data including cloud maps, precipitation maps, and temperature distribution maps. Preprocessing is performed on the meteorological, radar, and satellite data, including outlier removal and interpolation to fill in missing data, and spatial resolution alignment is applied to the meteorological, radar, and satellite data. The preprocessed meteorological, radar, and satellite data are then converted into discrete forms. Mutual information is calculated between the meteorological, radar, and satellite data, and the data with the highest mutual information value among the same meteorological type is retained to form feature data.

[0073] In the preprocessing stage, outlier removal and interpolation are used to fill in missing data and eliminate noise interference caused by equipment failure or transmission errors. Spatial resolution alignment uses a gridding method to unify radar and satellite data into the same spatial grid as meteorological data; for example, matching radar echo maps with the latitude and longitude coordinates of ground meteorological stations to ensure spatial consistency. Discrete transformation uses binning to divide continuous meteorological parameters into several intervals; for example, wind speed is divided into discrete levels such as 0-5 and 6-10. Mutual information calculation uses joint probability distribution to statistically analyze the correlation between different data sources, such as the correlation between radar echo intensity and precipitation, and selects data with mutual information values ​​higher than a preset threshold as valid features.

[0074] Specifically, in the outlier removal step, the standard deviation method is used to identify and remove data points exceeding three times the standard deviation of the mean, avoiding interference from extreme outliers. Kriging interpolation is used to fill in missing data from ground meteorological stations, and the spatial correlation of neighboring stations is used to generate a continuous data field. Spatial resolution alignment further adjusts the pixel resolution of satellite cloud images from the kilometer level to the hundred-meter level grid consistent with radar data through resampling; for example, the original satellite data is projected onto a 100m × 100m grid using bilinear interpolation. During the discrete form conversion process, precipitation is divided into four discrete levels: no rain, light rain, moderate rain, and heavy rain, and temperature is divided into intervals of 5℃. In mutual information calculation, the spatial distribution correlation between radar echo images and satellite cloud images is quantified using the entropy method, retaining strongly correlated data with mutual information values ​​higher than 0.8. For example, when the mutual information value between radar echo intensity and satellite precipitation map reaches 0.85, only radar echo data is retained as the main representation of precipitation characteristics in that area, thereby reducing data redundancy and improving the representation efficiency of feature data. As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0075] When the data acquisition unit processes meteorological data, radar data, and satellite data based on the mutual information method to form feature data, it includes:

[0076] Meteorological data includes temperature, humidity, air pressure, wind speed, and precipitation;

[0077] Radar data includes radar echo maps;

[0078] Satellite data includes cloud images, precipitation maps, and temperature distribution maps;

[0079] Preprocessing is performed on meteorological data, radar data, and satellite data. Preprocessing includes outlier removal and filling in missing data using interpolation methods. Spatial resolution alignment is also performed on the meteorological data, radar data, and satellite data.

[0080] Convert the preprocessed meteorological data, radar data, and satellite data into discrete form;

[0081] The mutual information between meteorological data, radar data, and satellite data is calculated separately. The data with the highest mutual information value among the same meteorological data is retained to form the feature data.

[0082] Specifically, the data acquisition unit first collects raw data from weather stations, radar, and satellites. Weather station data includes hourly temperatures, relative humidity, air pressure, wind speed, and precipitation. Radar data consists of radar echo maps every 6 minutes. Satellite data includes hourly cloud images, precipitation maps, and temperature distribution maps.

[0083] Furthermore, these raw data are preprocessed. Outliers are removed using the 3σ criterion, and missing data are imputed using linear interpolation. Then, all data are unified onto a 10km × 10km grid to achieve spatial resolution alignment.

[0084] Therefore, the preprocessed continuous data is discretized. For example, the temperature is divided into several intervals of 5°C, and the wind speed is divided into several intervals of 2 m / s. For image data, the pixel values ​​are discretized at certain intervals.

[0085] This involves calculating the mutual information between various data points. For example, calculating the mutual information values ​​between temperature and humidity, temperature and radar echo intensity, and radar echo and satellite cloud imagery. For each weather type (such as severe convection, heavy rain, etc.), the feature combination with the highest mutual information value is selected to form the final feature dataset.

[0086] Through the above technical solutions, this application achieves effective fusion of multi-source heterogeneous meteorological data, improving the representativeness and information content of feature data. Preprocessing and spatial alignment resolve the inconsistency problem of data from different sources. A feature selection method based on mutual information effectively reduces data redundancy and extracts the most representative feature combinations. This improves the accuracy of subsequent early warning models, reduces computational resource consumption, and enables the system to respond more quickly to sudden weather events.

[0087] In some of the solutions described above in this application, when the judgment unit determines the forecast weather type by comparing feature data with historical prior database, there is a problem of insufficient accuracy in judging the similarity of data features, which may lead to misjudgment of sudden severe convective weather warnings or slow risk warnings, affecting the reliability of the warning system.

[0088] This application further proposes determining the neighborhood radius based on radar echo range, initializing MinPts to 4; clustering the feature data with historical feature data in the historical prior database based on DBSCAN clustering to determine the cluster in which the feature data belongs; if there is a historical forecast type corresponding to the feature data in the cluster that is a historical severe convective weather sudden warning, the forecast meteorological type of the feature data is determined to be a severe convective weather sudden warning; if the historical forecast type corresponding to the historical feature data in the cluster is a historical slow risk warning, the forecast meteorological type of the feature data is determined to be a slow risk warning.

[0089] When determining the neighborhood radius based on radar echo range, the value of the neighborhood radius can be dynamically adjusted according to the area or intensity of the radar echo coverage area. For example, in mountainous terrain with a large radar echo coverage area, the neighborhood radius can be set to 10 kilometers; in plains areas, it can be set to 5 kilometers. Initializing MinPts to 4 defines the minimum number of samples required to form core objects in DBSCAN clustering. This value is verified through historical data and can balance the needs of cluster density and noise point filtering. During DBSCAN clustering, feature data and historical feature data are matched for similarity using spatial distance metrics. When the boundary distance between a feature data point and a historical feature data cluster is less than the neighborhood radius, it is determined to belong to that cluster.

[0090] Specifically, in the clustering process, the preprocessed feature data is first mapped to a multi-dimensional feature space, where each dimension corresponds to a meteorological parameter. Core points, boundary points, and noise points are identified by calculating the density reachability of feature data points with other historical data points in their neighborhood. For example, if a feature data point has at least four historical data points in its neighborhood, and at least one of these historical data points is marked as a historical severe convective weather warning, the feature data point is classified into a severe convective weather warning cluster. Furthermore, if multiple historical warning types exist within this cluster, the forecast type of the current feature data is determined through a majority voting mechanism. By dynamically adjusting the neighborhood radius, the system can adapt to the differences in meteorological data distribution across different regions. For example, the neighborhood radius can be expanded in areas with wide radar echo coverage to avoid over-segmentation, while the neighborhood radius can be reduced in densely distributed data areas to improve clustering accuracy. This effectively solves the misjudgment problem of traditional clustering algorithms in scenarios with complex spatial distribution of meteorological data, improving the accuracy of warning type determination.

[0091] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0092] When the judgment unit determines the forecast weather type based on the comparison results, it includes the following steps:

[0093] The neighborhood radius is determined based on the radar echo range. In this embodiment, the neighborhood radius is set to 10 kilometers, a value determined based on typical radar echo ranges. Simultaneously, the MinPts parameter is initialized to 4, a key parameter in the DBSCAN clustering algorithm that represents the minimum number of points required to form a cluster.

[0094] Next, the feature data is clustered with historical feature data in the historical prior database using the DBSCAN clustering algorithm. The specific implementation steps of the DBSCAN algorithm are as follows:

[0095] For each point in the feature dataset, calculate the number of points in its neighborhood (i.e., a circular area with a radius of 10 kilometers).

[0096] A point is marked as a core point if it has at least four neighboring points (including itself).

[0097] All directly density-reachable core points are grouped into one cluster.

[0098] For non-core points, if they are in the neighborhood of a core point, they are assigned to the cluster containing that core point.

[0099] Points not assigned to any cluster are considered noise points. Since both the historical prior data and current data used in this scheme are correlated with meteorological data, there are no cases where feature data is assigned as noise points. DBSCAN clustering is used to determine the cluster to which the feature data belongs. Then, it is determined whether there are any historical forecasts in that cluster corresponding to historical severe convective weather warnings. If so, the forecast meteorological type for the feature data is determined to be a severe convective weather warning.

[0100] If the historical forecast types corresponding to the historical feature data in the cluster to which the feature data belongs are all historical slow risk warnings, then the forecast weather type of the feature data is determined to be a slow risk warning.

[0101] Through the above technical solution, this application can effectively utilize historical data to determine the type of weather forecast. The DBSCAN clustering algorithm can group similar meteorological characteristic data together, thereby identifying potential severe convective weather or slow-risk weather. This density-based clustering method can handle irregularly shaped clusters and has good robustness to noisy data, making it suitable for analyzing complex and variable meteorological data. Furthermore, by setting appropriate neighborhood radii and MinPts parameters, the clustering accuracy can be flexibly adjusted to adapt to different weather forecasting needs. This method not only improves the accuracy of weather forecasts but also enables timely identification of potential severe convective weather, providing reliable technical support for weather warnings.

[0102] In some of the schemes mentioned above in this application, after determining that the forecast weather type is a severe convective weather emergency warning, it is necessary to fuse multi-source meteorological data to accurately assess the level of the emergency risk. However, radar data and satellite data contain complex spatial characteristics, and meteorological data have dynamic temporal characteristics. Traditional fusion methods are difficult to capture both spatial correlations and temporal evolution patterns simultaneously, resulting in limited data fusion effects and affecting the accuracy of warning level judgment.

[0103] This application further proposes a processing unit that inputs feature data into a deep learning model; a convolutional neural network layer is used to extract spatial features from radar and satellite data, including precipitation intensity, cloud morphology, and radar echoes; a long short-term memory network layer is used to process temporal features in meteorological data, capturing the temporal trends of temperature, humidity, and wind speed; a fusion layer fuses the outputs of the CNN and LSTM layers through weighted averaging to generate comprehensive features; a fully connected layer classifies the fused comprehensive features and outputs a severe convective weather sudden warning risk coefficient; and an output layer generates the final probability value and risk coefficient of the severe convective weather sudden warning.

[0104] The convolutional neural network layer employs multi-layered two-dimensional convolutional kernels, using a sliding window to scan radar echo maps and satellite cloud images to extract local spatial features. The long short-term memory network layer has 64 hidden units, receiving meteorological parameters arranged in a time series to construct temporal state transition relationships. The fusion layer has weight allocation coefficients of 0.6 and 0.4, corresponding to the contributions of spatial and temporal features, respectively. The fully connected layer contains three hidden layers, using ReLU as the activation function, and outputs the risk coefficient classification results through a softmax function. The output layer maps the risk coefficient to probability values ​​and uses a sigmoid function for normalization.

[0105] Specifically, when the forecast weather type is determined to be a severe convective weather alert, the preprocessed radar echo image and satellite cloud image are input into the convolutional neural network layer. A 3×3 convolutional kernel is used for feature extraction to capture the spatial distribution pattern of the precipitation area. Meteorological data is input into the long short-term memory network layer at 10-minute intervals, and a gating mechanism tracks the changes in humidity and wind speed. The fusion layer linearly weights the 128-dimensional feature vector output from the convolutional neural network layer and the 64-dimensional feature vector output from the long short-term memory network layer to generate a 192-dimensional comprehensive feature vector. The fully connected layer performs a non-linear transformation on the comprehensive feature vector, outputting a risk coefficient between 0 and 1. When the risk coefficient exceeds 0.7, a level-three alert is triggered. The output layer converts the risk coefficient into a probability value and calibrates it using historical disaster data for the preset area, ultimately generating an alert report containing both the probability value and the risk level.

[0106] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0107] When a severe convective weather warning is issued, the processing unit inputs the feature data into the deep learning model. The deep learning model includes the following hierarchical structure:

[0108] The convolutional neural network layer is used to extract spatial features from radar and satellite data. This layer uses multiple convolutional kernels to perform convolution operations on the input data to extract spatial features such as precipitation intensity, cloud morphology, and radar echoes. Specifically, a convolutional layer with a 3x3 kernel and a stride of 1 is used, followed by the ReLU activation function and a max-pooling layer.

[0109] The Long Short-Term Memory (LSTM) network layer is used to process the temporal features in meteorological data. This layer contains LSTM units, each with an input gate, a forget gate, and an output gate, used to capture the temporal trends of temperature, humidity, and wind speed. The hidden state dimension of the LSTM layer is set to 128.

[0110] The fusion layer combines the outputs of the CNN and LSTM layers using a weighted average to generate a comprehensive feature. The weighting coefficients are optimized using a backpropagation algorithm.

[0111] The fully connected layer classifies the fused integrated features and outputs a risk coefficient for severe convective weather emergencies. This layer uses the Softmax activation function, and the number of output nodes is the same as the number of predefined risk levels.

[0112] The output layer generates the final probability value and risk coefficient for severe convective weather emergencies. The probability value is mapped to the range of 0-1 using a sigmoid function, while the risk coefficient is output directly.

[0113] Through the above technical solution, this application achieves effective fusion and analysis of multi-source meteorological data. The deep learning model can automatically extract key features from meteorological data and capture complex spatiotemporal patterns. The convolutional neural network layer effectively processes radar and satellite image data, extracting spatial features; the long short-term memory network layer can capture the temporal changes of meteorological parameters. The fusion layer integrates features from different sources to generate a more comprehensive representation of meteorological conditions. The fully connected layer and output layer further classify the fused features and calculate warning probabilities, providing more accurate severe convective weather warnings. Compared with traditional methods, this deep learning-based approach can better handle large-scale multi-source meteorological data, improving the accuracy and timeliness of severe convective weather warnings.

[0114] In some of the solutions mentioned above in this application, a data fusion method based on deep learning was proposed to handle sudden severe convective weather warnings. However, when handling slow risk warnings, due to the temporal nature of meteorological data and the differences in the reliability of multi-source data, traditional methods are difficult to dynamically adjust the weights of forecast and observation data, resulting in inaccurate meteorological state estimation.

[0115] This application further proposes a processing unit that predicts the current meteorological state based on a historical prior database and a system dynamics model; updates the meteorological state estimate by comparing the feature data with the prediction results and adjusting the prediction results according to the Kalman gain; updates the current meteorological state by using a meteorological model to describe the change process of the meteorological data; calculates the Kalman gain based on the covariance of the prediction error and the observation error, and uses the weighted average result of the feature data and the prediction data as the final meteorological state estimate; and generates a slow risk warning coefficient based on the output of the Kalman filter.

[0116] The system dynamics model is used to establish the mathematical relationship between meteorological conditions and changes over time, such as describing the dynamic evolution of temperature and humidity through differential equations. The Kalman gain is calculated based on the covariance matrix, dynamically adjusting the weights of predicted and observed data by updating the statistical characteristics of prediction and observation errors in real time. During the weighted averaging process, the weight of observed data increases with its reliability; for example, when the real-time performance of radar data is superior to that of satellite data, radar data is given a higher weight.

[0117] Specifically, the meteorological model establishes the evolution patterns of meteorological parameters based on historical prior data to generate the predicted state at the current moment. The difference between the observed data and the predicted results is compensated for by Kalman gain, where the gain value is determined by the ratio of the prediction error covariance to the observation error covariance. For example, when the observed data error increases due to equipment noise, the Kalman gain automatically reduces the weight of the observed data, minimizing the impact of noise on the final estimate. By iteratively updating the covariance matrix and state estimate, the dynamic changes of the meteorological data are accurately modeled. The final output meteorological state estimate is used to calculate the slow risk warning coefficient, where the coefficient value reflects the cumulative degree of deviation of the meteorological state from the normal range. Therefore, the determination of the slow risk warning level is based on persistent meteorological anomalies, rather than instantaneous fluctuations, improving the stability and reliability of the warning results.

[0118] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0119] When the forecast weather type is a slow risk warning, the processing unit performs data fusion on the feature data based on Kalman filtering. Specifically, the processing unit first predicts the weather conditions at the current moment based on historical prior databases and system dynamics models. For example, a linear state-space model is used to describe the changes in meteorological parameters over time.

[0120] Furthermore, by comparing the feature data with the prediction results, the prediction results are adjusted according to the Kalman gain to update the meteorological state estimate. The feature data includes real-time observations of temperature, humidity, and air pressure.

[0121] Therefore, meteorological models are used to describe the changing processes of meteorological data and update the current meteorological state. Specifically, nonlinear Kalman filtering algorithms, such as extended Kalman filtering or unscented Kalman filtering, can be used to handle the nonlinear characteristics of the meteorological system.

[0122] Based on the covariance of prediction and observation errors, the Kalman gain is calculated, and the weighted average of the characteristic data and the prediction data is used as the final meteorological state estimate. The Kalman gain determines the weights of the observed and predicted values ​​in the final estimate.

[0123] Finally, a slow risk warning coefficient is generated based on the output of the Kalman filter. For example, the filtered meteorological state estimate can be compared with a preset threshold to calculate the risk coefficient.

[0124] Through the above technical solutions, this application achieves effective fusion of multi-source meteorological data, improving the accuracy and reliability of slow risk warnings. The Kalman filter algorithm can effectively handle noise and uncertainty in meteorological data, while considering the influence of historical data and real-time observations, making the warning results more stable and reliable. Furthermore, this method can adaptively adjust the prediction model to adapt to changes under different meteorological conditions, enhancing the flexibility and robustness of the warning system.

[0125] In some of the solutions described above in this application, the meteorological risk coefficient obtained through deep learning models or Kalman filtering is only used as an intermediate result for early warning judgment, but it is not clear how to convert the risk coefficient into a specific early warning level. Since the degree of damage from different meteorological disasters has a non-linear relationship, relying solely on the risk coefficient may lead to inaccurate early warning level classification and fail to provide a refined basis for emergency response decisions.

[0126] This application further proposes that the processing unit compares the risk coefficient of a sudden severe convective weather warning with the risk threshold of a sudden severe convective weather warning, or compares the risk coefficient of a slow risk warning with the risk threshold of a slow risk warning, and determines the level of a sudden warning or the level of a slow risk warning based on the comparison results; the level of a sudden warning is directly proportional to the difference between the risk coefficient of a sudden severe convective weather warning and the risk threshold of a sudden severe convective weather warning; the level of a slow risk warning is directly proportional to the difference between the risk coefficient of a slow risk warning and the risk threshold of a slow risk warning.

[0127] The determination of the warning level is based on the proportional relationship between the risk coefficient and the threshold. In severe convective weather warning scenarios, when the risk coefficient exceeds the threshold, a larger difference indicates stronger atmospheric instability, corresponding to a higher level of sudden warning. In slow-risk warning scenarios, the magnitude of the difference reflects the degree to which the meteorological system deviates from its normal state; for example, the extent of sustained higher temperatures is positively correlated with drought risk levels. In practice, multiple difference ranges can be set to correspond to different levels; for example, a difference exceeding the threshold by 20% is defined as a Level 1 warning, and exceeding 50% is defined as a Level 2 warning.

[0128] Specifically, after obtaining the risk coefficient, the processing unit first accesses a preset threshold database to retrieve the baseline threshold for the corresponding warning type. For a severe convective weather warning, if the risk coefficient is 0.75 and the threshold is 0.6, a difference of 0.15 corresponds to a Level 3 warning. If the risk coefficient rises to 0.9, a difference of 0.3 triggers a higher Level 1 warning. If the initial risk level is the highest, it remains unchanged. This difference ratio mechanism reflects the potential development trend of meteorological disasters. When the system detects a continuous increase in the difference, it can activate the upgraded response plan in advance. By establishing a strict mathematical ratio, the subjectivity of human experience-based judgment is avoided, ensuring that warning levels at different times and in different regions are comparable and operable.

[0129] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0130] The processing unit compares the risk coefficient of a sudden severe convective weather warning with the risk threshold of a sudden severe convective weather warning, or compares the risk coefficient of a slow risk warning with the risk threshold of a slow risk warning, and determines the level of a sudden warning or a slow risk warning based on the comparison results.

[0131] Specifically, the level of a sudden emergency warning is directly proportional to the difference between the risk coefficient and the risk threshold of a sudden emergency warning for severe convective weather. Similarly, the level of a slow-risk warning is directly proportional to the difference between the risk coefficient and the threshold of a slow-risk warning.

[0132] For example, for a sudden warning of severe convective weather, a risk threshold of 0.7 can be set. When the risk coefficient is 0.8, the difference is 0.1, corresponding to a Level 1 warning; when the risk coefficient is 0.9, the difference is 0.2, corresponding to a Level 2 warning; and when the risk coefficient is 1.0, the difference is 0.3, corresponding to a Level 3 warning.

[0133] For slow risk warnings, a risk threshold of 0.5 can be set. When the risk coefficient is 0.6, the difference is 0.1, corresponding to a Level 1 warning; when the risk coefficient is 0.7, the difference is 0.2, corresponding to a Level 2 warning; and when the risk coefficient is 0.8, the difference is 0.3, corresponding to a Level 3 warning.

[0134] Through the above technical solution, this application achieves a refined classification of sudden and slow-risk warnings for severe convection. This allows for dynamic adjustment of the warning level based on the difference between the risk coefficient and the threshold, improving the accuracy and relevance of the warnings. Furthermore, by establishing a direct proportional relationship between the difference between the risk coefficient and the threshold and the warning level, the classification of warning levels becomes more reasonable and scientific, helping meteorological departments to take appropriate preventative measures.

[0135] In some of the solutions mentioned above in this application, the determination of the emergency warning level is based solely on the difference between the risk coefficient and the threshold, without considering the impact of geographical location on the emergency warning level. This may lead to a mismatch between the warning level and the actual risk, especially in special terrain conditions such as mountainous areas where insufficient warnings are likely to occur.

[0136] This application further proposes an adjustment unit, which is configured to collect location information of a preset area after the emergency warning level is determined, and determine whether to adjust the emergency warning level based on the location information; when the location information is a mountainous area, it is determined to adjust the emergency warning level; when the location information is not a mountainous area, it is determined not to adjust the emergency warning level.

[0137] Location information is obtained through a geographic information system, including latitude and longitude coordinates and terrain classification data; the judgment logic is based on matching the mountain identification field in the terrain classification data; the adjustment operation is achieved by modifying the numerical range corresponding to the emergency warning level, such as shifting the risk coefficient range corresponding to the original level upward by a certain percentage; the criteria for determining mountainous areas can be combined with terrain slope, altitude or administrative division for comprehensive judgment, for example, areas with a slope greater than 15 degrees and an altitude higher than 500 meters are defined as mountainous areas.

[0138] Specifically, after the processing unit outputs a sudden warning level, the adjustment unit automatically triggers a location information collection process. This involves retrieving terrain data from the geographic information database to extract the terrain classification identifier for the current preset area. If the terrain classification identifier includes a mountain type, the level adjustment mechanism is triggered. During the adjustment process, the level increase is determined based on a comparison of the average elevation of the preset area with the first and second elevations. For example, if the first elevation is set at 500 meters and the second at 1500 meters, the level is increased by one level when the average elevation is below 500 meters; by two levels when it is between 500 and 1500 meters; and by three levels when it exceeds 1500 meters. This adjustment process is achieved by modifying the level threshold parameters, for example, reducing the original level threshold by 10%, 20%, or 30%, thus increasing the warning level corresponding to the same risk coefficient. The adjusted level data is transmitted to the warning release module via a communication interface, simultaneously updating historical warning records in the database. Therefore, in mountainous terrain conditions, the system can dynamically correct the warning level, compensating for the shortcomings of solely relying on risk coefficient assessment.

[0139] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0140] The intelligent duty and early warning system, which integrates meteorological operational data, also includes an adjustment unit. After determining the level of a sudden warning, the adjustment unit collects location information of a preset area and determines whether to adjust the warning level based on this information. If the location is in a mountainous area, the adjustment unit decides to adjust the warning level. If the location is not in a mountainous area, the adjustment unit decides not to adjust the warning level.

[0141] Specifically, the adjustment unit can obtain topographic data of a preset area through a Geographic Information System (GIS). For example, the adjustment unit can use digital elevation model (DEM) data to determine whether the preset area is mountainous. Furthermore, the adjustment unit can set an altitude threshold, such as defining areas with an altitude above 1000 meters as mountainous. Thus, when the average altitude of the preset area exceeds 1000 meters, the adjustment unit will determine that the area is mountainous and adjust the emergency warning level accordingly.

[0142] Through the aforementioned technical solution, this application enables dynamic adjustment of warning levels based on terrain features, improving the accuracy and relevance of warnings. Due to the complex terrain and more drastic changes in meteorological conditions in mountainous areas, adjusting warning levels in these areas allows the system to better address specific meteorological risks and provide more timely and effective warning information to residents and relevant departments. This location-based warning adjustment mechanism enhances the adaptability and practicality of the warning system, helps reduce warning errors caused by terrain factors, and thus improves the overall warning effectiveness.

[0143] In some of the solutions mentioned above in this application, after determining the level of the emergency warning, the adjustment unit can decide whether to adjust the level based on whether the location information is a mountainous area. However, it does not further consider the differences in the degree of impact of mountainous terrain at different altitudes on the sudden weather, resulting in insufficient precision in the adjustment of the warning level and difficulty in adapting to the actual warning needs under complex terrain conditions.

[0144] This application further proposes that when the adjustment unit determines to adjust the emergency warning level, it obtains the average altitude of a preset area, compares the average altitude with the first altitude and the second altitude respectively, and adjusts the emergency warning level according to the comparison results; the first altitude is less than the second altitude; when the average altitude is less than or equal to the first altitude, the adjustment unit raises the emergency warning level by one level; when the average altitude is greater than the first altitude and less than or equal to the second altitude, the adjustment unit raises the emergency warning level by two levels; when the average altitude is greater than the second altitude, the adjustment unit raises the emergency warning level by three levels.

[0145] The first and second elevations are set based on the differences in the response of different terrains to weather systems; for example, the first elevation can be set to 500 meters and the second to 1500 meters. The average elevation is obtained by extracting digital elevation model data of the preset area from a geographic information system and calculating the average value of the gridded elevation. Upgrading the emergency warning level is achieved through linear overlay or nonlinear mapping; for example, if the original level is level three, upgrading it by one level makes it level four.

[0146] Specifically, after the processing unit outputs the emergency warning level, the adjustment unit retrieves the location coordinates of a preset area from the geographic information database and determines whether the area belongs to the mountainous region category through spatial matching. If it is a mountainous region, the elevation data of the area is further extracted to calculate the average altitude. By dividing the average altitude into intervals with two preset thresholds, a correspondence between altitude and the adjustment amount of the warning level is established. For example, low-altitude mountainous areas, due to the weaker orographic lifting effect, only need to be raised by one level; mid-to-high-altitude mountainous areas, due to the intensified vertical airflow activity, need to be raised by two or three levels. This adjustment operation applies to the initial warning level generated by the processing unit, and the final corrected level is transmitted to the warning issuing terminal through the communication module. Thus, the warning system can dynamically adjust the warning intensity of sudden weather events based on the altitude gradient of mountainous areas, improving the accuracy of warnings under complex terrain conditions.

[0147] As a preferred embodiment, the solution of this application is implemented as follows: After determining the emergency warning level, the adjustment unit obtains the geographic coordinate information of a preset area and analyzes the average altitude of the area through a GIS system. When the system determines that the area belongs to mountainous terrain, the warning level adjustment mechanism is triggered. In specific implementation, the first altitude is set to 500 meters, and the second altitude is set to 1000 meters. If the average altitude of the area is 300 meters, the adjustment unit raises the emergency warning level from the original level three to level two; if the average altitude is 800 meters, it is raised from level three to level one; if the average altitude is 1500 meters, it is directly raised from level three to the highest level warning. The adjusted warning level is sent to the duty terminal in real time through the communication module, and the visual markers on the warning map are updated synchronously.

[0148] Through the above technical solution, this application can dynamically optimize the classification of early warning levels based on the altitude of mountainous terrain, effectively solving the problem of insufficient early warning sensitivity of traditional meteorological early warning systems in complex terrain areas due to altitude differences. The tiered adjustment mechanism avoids resource waste caused by excessively raising early warning levels in low-altitude areas, while ensuring higher safety redundancy in risk assessment of sudden meteorological events in high-altitude areas, significantly improving the accuracy of early warning decisions and the efficiency of emergency response.

[0149] In some of the schemes mentioned above in this application, the adjustment unit determines whether to adjust the emergency warning level based on the location information of the preset area, but does not consider the impact of different altitudes on meteorological risks, resulting in a lack of terrain adaptability in the adjustment of the warning level and an inability to accurately reflect the amplification effect of complex mountain terrain on severe convective weather.

[0150] This application further proposes a method for adjusting the emergency warning level based on the adjustment unit determination, including obtaining the average altitude of a preset area, comparing the average altitude with a first altitude and a second altitude respectively, and adjusting the emergency warning level according to the comparison results, wherein the first altitude is lower than the second altitude.

[0151] The adjustment unit obtains the average altitude data of a preset area through a geographic information system or elevation database, quantified in meters. The first altitude is set at 500 meters, and the second at 1500 meters; these thresholds are determined through correlation analysis between historical meteorological disaster events and altitude. When the average altitude is less than or equal to 500 meters, the emergency warning level is raised by one level; when the average altitude is between 500 and 1500 meters, the emergency warning level is raised by two levels; and when the average altitude exceeds 1500 meters, the emergency warning level is raised by three levels. Altitude data and meteorological data are correlated and matched using spatial coordinates to ensure data consistency.

[0152] Specifically, when the location information of the preset area is identified as mountainous, the adjustment unit initiates the altitude data acquisition process. The average altitude is obtained by calculating the arithmetic mean of all elevation points within the preset area, retaining two decimal places of precision. During the altitude threshold comparison process, an interval division method is used to determine the level upgrade range. For example, in an area with an average altitude of 800 meters, since it is between the first and second altitudes, the emergency warning level is upgraded by two levels. This adjustment mechanism is based on the physical law that the increase in altitude in mountainous areas leads to airflow lifting and intensifies the intensity of severe convective weather, improving the accuracy of the warning by quantifying the impact of altitude. For each increase in altitude gradient, the warning level upgrade range increases progressively, forming a non-linear correlation. The 500-meter and 1500-meter boundary values ​​are verified through regression analysis of disaster loss data from different altitude areas in historical cases. The adjusted warning level is transmitted to the warning release terminal via the communication module, simultaneously triggering the update of the risk label in the corresponding area in the map visualization module.

[0153] As a preferred embodiment, see [reference] Figure 3 As shown, the specific implementation of the solution in this application is as follows: The intelligent question-answering module integrates a speech recognition and natural language processing engine. When a forecaster inputs "query cases of severe convective weather in the eastern coastal area over the past three months" via voice, the module first converts the voice command into text and uses a semantic parsing engine to identify the time range, geographical region, and weather type in the query elements. Then, it calls a semantic matching algorithm to filter historical case sets from the prior historical database that meet the time condition of the past three months, have geographical tags including eastern coastal cities, and are labeled as severe convective weather emergencies. The query results are displayed in a structured table format, along with associated radar echo evolution maps and precipitation intensity change curves.

[0154] The feedback module automatically extracts key parameters such as maximum wind speed, peak precipitation, and warning response time from the sixteen returned matching cases, and generates a decision support report through statistical analysis. The report includes a comparative analysis module that compares the currently collected radar echo intensity with historical data from the same period, generating a risk level trend chart. The decision reference content update module scans newly added historical warning data every five minutes. When similar meteorological characteristics are detected in the same geographical area, it automatically adds the latest response strategy reference entries to the report's suggestion column.

[0155] Through the above technical solutions, this application realizes intelligent support for human-machine collaborative decision-making in the meteorological early warning process, solving the technical problems of low efficiency in manual querying and lagging updates of decision-making basis in traditional early warning systems. The natural language interaction mechanism significantly improves the retrieval efficiency for complex meteorological conditions, the dynamically updated decision suggestions effectively enhance the response speed to sudden weather events, and the multi-dimensional data correlation analysis provides a more comprehensive reference for determining the early warning level.

[0156] The above embodiments improve the accuracy and real-time performance of the meteorological early warning system by introducing a historical prior database, multi-source data fusion, and a combination of deep learning and Kalman filtering. The historical prior database integrates meteorological, radar, and satellite data from many years into a unified data source, facilitating effective comparison and analysis when facing complex meteorological events. The acquisition unit accurately extracts relevant features from meteorological, radar, and satellite data using mutual information, reducing data redundancy and improving data quality. The judgment unit intelligently identifies the meteorological type based on the comparison results, ensuring that the early warning system adopts different data fusion strategies according to severe convective weather or slowly changing risk conditions. The deep learning model focuses on sudden warnings of severe convective weather, enabling rapid response and complex pattern recognition. Kalman filtering effectively handles slowly changing weather, accurately predicting long-term trends and subtle changes. Through this technological combination, efficient and accurate early warnings for complex and variable meteorological phenomena are achieved, improving the intelligence and practicality of meteorological early warning systems.

[0157] In another preferred embodiment based on the above embodiments, see [reference] Figure 4 As shown, this embodiment provides an intelligent duty and early warning method that integrates meteorological operational data, applied to the aforementioned intelligent duty and early warning system that integrates meteorological operational data, including:

[0158] S100: Collect historical prior data, which includes several historical forecast types and several historical feature data. Each historical forecast type corresponds to a historical feature data, which includes historical meteorological data, historical radar data, and historical satellite data.

[0159] S200: Collects meteorological data, radar data, and satellite data of a preset area, and processes the meteorological data, radar data, and satellite data based on the mutual information method to form feature data.

[0160] S300: Collects feature data and compares it with historical prior databases. Based on the comparison results, it determines the forecast weather type, which includes severe convective weather sudden warnings and slow risk warnings.

[0161] S400: When the forecast weather type is a severe convective weather sudden warning, the feature data is fused based on deep learning, and the sudden warning level is determined based on the fusion result. When the forecast weather type is a slow risk warning, the feature data is fused based on Kalman filtering, and the slow risk warning level is determined based on the fusion result.

[0162] Understandably, the accuracy and real-time performance of the meteorological early warning system have been improved by introducing a historical prior database, multi-source data fusion, and a combination of deep learning and Kalman filtering. The historical prior database integrates years of meteorological, radar, and satellite data into a unified data source, facilitating effective comparison and analysis when facing complex meteorological events. The acquisition unit accurately extracts relevant features from meteorological, radar, and satellite data using mutual information, reducing data redundancy and improving data quality. The judgment unit intelligently identifies the weather type based on the comparison results, ensuring that the early warning system adopts different data fusion strategies according to severe convective weather or slowly changing risk scenarios. The deep learning model focuses on sudden warnings of severe convective weather, enabling rapid response and complex pattern recognition. Kalman filtering effectively handles slowly changing weather, accurately predicting long-term trends and subtle changes. Through this technological combination, efficient and accurate early warnings for complex and variable meteorological phenomena have been achieved, improving the intelligence and practicality of meteorological early warning systems.

[0163] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program goods. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0164] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program goods according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0165] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0166] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0167] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. An intelligent watch-keeping early warning system that fuses meteorological operational data, characterized in that, include: The historical prior database includes several historical forecast types and several historical feature data, with each historical forecast type corresponding to a historical feature data, which includes historical meteorological data, historical radar data, and historical satellite data. The acquisition unit is used to acquire meteorological data, radar data, and satellite data of a preset area, and to process the meteorological data, radar data, and satellite data based on the mutual information method to form feature data. The judgment unit is configured to collect the feature data and compare the feature data with the historical prior database, and determine the forecast weather type based on the comparison result. The forecast weather type includes severe convective weather sudden warning and slow risk warning. The processing unit is configured to, when the forecast weather type is a severe convective weather sudden warning, perform data fusion on the feature data based on deep learning and determine the sudden warning level based on the fusion result; and when the forecast weather type is a slow risk warning, perform data fusion on the feature data based on Kalman filtering and determine the slow risk warning level based on the fusion result. When the judgment unit determines the forecast weather type based on the comparison results, it includes: The neighborhood radius is determined based on the radar echo range, and MinPts is initialized to 4. Based on DBSCAN clustering, the feature data is clustered with historical feature data in the historical prior database to determine the cluster to which the feature data belongs. If there is a historical forecast type corresponding to historical feature data in the cluster to which the feature data belongs that is a historical severe convective weather sudden warning, the forecast meteorological type of the feature data is determined to be the severe convective weather sudden warning; if the historical forecast type corresponding to historical feature data in the cluster to which the feature data belongs is a historical slow risk warning, the forecast meteorological type of the feature data is determined to be the slow risk warning. When the forecast weather type is a severe convective weather warning, the processing unit performs data fusion on the feature data based on deep learning, including: The processing unit inputs the feature data into the deep learning model; Convolutional neural network layer: used to extract spatial features from radar and satellite data, including precipitation intensity, cloud morphology, and radar echoes; Long Short-Term Memory (LSTM) network layer: used to process the temporal features in meteorological data and capture the temporal trends of temperature, humidity, and wind speed; Fusion layer: The outputs of the CNN layer and the LSTM layer are fused by weighted averaging to generate comprehensive features; Fully connected layer: Classifies the fused integrated features and outputs a risk coefficient for sudden strong convection warning; Output layer: Generates the final probability value and risk coefficient for severe convective weather emergencies; When the forecast weather type is a slow risk warning, the processing unit performs data fusion on the feature data based on Kalman filtering, including: The processing unit predicts the current weather conditions based on the historical prior database and the system dynamics model. By comparing the feature data with the prediction results, adjusting the prediction results according to the Kalman gain, and updating the weather state estimate; Use meteorological models to describe the changes in meteorological data and update the current meteorological status. Based on the covariance of prediction error and observation error, the Kalman gain is calculated, and the weighted average of the feature data and the prediction data is used as the final meteorological state estimate. A slow risk warning coefficient is generated based on the output of the Kalman filter.

2. The intelligent watch warning system fusing meteorological operational data according to claim 1, characterized in that, When the acquisition unit processes the meteorological data, radar data, and satellite data based on the mutual information method to form feature data, it includes: The meteorological data includes temperature, humidity, air pressure, wind speed, and precipitation. The radar data includes radar echo maps; The satellite data includes cloud images, precipitation maps, and temperature distribution maps; The meteorological data, radar data, and satellite data are preprocessed, including outlier removal and interpolation to fill in missing data, and spatial resolution alignment is performed on the meteorological data, radar data, and satellite data. Convert the preprocessed meteorological data, radar data, and satellite data into discrete form; The mutual information between meteorological data, radar data, and satellite data is calculated separately, and the data with the highest mutual information value among the same type of meteorological data is retained to form the feature data.

3. The intelligent watch warning system that fuses weather business data according to claim 2, characterized in that, When the processing unit determines the sudden warning level or the slow risk warning level based on the fusion result, it includes: The processing unit compares the strong convection sudden warning risk coefficient with the strong convection sudden warning risk threshold, or compares the slow risk warning coefficient with the slow risk warning threshold, and determines the sudden warning level or the slow risk warning level based on the comparison result. The level of the emergency warning is directly proportional to the difference between the risk coefficient of the emergency warning for severe convective weather and the risk threshold for the emergency warning for severe convective weather. The slow risk warning level is directly proportional to the difference between the slow risk warning coefficient and the slow risk warning threshold.

4. The intelligent watch warning system fusing meteorological operational data according to claim 3, characterized in that, Also includes: The adjustment unit is configured to, after determining the emergency warning level, collect the location information of the preset area and determine whether to adjust the emergency warning level based on the location information; When the location information is in a mountainous area, the adjustment unit determines to adjust the emergency warning level; when the location information is not in a mountainous area, the adjustment unit determines not to adjust the emergency warning level.

5. The intelligent watch warning system fusing meteorological operational data according to claim 4, characterized in that, When the adjustment unit determines to adjust the emergency warning level, it includes: The adjustment unit obtains the average altitude of the preset area, compares the average altitude with a first altitude and a second altitude respectively, and adjusts the emergency warning level according to the comparison results; the first altitude is less than the second altitude; When the average altitude is less than or equal to the first altitude, the adjustment unit raises the emergency warning level by one level; when the average altitude is greater than the first altitude and less than or equal to the second altitude, the adjustment unit raises the emergency warning level by two levels; when the average altitude is greater than the second altitude, the adjustment unit raises the emergency warning level by three levels.

6. The intelligent watch warning system fusing weather business data according to claim 5, characterized in that, Also includes: The intelligent question-answering module is used to support forecasters to interact via natural language and quickly query the historical prior database; The feedback module provides suggestions based on forecasters' query results, generates decision support reports, and automatically updates decision reference content based on historical prior data and the current warning status.

7. An intelligent duty officer warning method of fusing meteorological service data, applied to the intelligent duty officer warning system of fusing meteorological service data according to any one of claims 1-6, characterized in that, include: Collect historical prior data, which includes several historical forecast types and several historical feature data, and each historical forecast type corresponds to a historical feature data, which includes historical meteorological data, historical radar data and historical satellite data; Meteorological data, radar data, and satellite data of a preset area are collected, and feature data are formed by processing the meteorological data, radar data, and satellite data based on the mutual information method. The feature data is collected and compared with the historical prior database. The forecast weather type is determined based on the comparison result. The forecast weather type includes severe convective weather sudden warning and slow risk warning. When the forecast weather type is a severe convective weather emergency warning, the feature data is fused based on deep learning, and the emergency warning level is determined according to the fusion result; When the forecast weather type is a slow risk warning, the feature data is fused based on Kalman filtering, and the slow risk warning level is determined according to the fusion result.

Images