Meteorological AI abnormity early warning method and system based on extreme weather risk perception

Abstract

The invention relates to the technical field of meteorological early warning, and discloses a meteorological AI abnormity early warning method and system for extreme weather risk perception, and the method comprises the steps: collecting and preprocessing multi-source meteorological data, and obtaining standardized multi-source meteorological data; feature extraction and anomaly recognition are carried out on the standardized multi-source meteorological data in the continuous time period, and preliminary risk information and a first confidence value are obtained; and constructing a space-time diagram model and applying consistency constraint in the training and reasoning process of the space-time diagram model to obtain a risk identification result. And according to the risk identification result, determining a dynamic alarm threshold value to generate graded meteorological anomaly alarm information, and adjusting the sampling frequency. And sending the meteorological abnormity alarm information and the sampling adjustment instruction to a cloud platform and a terminal display device, and carrying out risk perception and early warning release of extreme weather. According to the invention, high-credibility, low-time-delay and full-period dynamic perception and hierarchical response of extreme weather risks are realized.

Description

Technical Field

[0001] This invention relates to the field of meteorological early warning technology, and more specifically, to a meteorological AI-based method and system for extreme weather risk perception and early warning. Background Technology

[0002] The frequent occurrence of extreme weather events has placed higher demands on the real-time and intelligent capabilities of disaster prevention and mitigation systems. Traditional meteorological early warning systems mainly rely on fixed monitoring stations and periodic data reporting mechanisms, which are insufficient to meet the needs for rapid perception and dynamic dissemination of information regarding regional emergencies (such as short-duration heavy rainfall, tornadoes, and hail). With the integrated development of edge computing, artificial intelligence, and the Internet of Things, meteorological monitoring is gradually evolving from single-point observation to multi-source fusion and intelligent identification.

[0003] While existing technologies have incorporated 5G communication and IoT architecture to enhance the communication and dissemination capabilities of local weather warnings, some problems remain. For example, patent CN116071892A, "A Method for Publishing Local Disastrous Weather Warnings via IoT Based on 5G+IoT," proposes a rapid warning system with cloud-edge-device collaboration, enabling the identification of severe weather within 10 seconds and real-time LED dissemination. However, it focuses on optimizing information transmission latency and lacks a reliability assessment and anomaly correction mechanism for the quality of multi-source meteorological data. Edge identification results are not corrected for global spatiotemporal consistency, resulting in limited accuracy in risk identification. CN116721517A, "A Meteorological Data Monitoring and Warning Method and Device Based on Tianjing," improves the flexibility of the monitoring cycle through temporary frequency upsampling and downsampling mechanisms to enhance warning timeliness without increasing system load. However, this solution still relies on preset cycle triggering and does not achieve dynamic sampling scheduling based on risk scores and confidence values, nor does it form a closed loop of cloud-edge collaborative AI risk identification and hierarchical alarm.

[0004] Therefore, it is necessary to design a meteorological AI-based method and system for extreme weather risk perception to address the problems existing in current technologies. Summary of the Invention

[0005] In view of this, the present invention proposes a meteorological AI anomaly early warning method and system for extreme weather risk perception, aiming to solve the problems of insufficient accuracy and timeliness of early warning results caused by the lack of a reliable multi-source data correction mechanism, collaborative recognition capability of edge and cloud spatiotemporal graph models, and dynamic early warning control mechanism in the existing technology.

[0006] In one aspect, this invention proposes a meteorological AI-based method for extreme weather risk perception and early warning, comprising: Multi-source meteorological data is collected and preprocessed to obtain standardized multi-source meteorological data. The preprocessing includes normalization and screening. The screening includes calculating a confidence score based on the data quality, proximity consistency and historical stability of each monitoring node. When the confidence score is lower than the confidence score threshold, data correction or node replacement is performed. The multi-source meteorological data includes ground monitoring station data, radar observation data, satellite remote sensing data and video surveillance data. The space-time anomaly detection model is run in the edge computing node to extract features and identify anomalies in standardized multi-source meteorological data over a continuous time period, thereby obtaining preliminary risk information and a first confidence value. The preliminary risk information is uploaded to the cloud, a spatiotemporal graph model is constructed, and consistency constraints are applied during the training and inference of the spatiotemporal graph model to obtain risk identification results. The consistency constraints are used in the training as model constraint terms and are used to correct the risk identification results during the inference stage. The spatiotemporal graph model includes meteorological monitoring nodes and their spatial relationships. The consistency constraints include meteorological energy conservation constraints, water vapor balance constraints, and cross-modal consistency constraints. The risk identification results include a risk score and a second confidence value. A dynamic alarm threshold is determined based on the risk identification results. When the risk score exceeds the dynamic alarm threshold and the comprehensive confidence score meets the comprehensive confidence score threshold, a graded meteorological anomaly alarm is generated. The sampling frequency of the monitoring nodes is adjusted based on the risk score, the second confidence value, and the comprehensive confidence score. The comprehensive confidence score is obtained by weighting the confidence scores of the monitoring nodes covered by the alarm object. The meteorological anomaly alarm information and sampling adjustment instructions are sent to the cloud platform and terminal display device to conduct risk perception and early warning of extreme weather.

[0007] Furthermore, the collection and preprocessing of multi-source meteorological data includes: Raw meteorological data were collected from ground monitoring stations, radar observations, satellite remote sensing, and video monitoring terminals, respectively. Data from different sources, with different resolutions and different sampling periods are formatted and normalized in a unified manner so that physical quantities can be compared on a unified numerical scale. These physical quantities include temperature, humidity, wind speed, rainfall and reflectivity. The credibility score of each monitoring node is calculated. The credibility score is determined based on the data integrity of the monitoring node, the consistency of the observation results of adjacent nodes, and the historical fluctuation stability of the node. When the credibility score is lower than the preset threshold, the abnormal data is corrected by interpolation substitution of the neighboring high credibility nodes or by backtracking reconstruction of the historical stable interval. If the node is continuously lower than the credibility score threshold, the corresponding monitoring task is taken over by the backup node. The corrected multi-source meteorological data is the standardized multi-source meteorological data.

[0008] Furthermore, when running a spacetime anomaly detection model in an edge computing node, it includes: The standardized multi-source meteorological data is divided into time sliding windows and spatial grids, and time alignment and spatial registration are performed on the standardized multi-source meteorological data within the same window; The motion change characteristics of the video surveillance data, the echo morphology characteristics of the radar observation data, and the cloud structure characteristics of the satellite remote sensing data are obtained respectively. The features from different sources are weighted and fused according to the credibility score of the monitoring node. During the fusion process, occlusion detection, low-light enhancement, and echo interference removal are performed. The fusion features are judged to obtain preliminary risk information including anomaly type and anomaly intensity, and the first confidence value is generated based on the score output by the spatiotemporal anomaly detection model, combined with the consistency of neighboring nodes and the confidence score of monitoring nodes.

[0009] Furthermore, the construction of the spatiotemporal graph model includes: Using monitoring nodes as graph nodes and the geographical proximity, wind direction propagation path and topographic influence between nodes as edges, a dynamic spatial relationship is formed. During the model training phase, the connection strength between nodes is periodically updated based on the changes in edge weights, so that the spatial relationship can be adjusted according to changes in airflow direction and geographical conditions.

[0010] Furthermore, the consistency constraints include: The meteorological energy conservation constraint and water vapor balance constraint, during the model training phase, jointly constrain the predicted results of temperature, air pressure, humidity, and precipitation in the spatiotemporal map model by introducing physical boundary conditions of energy and humidity. This joint constraint includes checking the balance of energy budget and water vapor flux in the predicted results generated in each iteration of the spatiotemporal map model and feeding back deviation information to adjust the weights of the spatiotemporal map model. During the inference phase, when a discrepancy is detected between the predicted results and the physical boundary conditions, the weight allocation of the corresponding region in the risk score is corrected. The cross-modal consistency constraint, during the model training phase, performs temporal alignment and spatial registration of the echo distribution of radar observation data, cloud morphology of satellite remote sensing data, and real-time meteorological parameters from ground monitoring station data, and extracts common features between different modes based on a feature mapping network. During the inference phase, when any mode of data shows drift or is missing, the weight of that mode in the risk identification result is adjusted based on the high confidence results of the remaining mode data.

[0011] Furthermore, when obtaining risk identification results, the following are included: The edge recognition results and the spatiotemporal graph model output results are received. Based on the comprehensive credibility score, the node-level recognition results and the region-level recognition results are weighted and fused to obtain the candidate risk score. The candidate risk scores are tested for deviation based on the meteorological energy conservation constraints and water vapor balance constraints. When the deviation exceeds the allowable range, the risk scores of the corresponding areas are adjusted so that the risk scores are consistent with the changing trends of meteorological elements. Comparisons were made between different data modalities, and data modalities that were drifting or missing were corrected by combining radar observation data, satellite remote sensing data and ground monitoring station data. The risk changes of adjacent monitoring nodes in space are examined. When the risk changes of adjacent nodes show a continuous or synchronous trend, the second confidence value of the risk identification result is increased, and the risk identification result is finally obtained.

[0012] Furthermore, when determining the dynamic alarm threshold, the following factors are considered: Historical risk scores are archived according to monitoring area and season, and reference thresholds are extracted based on similar weather events and similar underlying surface conditions. The reference threshold is adjusted based on the second confidence value and the comprehensive reliability score, such that the trigger threshold is increased when the second confidence value is lower than the second confidence value threshold, and decreased when the second confidence value is higher than the second confidence value threshold. The risk score over a continuous time period is continuously checked and hysteresis controlled. When the risk score continuously exceeds the dynamic alarm threshold and the duration exceeds the duration threshold, the graded meteorological anomaly alarm information is generated. The alarm level is determined based on the extent and duration of the risk score exceeding the limit. Simultaneously, adjacent monitoring nodes within the spatial correlation are merged to avoid duplicate alarms.

[0013] Furthermore, when adjusting the sampling frequency of the monitoring nodes based on the risk score, the second confidence value, and the comprehensive confidence score, the following steps are included: The sampling frequency of the monitoring node is adjusted based on the combined result of the risk score, the second confidence value, and the comprehensive confidence score. When the risk score exceeds the dynamic alarm threshold and the second confidence value is less than or equal to the second confidence value threshold, the sampling frequency of the corresponding monitoring node is set to a high-frequency sampling value. When the risk score is lower than the dynamic alarm threshold and the second confidence value is higher than the second confidence value threshold, the sampling frequency of the corresponding monitoring node is set to a low-frequency sampling value. When the risk score is below the dynamic alarm threshold and remains stable within a continuous fixed time window, the sampling frequency of the monitoring node will be restored to the normal sampling value. The sampling frequency of monitoring nodes in the shared observation area within the spatial correlation remains consistent.

[0014] Furthermore, when sending the meteorological anomaly alarm information and sampling adjustment instructions to the cloud platform and terminal display device, the process includes: generating an alarm load containing alarm level, alarm object, spatial correlation, validity period, sampling adjustment parameters and timestamp, and assigning a unique alarm identifier to each alarm load; Alarms with the same level within the same spatial relationship and their validity period are merged to form a single alarm load and a single sampling adjustment command; The sending order is established according to the alarm level from high to low. If a receipt is not received within the confirmation time limit, the transmission is resent. If the resentment is still unsuccessful, the transmission is switched to the backup transmission channel. The alarm load is displayed in a tiered manner and accompanied by audio-visual prompts on the terminal display device side.

[0015] Compared with existing technologies, the beneficial effects of this invention are as follows: By constructing a meteorological AI early warning system based on edge recognition, cloud-based spatiotemporal reasoning, and dynamic alarm control, it achieves highly reliable, low-latency, full-cycle dynamic perception and graded response to extreme weather risks; by introducing a reliable scoring mechanism based on data quality, proximity consistency, and historical stability, it adaptively corrects and replaces nodes in multi-source meteorological data, improving the reliability of input data; by deploying a spatiotemporal anomaly detection model at the edge, it can achieve local real-time feature extraction and preliminary risk identification, shortening the early warning response time; by integrating spatiotemporal graph models with physical and semantic constraints such as meteorological energy conservation, water vapor balance, and cross-modal consistency, it achieves physical consistency and global correction of risk identification results in the cloud, improving the accuracy and stability of early warnings; based on the dynamic threshold determination and sampling frequency control mechanism of risk scoring, second confidence value, and comprehensive reliable scoring, it can adjust the monitoring density and alarm level according to the risk level, ultimately realizing intelligent, collaborative, and graded dynamic early warning of extreme weather events, improving the real-time and intelligent response capabilities in the field of meteorological disaster monitoring.

[0016] On the other hand, this application also provides a meteorological AI anomaly early warning system for extreme weather risk perception, used to apply the above-mentioned meteorological AI anomaly early warning method for extreme weather risk perception, including: The acquisition unit is configured to acquire multi-source meteorological data and perform preprocessing to obtain standardized multi-source meteorological data. The preprocessing includes normalization processing and filtering processing. The identification unit is configured to run a space-time anomaly detection model in an edge computing node to perform feature extraction and anomaly identification on standardized multi-source meteorological data within a continuous time period, and obtain preliminary risk information and a first confidence value. The prediction unit is configured to upload the preliminary risk information to the cloud, construct a spatiotemporal graph model, and apply consistency constraints during the training and inference process of the spatiotemporal graph model to obtain risk identification results; The judgment and adjustment unit is configured to determine a dynamic alarm threshold based on the risk identification result. When the risk score exceeds the dynamic alarm threshold and the comprehensive confidence score meets the comprehensive confidence score threshold, a graded meteorological anomaly alarm message is generated, and the sampling frequency of the monitoring node is adjusted according to the risk score, the second confidence value and the comprehensive confidence score. The early warning unit is configured to send the meteorological anomaly alarm information and sampling adjustment instructions to the cloud platform and terminal display device to conduct risk perception and early warning issuance for extreme weather.

[0017] It is understandable that the above-mentioned meteorological AI anomaly early warning methods and systems for extreme weather risk perception have the same beneficial effects, and will not be elaborated further here. Attached Figure Description

[0018] 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: Figure 1 A flowchart of a meteorological AI anomaly early warning method for extreme weather risk perception provided in an embodiment of the present invention; Figure 2 A functional block diagram of a meteorological AI anomaly early warning system for extreme weather risk perception provided in an embodiment of the present invention. Detailed Implementation

[0019] 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.

[0020] Traditional meteorological early warning systems lack quality assessment and anomaly correction mechanisms for multi-source meteorological data, making it difficult to guarantee the reliability of data from ground monitoring stations, radar observations, satellite remote sensing, and video surveillance. The preliminary identification results of spatiotemporal anomalies by edge computing nodes lack global spatiotemporal consistency constraints, failing to eliminate the risk of misjudgment caused by local data drift or node failure. Furthermore, the adjustment of dynamic alarm thresholds and monitoring node sampling frequencies is not combined with joint analysis of risk scores and confidence values, resulting in unreasonable resource allocation and limited early warning timeliness.

[0021] For example, in urban short-term heavy rainfall warning scenarios, ground monitoring stations experience abnormal temperature data fluctuations due to equipment aging, while nearby radar observations suffer from echo distortion due to electromagnetic interference, and video surveillance exhibits motion artifacts due to low-light environments. Edge computing nodes, extracting preliminary risk information based on a single time window, fail to consider the physical constraints of meteorological energy conservation and water vapor balance, mistakenly identifying local data anomalies as heavy rainfall risks. The cloud platform uses fixed alarm thresholds and uniform sampling frequencies, failing to dynamically adjust the data acquisition strategies of monitoring nodes according to the spatial distribution of risk scores, resulting in insufficient data sampling in high-value areas and resource overload in low-risk areas.

[0022] If the above issues are not addressed, the lack of anomaly correction mechanisms will reduce the reliability of input data to the spatiotemporal graph model, thereby affecting the accuracy of risk identification results. The lack of cross-modal consistency constraints in the edge and cloud collaborative analysis process may amplify local data errors, leading to false alarms or missed alarms. The decoupling of dynamic sampling strategies from risk scoring will prevent monitoring resources from accurately matching actual risk changes, reducing the timeliness of extreme weather warnings and the overall system efficiency.

[0023] For this, please refer to Figure 1 As shown, this application proposes a meteorological AI-based method for extreme weather risk perception and early warning, including: S100: Collects and preprocesses multi-source meteorological data to obtain standardized multi-source meteorological data. Preprocessing includes normalization and screening. Screening includes calculating a reliability score based on the data quality, proximity consistency, and historical stability of each monitoring node. When the reliability score is lower than the reliability score threshold, data correction or node replacement is performed. Multi-source meteorological data includes ground monitoring station data, radar observation data, satellite remote sensing data, and video surveillance data.

[0024] S200: Runs a spacetime anomaly detection model in an edge computing node to extract features and identify anomalies from standardized multi-source meteorological data over a continuous time period, obtaining preliminary risk information and a first confidence value.

[0025] S300: Upload preliminary risk information to the cloud, construct a spatiotemporal graph model, and apply consistency constraints during the training and inference of the spatiotemporal graph model to obtain risk identification results. Among them, consistency constraints participate in training in the form of model constraint terms and are used to correct the risk identification results during the inference stage. The spatiotemporal graph model includes meteorological monitoring nodes and their spatial relationships. Consistency constraints include meteorological energy conservation constraints, water vapor balance constraints, and cross-modal consistency constraints. The risk identification results include risk scores and second confidence values.

[0026] S400: Determine the dynamic alarm threshold based on the risk identification results. When the risk score exceeds the dynamic alarm threshold and the comprehensive confidence score meets the comprehensive confidence score threshold, generate a graded meteorological anomaly alarm information. Adjust the sampling frequency of the monitoring nodes based on the risk score, the second confidence value, and the comprehensive confidence score. The comprehensive confidence score is obtained by weighting the confidence scores of the monitoring nodes covered by the alarm object.

[0027] S500: Sends meteorological anomaly alarm information and sampling adjustment instructions to the cloud platform and terminal display device to conduct risk perception and early warning of extreme weather.

[0028] Specifically, multi-source meteorological data includes data from ground monitoring stations, radar observations, satellite remote sensing data, and video surveillance data. This can be achieved using multi-sensor fusion technology, integrating observation data with different spatial resolutions and temporal frequencies to improve the comprehensiveness and redundancy of meteorological anomaly detection. The screening process involves calculating a reliability score based on the data quality, proximity consistency, and historical stability of each monitoring node. This can be achieved using data quality assessment algorithms and interpolation substitution strategies, dynamically correcting or replacing low-reliability node data to ensure the reliability of the input data. The spatiotemporal anomaly detection model runs on edge computing nodes, specifically using convolutional neural networks combined with long short-term memory networks. By extracting anomaly patterns from time series and spatial distributions, it reduces cloud computing load and improves real-time performance. The spatiotemporal graph model is built in the cloud and applies consistency constraints. This can be achieved using graph neural networks combined with physical constraints. By introducing meteorological energy conservation, water vapor balance, and cross-modal consistency constraints, it improves the physical rationality and regional coordination of risk identification results. The dynamic alarm threshold is determined based on the risk identification results, specifically using an adaptive threshold algorithm based on historical data. This dynamically adjusts the alarm triggering conditions by combining the current risk score and the comprehensive credibility score, avoiding false alarms or missed alarms caused by local data anomalies. The comprehensive credibility score is obtained by weighting the credibility scores of the monitoring nodes covered by the alarm object, specifically using a spatial weighted average algorithm. This prioritizes data from high-credibility nodes, improving the overall credibility of alarm decisions. The sampling frequency adjustment of monitoring nodes is based on the risk score, the second confidence value, and the comprehensive credibility score, specifically using a feedback control algorithm. This dynamically adjusts the monitoring density of different risk areas, optimizing resource allocation and improving the timeliness of early warnings.

[0029] This application constructs a cloud-edge collaborative meteorological anomaly early warning closed-loop system. It realizes a collaborative mechanism of preliminary anomaly detection and global correction of cloud spatiotemporal map model through edge computing nodes. Combining multi-source data credibility assessment, physical constraint-driven risk identification and dynamic threshold alarm strategy, it solves the problem of low early warning accuracy caused by data quality defects, local identification bias and fixed alarm threshold in traditional methods.

[0030] The working process and principle of this application are as follows: First, multi-source meteorological data is collected and preprocessed to obtain standardized multi-source meteorological data. Preprocessing includes normalization and screening. Screening assesses data quality by calculating the reliability score of each monitoring node; when the reliability score falls below a threshold, data correction or node replacement is performed. Multi-source meteorological data includes data from ground monitoring stations, radar observations, satellite remote sensing, and video surveillance.

[0031] Then, a space-time anomaly detection model is run on the edge computing node to extract features and identify anomalies in standardized multi-source meteorological data over a continuous time period, thereby obtaining preliminary risk information and a first confidence value.

[0032] Next, preliminary risk information is uploaded to the cloud, a spatiotemporal graph model is constructed, and consistency constraints are applied during training and inference to obtain risk identification results. Consistency constraints participate in training as model constraint terms and are used to correct the risk identification results during the inference phase. The spatiotemporal graph model includes meteorological monitoring nodes and their spatial relationships. Consistency constraints include meteorological energy conservation constraints, water vapor balance constraints, and cross-modal consistency constraints. Risk identification results include risk scores and second confidence values.

[0033] Based on the risk identification results, a dynamic alarm threshold is determined. When the risk score exceeds the dynamic alarm threshold and the comprehensive confidence score meets the threshold, a tiered meteorological anomaly alarm is generated. Simultaneously, the sampling frequency of the monitoring nodes is adjusted based on the risk score, the second confidence value, and the comprehensive confidence score. The comprehensive confidence score is obtained by weighting the confidence scores of the monitoring nodes covering the alarm object.

[0034] Finally, meteorological anomaly warning information and sampling adjustment instructions are sent to the cloud platform and terminal display device to conduct risk perception and early warning of extreme weather.

[0035] This solution improves data reliability through multi-level screening and correction mechanisms, achieves rapid initial identification using edge computing, and ensures the accuracy of risk identification by combining cloud-based global analysis. Dynamically adjusting alarm thresholds and sampling frequency enhances flexibility.

[0036] As a preferred embodiment, the solution of this application is specifically implemented as follows: When collecting multi-source meteorological data, conventional meteorological parameters such as temperature, humidity, and wind speed are obtained from ground monitoring stations; echo data such as reflectivity and radial velocity are obtained from radar; remote sensing data such as cloud top temperature and cloud structure are obtained from meteorological satellites; and real-time image sequences are obtained from video surveillance. These heterogeneous data are then normalized to unify different physical quantities into a 0-1 value range.

[0037] When calculating the reliability score of a monitoring node, factors such as data integrity, correlation with neighboring nodes, and stability of historical data are considered. For example, if the temperature data of a ground station changes abruptly and differs significantly from that of surrounding stations, its reliability score will decrease. When the score is below 0.7, it is corrected through interpolation or reconstruction of historical data.

[0038] Edge computing nodes employ lightweight convolutional neural networks to extract spatiotemporal features, combined with decision trees for anomaly classification. For video data, optical flow features are extracted to characterize cloud movement. For radar data, echo morphology features are extracted. For satellite data, cloud top temperature gradient features are extracted. After fusing multi-source features, preliminary risk information and confidence levels are output.

[0039] The spatiotemporal graph model constructed in the cloud employs a graph neural network structure, where nodes represent monitoring stations and edges represent the geographical relationships between stations. During training, energy conservation and water vapor balance are introduced as physical constraints. For example, the temperature and humidity changes of adjacent nodes are required to satisfy the heat and water vapor transport equations. Cross-modal consistency constraints are achieved through a feature mapping network to align different data sources.

[0040] The dynamic alarm threshold is determined based on historical data from the same period and the current risk score. The sampling frequency adjustment strategy is as follows: when the risk score exceeds the threshold but the confidence level is low, the sampling frequency is increased to obtain more data. When the risk is low and the confidence level is high, the sampling frequency is decreased to save resources.

[0041] The final alarm information includes elements such as risk level, scope of impact, and duration. It is pushed to relevant department terminals through the cloud platform and published on public facilities such as LED displays.

[0042] Through the above-described scheme, this application achieves quality assessment and anomaly correction of multi-source heterogeneous meteorological data, improving the reliability of input data. The combination of edge computing and cloud-based collaborative analysis ensures both the real-time nature of initial identification and enhances the accuracy of risk assessment through global constraints. Dynamically adjusting alarm thresholds and sampling strategies enables monitoring resources to more accurately match actual risk changes, improving the timeliness of extreme weather warnings. In scenarios of sudden meteorological disasters such as short-duration heavy rainfall in cities, this scheme can quickly identify local anomalies and reduce false alarm rates through cross-validation with multimodal data.

[0043] In some of the schemes mentioned above in this application, when collecting and preprocessing multi-source meteorological data, the differences in format, resolution and sampling period of data from different sources make it difficult to compare physical quantities. At the same time, the lack of a dynamic evaluation and correction mechanism for the data quality of monitoring nodes may introduce abnormal data into the risk identification process.

[0044] This application further proposes a method for collecting and preprocessing multi-source meteorological data, including: collecting raw meteorological data from ground monitoring stations, radar observations, satellite remote sensing, and video monitoring terminals. Data from different sources, resolutions, and sampling periods are uniformly formatted and normalized to ensure that physical quantities, including temperature, humidity, wind speed, rainfall, and reflectivity, are compared on a unified numerical scale. A reliability score is calculated for each monitoring node, determined based on data integrity, consistency of observations from adjacent nodes, and the node's historical stability. When the reliability score falls below a preset threshold, abnormal data is corrected through interpolation from nearby high-reliability nodes or by backtracking and reconstructing historical stable intervals. If a node consistently falls below the reliability score threshold, a backup node takes over the corresponding monitoring task. The corrected multi-source meteorological data is denoted as standardized multi-source meteorological data.

[0045] The unified formatting and normalization process employs data transformation algorithms to map data from different timestamps, spatial coordinate systems, and units to a unified spatiotemporal reference. For example, the raster resolution of satellite remote sensing data is resampled to the same spatial grid as that of ground monitoring station data. During the reliability score calculation, data integrity is quantified by the missing value rate and the proportion of outliers; consistency between adjacent nodes is measured by Euclidean distance or Pearson correlation coefficient; and historical fluctuation stability is assessed by the standard deviation or coefficient of variation within a sliding window. When correcting outlier data, interpolation substitution uses Kriging interpolation or inverse distance weighting to generate replacement values, and the backtracking reconstruction of historical stable intervals involves extracting trend and periodic terms through time series decomposition for data filling.

[0046] Specifically, after the raw meteorological data is collected, it undergoes format conversion and time synchronization, for example, aligning the frame rate of video surveillance data with the scanning period of radar observation data. Normalization is performed using maximum / minimum scaling or Z-score standardization to ensure that physical quantities such as temperature and humidity are within the same numerical range. The reliable scoring module periodically updates the scores of each node, triggering a correction process when the score falls below a threshold: if the data missing rate exceeds the threshold, neighboring nodes are used for interpolation as a substitute. If data fluctuations exceed the historical stable range, reasonable data is reconstructed based on an autoregressive model or a long short-term memory network. A backup node takeover mechanism is activated when multiple consecutive scoring cycles fail to meet the target, allocating monitoring tasks through a load balancing algorithm. The corrected data is verified and stored in a standardized database for use by the spatiotemporal anomaly detection model. This process, through dynamic quality assessment and multi-strategy correction, ensures the reliability and consistency of the input data, reducing the impact of anomalous data on the risk identification accuracy of edge computing nodes.

[0047] As a preferred embodiment, the solution of this application is specifically implemented as follows: The process of collecting and preprocessing multi-source meteorological data includes the following steps: First, raw meteorological data were collected from ground monitoring stations, radar observations, satellite remote sensing, and video surveillance terminals. Ground monitoring station data included parameters such as temperature, humidity, air pressure, wind speed, and wind direction, sampled every 10 minutes. Radar observation data included reflectivity and radial velocity, with a scanning cycle of 6 minutes. Satellite remote sensing data included cloud top brightness temperature and cloud structure, with a temporal resolution of 15 minutes. Video surveillance data was acquired at a frequency of 30 frames per second.

[0048] Secondly, data from different sources, with different resolutions, and different sampling periods were uniformly formatted and normalized. Temperature data was converted to degrees Celsius, humidity to relative humidity percentage, wind speed to meters per second, rainfall to millimeters per hour, and radar reflectivity to dBZ. Each physical quantity was normalized to a value range of 0-1.

[0049] Further, a reliability score is calculated for each monitoring node. The reliability score consists of three parts: data integrity (weight 0.4), neighbor consistency (weight 0.3), and historical stability (weight 0.3). Data integrity is calculated based on the proportion of missing data, neighbor consistency is calculated based on the deviation of observations from the five neighboring nodes, and historical stability is calculated based on the standard deviation of data from the past 24 hours. The final reliability score is obtained by weighted summing of the three scores, with a value ranging from 0 to 100.

[0050] When the credibility score falls below the preset threshold of 80 points, abnormal data is corrected in the following ways: For missing data, interpolation is used to replace it with the scores of the five nearest high-credibility nodes (credibility score_90 points); for abnormally fluctuating data, the median of the node's stable interval over the past 24 hours (standard deviation_10% of the historical average) is used for backtracking and reconstruction. If a node's score is below 80 points for three consecutive times, the corresponding monitoring task is taken over by a backup node.

[0051] Finally, the corrected multi-source meteorological data is recorded as standardized multi-source meteorological data.

[0052] Through the above technical solutions, this application achieves unified preprocessing of multi-source heterogeneous meteorological data. A reliable scoring mechanism identifies and corrects abnormal data, improving data quality. Interpolation substitution and historical reconstruction methods ensure data continuity and consistency. Dynamic adjustment of monitoring nodes further enhances the reliability of data acquisition.

[0053] In some of the solutions mentioned above in this application, multi-source meteorological data suffers from insufficient spatiotemporal alignment accuracy during feature extraction. Feature fusion of data from different sources does not consider the differences in credibility and is prone to introducing noise under complex environmental interference, leading to a decrease in the reliability of anomaly detection results.

[0054] This application further proposes dividing standardized multi-source meteorological data into temporal sliding windows and spatial grids, and performing temporal alignment and spatial registration of the standardized multi-source meteorological data within the same window. Motion change characteristics from video surveillance data, echo morphology characteristics from radar observation data, and cloud structure characteristics from satellite remote sensing data are acquired separately. These features from different sources are then weighted and fused according to the confidence scores of monitoring nodes. During the fusion process, occlusion detection, low-light enhancement, and echo interference removal are performed. The fused features are then discriminated to obtain preliminary risk information including anomaly type and intensity. Finally, a first confidence value is generated based on the score output by the spatiotemporal anomaly detection model, combined with the consistency of neighboring nodes and the confidence scores of monitoring nodes.

[0055] The time sliding window is set to a fixed-length time period, and the spatial grid is divided into grids based on geographic coordinates to ensure that the data within the same window has temporal continuity and spatial proximity. Time alignment is compensated for the sampling time difference of data from different sources through interpolation algorithms, and spatial registration is achieved by geographic coordinate transformation to eliminate projection differences. Motion change characteristics of video surveillance data are extracted using optical flow methods, radar echo morphology characteristics are based on reflectivity gradient analysis, and satellite cloud structure characteristics are identified using convolutional neural networks. In the weighted fusion process, monitoring nodes with high confidence scores are assigned higher weights, occlusion detection is achieved through multi-view data complementarity, low-light enhancement uses a histogram equalization algorithm, and echo interference removal is based on Doppler velocity consistency checks.

[0056] Specifically, standardized multi-source meteorological data is first segmented into combined units of time windows and spatial grids. Data within each unit undergoes temporal interpolation and coordinate transformation to achieve spatiotemporal alignment, eliminating biases caused by differences in sampling frequencies and spatial resolutions across different devices. Video surveillance data is analyzed using optical flow to calculate pixel displacement vectors and extract storm movement trends. Radar data is analyzed using reflectivity gradients to identify areas of abrupt echo morphological changes. Satellite data is segmented using convolutional neural networks to define cloud structure features. In the feature fusion stage, monitoring node data with higher confidence scores are assigned higher weights; for example, when the confidence score of a ground station is higher than that of video surveillance, the feature weight of its corresponding area is increased. Obscured areas are filled with missing information using multi-view data, video data under low-light conditions undergoes histogram equalization to enhance contrast, and echo interference is filtered through Doppler velocity and consistency analysis with neighboring areas. The fused features are input into a classification model to determine the type and intensity of anomalies. Simultaneously, confidence values ​​are generated by combining the consistency of neighboring node data with the confidence level of the monitoring node; for example, when a node detects a strong echo but neighboring nodes show no anomalies, the confidence value is correspondingly reduced. The resulting preliminary risk information has higher confidence in the spatiotemporal dimensions.

[0057] As a preferred embodiment, the solution of this application is specifically implemented as follows: When running a spacetime anomaly detection model in an edge computing node, the following steps are included: First, the standardized multi-source meteorological data is divided into time-sliding windows and spatial grids. The time-sliding window is set to 30 minutes, sliding every 5 minutes. The spatial grid is divided into 10km × 10km grids. Within the same window, the standardized multi-source meteorological data is time-aligned and spatially registered.

[0058] Secondly, motion change characteristics of video surveillance data, echo morphology characteristics of radar observation data, and cloud structure characteristics of satellite remote sensing data are obtained respectively. For video surveillance data, optical flow field characteristics between consecutive frames are extracted, and motion vectors of local areas are calculated. For radar observation data, features such as echo intensity, echo top height, and vertically integrated liquid water content are extracted. For satellite remote sensing data, features such as cloud top temperature, cloud top height, and cloud development speed are extracted.

[0059] Then, features from different sources are weighted and fused according to the reliability scores of the monitoring nodes. Occlusion detection is performed during the fusion process, identifying occlusion areas by comparing depth information from adjacent frames. For low-light areas, an adaptive histogram equalization method is used for image enhancement. Ground clutter in radar echoes is removed using Doppler filtering.

[0060] Finally, the fused features are discriminated to obtain preliminary risk information including anomaly type and intensity. A first confidence value is generated based on the score output by the spatiotemporal anomaly detection model, combined with the consistency of neighboring nodes and the confidence score of the monitoring node. Specifically, a Long Short-Term Memory (LSTM) network is used to model the temporal features, and a Graph Convolutional Network (GCN) is used to model the spatial features. The outputs of both are fused to obtain the anomaly score. The first confidence value is calculated by weighted averaging the anomaly scores of neighboring nodes and the confidence score of the current node.

[0061] Through the above technical solutions, this application achieves the fusion and anomaly detection of multi-source heterogeneous meteorological data. By extracting spatiotemporal features and fusing multimodal data, the accuracy of identifying abnormal weather events is improved. Edge computing is used for preliminary risk identification, reducing data transmission burden and improving response speed. A reliable scoring mechanism is introduced to reduce interference from unreliable data sources on the identification results. Special processing for complex scenarios such as obstruction and low light enhances adaptability in various environments. Overall, the timeliness and reliability of extreme weather risk perception are improved.

[0062] In some of the schemes mentioned above in this application, fixed spatial relationships are used when constructing spatiotemporal map models, which cannot dynamically reflect the impact of changes in airflow direction and geographical conditions on the propagation of meteorological elements, thus limiting the accuracy of the model in predicting abnormal weather.

[0063] This application further proposes that when constructing a spatiotemporal graph model, monitoring nodes are used as graph nodes, and the geographical proximity, wind propagation path, and terrain influence between nodes are used as edges to form dynamic spatial relationships. During the model training phase, the connection strength between nodes is periodically updated based on changes in edge weights, so that the spatial relationships are adjusted according to changes in airflow direction and geographical conditions.

[0064] In this system, geographical proximity is calculated using straight-line distances between nodes; wind propagation paths are generated based on real-time wind direction data to establish upstream and downstream connections; and terrain influence relationships are established based on elevation data and obstacle distribution. Boundary weight changes are calculated through backpropagation of meteorological element propagation errors during the training phase, with periodic update intervals set at the minute or hour level. During dynamic adjustment, when sudden changes in wind direction or increased terrain shading effects are detected, invalid connections are automatically weakened, and the boundary weights corresponding to the dominant propagation path are strengthened.

[0065] Specifically, during the model training phase, after each training cycle, the edge weight gradient is calculated based on the propagation error of meteorological elements between nodes, and the edge weight matrix is ​​updated using the gradient descent method. For example, when strong winds occur in a certain area, wind direction propagation paths between nodes are generated based on wind direction sensor data. The edge weights of connections in the upwind direction are reduced to 30% of their original values, while the edge weights of connections in the downwind direction are increased to 150% of their original values. Simultaneously, combined with terrain elevation data, an attenuation coefficient is applied to connections between nodes located on the leeward side of mountains. The coefficient value is dynamically adjusted between 0.5 and 0.8 based on the terrain elevation difference. The resulting dynamic spatial correlation accurately reflects the propagation patterns of meteorological elements in complex terrain and changing airflows, improving the prediction accuracy of the spatiotemporal map model for extreme weather events.

[0066] As a preferred embodiment, the solution of this application is specifically implemented as follows: When constructing the spatiotemporal graph model, monitoring nodes are used as graph nodes, and geographical proximity, wind propagation paths, and topographic influences between nodes are used as edges to form dynamic spatial relationships. Specifically, firstly, an initial graph structure is constructed based on the geographical coordinates of the monitoring stations, with each monitoring station serving as a graph node. Then, proximity edges are set based on geographical distance, for example, connecting monitoring stations within 50 kilometers. Secondly, directed edges are added along the prevailing wind direction using wind field data from the numerical weather prediction model to characterize the propagation paths of meteorological elements. Furthermore, a digital elevation model is introduced to consider the blocking or guiding effects of topographic relief on the propagation of meteorological elements, and the edge weights are adjusted accordingly.

[0067] Furthermore, during the model training phase, the connection strength between nodes is periodically updated based on changes in edge weights, allowing spatial relationships to adjust with changes in airflow direction and geographical conditions. Specifically, edge weights are recalculated every 6 hours using the following methods: 1) updating the weights of wind direction propagation edges based on the latest wind field data; 2) adjusting the weights of geographically adjacent edges based on the correlation of actual observation data; and 3) fine-tuning the edge weights in conjunction with terrain changes. Thus, the spatiotemporal map model can dynamically adapt to the evolution of weather systems and more accurately depict the relationships between monitoring nodes.

[0068] Through the aforementioned technical solution, this application constructs a dynamically updated spatiotemporal map model capable of accurately capturing the spatial propagation characteristics and mutual influences of meteorological elements. This model fully considers factors such as geographical location, wind direction, and topography, making the risk identification results more consistent with actual meteorological processes. By periodically updating edge weights, the model can adapt to the dynamic changes of weather systems, improving the accuracy and timeliness of extreme weather event warnings. Furthermore, dynamically adjusted spatial relationships help optimize the allocation of monitoring resources, achieving more efficient meteorological monitoring and early warning.

[0069] In some of the schemes mentioned above in this application, the spatiotemporal graph model may cause the risk identification results to deviate from the actual meteorological evolution law due to differences in modalities of multi-source data or mismatch of physical laws during the training and inference process. Specifically, this may manifest as energy deficit imbalance, abnormal water vapor flux, or cross-modal data conflict, affecting the accuracy of risk scoring.

[0070] This application further proposes a joint constraint on the predicted results of temperature, air pressure, humidity, and precipitation in the spatiotemporal map model during the model training phase by introducing physical boundary conditions of energy and humidity. This joint constraint includes checking the energy budget and water vapor flux balance of the predicted results generated in each iteration of the spatiotemporal map model and feeding back deviation information to adjust the weights of the spatiotemporal map model. During the inference phase, when a discrepancy is detected between the predicted results and the physical boundary conditions, the weight allocation of the corresponding region in the risk score is corrected. Cross-modal consistency constraints are implemented during the model training phase by temporally aligning and spatially registering the echo distribution of radar observation data, cloud morphology of satellite remote sensing data, and real-time meteorological parameters from ground monitoring station data, and extracting common features between different modes based on a feature mapping network. During the inference phase, when any mode of data shows drift or is missing, the weight of that mode in the risk identification result is adjusted based on the high confidence results of the remaining mode data.

[0071] The energy budget and water vapor flux balance detection is achieved by calculating the matching degree between the gradient changes of the temperature and humidity fields and the energy conservation equation. For example, in each iteration cycle during the training phase, the closure error of the first law of thermodynamics is calculated for the temperature prediction value. When the error exceeds a set threshold, a backpropagation signal is generated to adjust the model parameters. The water vapor balance constraint compares the flux difference between the humidity prediction value and the measured precipitation. If the difference exceeds the allowable range, the model weight is updated. The feature mapping network for cross-modal consistency constraints uses a convolutional neural network to spatially align data from different modalities. Radar echo data and satellite cloud images are matched at multiple scales through feature pyramids, and ground monitoring data is interpolated and mapped to a grid with the same spatial resolution. After common features are extracted, a cross-modal correlation matrix is ​​generated, which is used for multi-modal feature fusion during the training phase and abnormal modality weight adjustment during the inference phase.

[0072] Specifically, during the model training phase, the energy conservation constraint imposes hard boundary conditions on the predicted temperature and pressure results through physical equations. For example, the predicted temperature value is substituted into the energy conservation equation to calculate the theoretical energy change. If the deviation from the model's predicted energy change exceeds 5%, a penalty term is added to the loss function to adjust the model weights. The water vapor balance constraint calculates water vapor flux using measured data of humidity field and precipitation. If the relative error between the predicted flux and the measured flux exceeds 10%, model parameter updates are triggered. The cross-modal consistency constraint performs time synchronization and spatial registration of radar, satellite, and ground data during training. For example, feature matching is performed between radar echo intensity maps and satellite cloud images at the same time, and a feature mapping network is used to extract the spatial correlation features between the two to generate a cross-modal correlation matrix as model input. During the inference phase, when radar data is missing echoes due to equipment failure, the weight of radar mode in risk scoring is reduced to 30% based on the cloud morphology characteristics of high-confidence areas in satellite cloud images and measured precipitation data from ground monitoring stations. At the same time, the weight of satellite and ground data is increased to 70%, thereby ensuring the physical rationality and cross-modal consistency of risk identification results.

[0073] As a preferred embodiment, the solution of this application is specifically implemented as follows: During the model training phase, meteorological energy conservation constraints and water vapor balance constraints are jointly applied to the predicted results of temperature, air pressure, humidity, and precipitation in the spatiotemporal map model by introducing physical boundary conditions for energy and humidity. Specifically, in the spatiotemporal map model, the energy budget and water vapor flux balance of the predicted results generated in each iteration are checked, and deviation information is fed back to adjust the weights of the spatiotemporal map model. For example, the energy budget check includes calculating the net radiation flux, sensible heat flux, latent heat flux, and surface heat flux within the prediction area, ensuring that their sum is close to zero. The water vapor flux balance check includes calculating the changes in water vapor input, output, and storage within the prediction area, ensuring that their sum is close to zero.

[0074] During the inference phase, when a discrepancy is detected between the predicted result and the physical boundary conditions, the weighting of the corresponding region in the risk score is adjusted. For example, if the predicted temperature for a region is abnormally high without a corresponding increase in energy input, the weighting of the temperature anomaly in that region on the risk score is reduced.

[0075] Cross-modal consistency constraints align the echo distribution of radar observation data, cloud morphology of satellite remote sensing data, and real-time meteorological parameters from ground monitoring stations in terms of time and space during the model training phase. Common features among different modalities are then extracted based on a feature mapping network. Specifically, timestamp alignment and spatial interpolation methods are used to synchronize multi-source data. Then, local features of each modality are extracted using a convolutional neural network, and finally, an attention mechanism is used to fuse the feature representations of different modalities.

[0076] During the inference phase, when any modality data shows drift or is missing, the weight of that modality in the risk identification result is adjusted based on the high confidence results of the remaining modal data. For example, if satellite remote sensing data shows vigorous cloud development in a certain area, but ground monitoring station data does not show corresponding precipitation signals, the weight of ground monitoring station data in the precipitation risk assessment of that area is reduced.

[0077] Through the aforementioned technical solutions, this application improves the accuracy and reliability of meteorological risk identification. By introducing physical constraints and cross-modal consistency checks, it reduces potential misjudgments from single data sources or models, making risk assessment results more consistent with actual meteorological processes. Simultaneously, by dynamically adjusting the weights of different data sources, it enhances adaptability to data quality fluctuations and improves the stability of risk identification under complex meteorological conditions. Furthermore, the model training and inference mechanism based on physical constraints makes the AI ​​model's predictions more consistent with meteorological principles, improving the model's interpretability and credibility.

[0078] In some of the solutions described above in this application, after receiving the edge recognition results and the output results of the spatiotemporal map model, a weighted fusion is directly performed to generate candidate risk scores. However, the constraints of the physical laws governing meteorological energy conservation and water vapor balance are not considered, which may lead to deviations in the candidate risk scores that contradict the changing trends of meteorological elements. Furthermore, no cross-validation mechanism is established between different data modalities. When data from one modality drifts or is missing, it is impossible to correct it using data from other modalities. In addition, the risk changes of adjacent monitoring nodes are not synchronized, which may reduce the confidence level of the risk identification results due to local data anomalies.

[0079] This application further proposes receiving edge identification results and spatiotemporal map model output results, and weighting and fusing node-level and region-level identification results based on a comprehensive confidence score to obtain candidate risk scores. Deviation detection is performed on the candidate risk scores based on meteorological energy conservation constraints and water vapor balance constraints; when the deviation exceeds the allowable range, the risk score for the corresponding region is adjusted. Comparisons are performed between different data modalities, and drifting or missing data modalities are corrected using the joint results of radar observation data, satellite remote sensing data, and ground monitoring station data. Risk changes of spatially adjacent monitoring nodes are examined; when the risk changes of adjacent nodes show a continuous or synchronous trend, the second confidence value of the risk identification result is increased, ultimately obtaining the risk identification result.

[0080] In the weighted fusion process, the comprehensive credibility score is used to dynamically adjust the weight allocation between node-level and regional-level results. For example, when the node-level credibility score is higher than the regional-level score, the node-level weight is increased to 0.7. The deviation detection stage sets the allowable deviation range at ±5%. When the energy expenditure deviation of a candidate risk score exceeds ±5%, a weight adjustment based on physical constraints is triggered. Data modality correction employs multimodal joint comparison. When satellite remote sensing data drifts, interpolation compensation is performed using the matching results of radar echo distribution and ground monitoring parameters. Adjacent node verification calculates the spatial gradient change of the risk score. When the gradient value is lower than a preset threshold, it is determined to be a continuous trend, and the second confidence value is increased by 10%.

[0081] Specifically, after candidate risk scores are generated, their physical rationality is first verified through energy conservation and water vapor balance constraints. For example, the energy budget of temperature and precipitation is checked for consistency. If a surge in precipitation is detected in a certain area but the temperature does not decrease synchronously, it is determined that the energy deviation exceeds the allowable range, and the risk score of that area is downweighted. Subsequently, multimodal data cross-validation is performed. When there are contradictions between satellite cloud morphology data and ground humidity monitoring data, radar echo distribution data is used as the primary arbitration method, and historical stability corrections are applied to the satellite data. Finally, spatial continuity analysis is performed on the risk scores of adjacent nodes. If multiple adjacent nodes show an increasing risk trend along the wind propagation path, it is determined to be a systemic risk signal, and the second confidence value is increased from 0.8 to 0.9. Through these steps, the risk identification results are ensured to conform to physical constraints and have the support of multimodal data consistency and spatial continuity, thus improving the accuracy of extreme weather warnings.

[0082] As a preferred embodiment, the solution of this application is specifically implemented as follows: The system receives the edge recognition results and the output of the spatiotemporal graph model. It then weights and fuses the node-level and region-level recognition results based on a comprehensive credibility score to obtain a candidate risk score. Specifically, for each monitoring node, the risk score from the edge recognition results and the risk score output by the spatiotemporal graph model are weighted and averaged according to the comprehensive credibility score to obtain the candidate risk score for that node.

[0083] Furthermore, deviation detection is performed on candidate risk scores based on meteorological energy conservation constraints and water vapor balance constraints. For example, the predicted values ​​of temperature, air pressure, humidity, and precipitation corresponding to the candidate risk scores are calculated, and it is checked whether these predicted values ​​meet the energy budget balance and water vapor flux balance. When the deviation exceeds the allowable range, the risk score for the corresponding area is adjusted to ensure that the risk score is consistent with the changing trends of meteorological elements.

[0084] Therefore, comparisons are made between different data modalities. Data modalities with drift or missing information are corrected using the combined results of radar observation data, satellite remote sensing data, and ground monitoring station data. Specifically, when the risk score of a certain data modality differs significantly from other modalities, the risk score of that modality is corrected based on the high-confidence results of the other modalities.

[0085] The risk changes of spatially adjacent monitoring nodes are examined. When the risk changes of adjacent nodes show a continuous or synchronous trend, the second confidence value of the risk identification result is increased. For example, the spatial correlation of risk scores of adjacent nodes is calculated, and when the correlation exceeds a preset threshold, the second confidence value is increased. The final risk identification result is obtained, including the corrected risk score and the updated second confidence value.

[0086] Through the above technical solutions, this application achieves the fusion and correction of multi-source data, improving the accuracy and reliability of risk identification. The introduction of meteorological energy conservation and water vapor balance constraints ensures the consistency between risk scores and physical laws. Cross-modal comparison and spatial correlation testing further enhance the stability of risk identification results. Therefore, this solution can more accurately capture the occurrence and evolution of extreme weather events.

[0087] In some of the solutions mentioned above in this application, there is a problem that fixed thresholds cannot adapt to regional differences and data fluctuations when determining dynamic alarm thresholds, which increases the risk of false alarms or missed alarms.

[0088] This application further proposes archiving historical risk scores by monitoring area and season, and extracting reference thresholds based on similar weather events and underlying surface conditions. The reference thresholds are adjusted based on a second confidence value and a comprehensive confidence score, raising the trigger threshold when the second confidence value is below the second confidence value threshold and lowering it when the second confidence value is above the second confidence value threshold. Continuous verification and hysteresis control are performed on risk scores over continuous time periods. When the risk score continuously exceeds the dynamic alarm threshold and the duration exceeds the duration threshold, a graded meteorological anomaly alarm is generated. The alarm level is determined based on the magnitude and duration of the risk score exceeding the limit. Simultaneously, adjacent monitoring nodes within spatial relationships are merged to avoid duplicate alarms.

[0089] When archiving historical risk scores, the monitoring area is divided into latitude and longitude grids, and the seasons are divided into four to six stages according to meteorological standards. Similar weather events include typhoons, rainstorms, or cold waves, and similar underlying surface conditions cover terrain type and land cover category. Reference thresholds are determined by statistically analyzing the percentiles of historical risk scores under similar conditions; for example, the top 10% percentile is used as the initial trigger threshold. During the adjustment process, the second confidence threshold is set to 0.7, and the adjustment range of the trigger threshold is linearly related to the confidence value deviation; for every 0.1 decrease in confidence value, the trigger threshold is increased by 5%. A sliding window mechanism is used for persistence testing, with a window length set from 30 minutes to 2 hours. Hysteresis control is achieved by setting the difference between the rising trigger threshold and the falling release threshold, with the difference ranging from 10% to 20% of the trigger threshold. Alarm levels are divided into three levels; each 20% increase in the exceedance or each 30-minute increase in duration raises the level by one. The merging of adjacent monitoring nodes is based on the geographical proximity and wind propagation path in spatial relationships. After merging, only the alarm information of the central node is retained.

[0090] Specifically, the process of determining the dynamic alarm threshold first establishes regional and seasonal reference benchmarks through historical data archiving. For example, historical risk score distributions are extracted for typhoon season rainstorms in South China, and the top 10% quantiles are used as the initial threshold. When the second confidence value is below 0.7, the trigger threshold is increased proportionally to avoid false triggers under low confidence levels; for example, the trigger threshold is increased by 5% when the confidence value is 0.6. In the continuous verification phase, a 30-minute sliding window is used to continuously detect whether the risk score exceeds the adjusted threshold. If it exceeds the limit for three consecutive windows, an alarm is triggered. At the same time, a release threshold 10% higher than the trigger threshold is set to prevent false alarms caused by instantaneous fluctuations. The alarm level is dynamically divided according to the extent of exceeding the limit; for example, a level 2 alarm is triggered when the risk score exceeds the threshold by 20% for 60 minutes. When merging adjacent nodes, spatially related monitoring node groups are identified based on wind direction propagation paths. Only the alarm information of the node with the highest risk score within the group is retained to eliminate duplicate alarms caused by spatial continuity. This process improves alarm accuracy and the operability of alarm information by combining historical data to dynamically adjust thresholds, confidence-driven threshold adjustment, and intelligent merging of spatial correlations.

[0091] As a preferred embodiment, the solution of this application is specifically implemented as follows: Historical risk scores are archived according to monitoring area and season, and reference thresholds are extracted based on similar weather events and underlying surface conditions. For example, for severe convective weather in a coastal city during the summer, risk score data for the same period over the past 5 years can be classified and stored according to factors such as land-sea distribution and topographic features, and reference thresholds corresponding to different risk levels can be extracted.

[0092] The reference threshold is adjusted based on the second confidence level and the overall confidence score. Specifically, when the second confidence level is below 0.7, the reference threshold is increased by 10%. When the second confidence level is above 0.9, the reference threshold is decreased by 5%. Simultaneously, the threshold is fine-tuned by ±3% based on the variation of the overall confidence score between 0.8 and 1.0.

[0093] The risk score is continuously tested and hysteresis controlled over a continuous time period. Furthermore, a testing time window is set at 30 minutes. When the risk score continuously exceeds the dynamic alarm threshold within this window for a duration of 15 minutes, the generation process for a graded meteorological anomaly alarm is triggered.

[0094] The alarm level is determined based on the extent and duration of the risk score exceeding the limit. Thus, an exceedance of 10%–30%, 30%–50%, and above 50% can be respectively classified as Level 3, Level 2, and Level 1 alarms, with the alarm level increasing by one level for every 30 minutes of duration.

[0095] To avoid duplicate alarms, adjacent monitoring nodes within a spatial relationship are merged. For example, adjacent nodes within a 10-kilometer range are treated as a single alarm unit, and only the highest-level alarm information is retained.

[0096] Through the above technical solutions, this application achieves dynamic alarm threshold adjustment based on historical data and real-time reliability, improving the accuracy and timeliness of extreme weather warnings. Simultaneously, through continuous verification and spatial correlation processing, false alarms and duplicate alarms are reduced, optimizing the quality of warning information. Furthermore, the tiered alarm mechanism makes emergency response more targeted, contributing to improved disaster prevention and mitigation efficiency.

[0097] In some of the solutions mentioned above in this application, the sampling frequency of monitoring nodes is adjusted based on the joint results of risk score, second confidence value and comprehensive credibility score. However, in the process of dynamic adjustment, if the synergistic effect of confidence level and data credibility under different risk states is not considered, the sampling frequency adjustment may lack pertinence and fail to achieve a balance between monitoring efficiency and data quality under resource constraints.

[0098] This application further proposes adjusting the sampling frequency of monitoring nodes based on the joint results of risk score, second confidence value, and comprehensive confidence score. When the risk score exceeds the dynamic alarm threshold and the second confidence value is less than or equal to the second confidence value threshold, the sampling frequency of the corresponding monitoring node is set to a high-frequency sampling value. When the risk score is lower than the dynamic alarm threshold and the second confidence value is higher than the second confidence value threshold, the sampling frequency of the corresponding monitoring node is set to a low-frequency sampling value. When the risk score is lower than the dynamic alarm threshold and remains stable within a continuous fixed time window, the sampling frequency of the monitoring node is restored to a normal sampling value. The sampling frequency of monitoring nodes in spatially correlated areas that share observation regions is kept consistent.

[0099] The high-frequency sampling value is set within a range of 2-5 times the regular sampling value, while the low-frequency sampling value is set within a range of 0.2-0.5 times the regular sampling value. The duration of the continuous fixed time window is set to 10-30 minutes based on seasonal characteristics. Monitoring nodes in the shared observation area form overlapping coverage areas through geographical proximity or wind direction propagation paths.

[0100] Specifically, when the risk score of a monitoring node exceeds the dynamic alarm threshold and the second confidence value fails to reach the preset threshold, it indicates that there is a potential risk in the area, but the reliability of the model's judgment is insufficient. In this case, the sampling frequency is increased to a high-frequency mode to obtain more dense observation data, for example, shortening the regular 5-minute sampling interval to 1 minute. When the risk score is below the threshold and the second confidence value is high, it indicates that the judgment of the current state is highly certain. In this case, the sampling frequency is reduced to a low-frequency mode to reduce resource consumption, for example, extending the sampling interval to 10 minutes. When the risk score remains below the threshold and there are no fluctuations within three consecutive time windows, it automatically reverts to the regular sampling interval. For node groups sharing the observation area, their sampling time and interval are forcibly synchronized to avoid data conflicts caused by asynchronous sampling. For example, three nodes on the same wind propagation path use the same sampling timestamp to ensure the continuity of data in the spatiotemporal dimension. This scheme optimizes the resource allocation efficiency of edge computing nodes while ensuring the accuracy of risk identification through a multi-parameter joint decision-making mechanism.

[0101] As a preferred embodiment, the specific implementation of this application's solution is as follows: During the adjustment of the sampling frequency of monitoring nodes, the risk score, second confidence value, and comprehensive reliability score of each monitoring node within the current time window are first obtained. When the risk score exceeds the dynamic alarm threshold and the second confidence value is less than or equal to a preset 0.7, the sampling frequency of the corresponding monitoring node is increased from the usual 5 minutes / time to 30 seconds / time. When the risk score is lower than the dynamic alarm threshold and the second confidence value is higher than 0.85, the sampling frequency is reduced from the usual value to 15 minutes / time. If the risk score is lower than the dynamic alarm threshold for six consecutive time windows and the fluctuation range is less than 5%, the sampling frequency is restored to 5 minutes / time. For node groups with shared observation areas in spatial relationships, such as three adjacent monitoring nodes located on the same wind propagation path, when the sampling frequency of any one node is adjusted, the sampling frequencies of the other two nodes are synchronously set to the same value to ensure that the data acquisition time of the shared area is synchronized.

[0102] Through the above technical solution, this application achieves dynamic sampling optimization based on real-time risk assessment. In low-risk scenarios, it reduces data acquisition frequency to save computational resources, while in high-risk scenarios, it improves data timeliness through high-frequency sampling. By employing a sampling synchronization mechanism for spatially related nodes, it avoids data comparison errors caused by inconsistent sampling periods between adjacent nodes, ensuring the accurate capture of the spatial continuity of meteorological elements. This solution balances the contradiction between resource consumption and early warning accuracy, and solves the problem that fixed sampling mechanisms in existing technologies cannot adapt to dynamic risk changes.

[0103] In some of the solutions mentioned above in this application, the difference in sampling frequency between nodes in the shared observation area within the spatial correlation is not considered when dynamically adjusting the sampling frequency of monitoring nodes. This leads to inconsistent sampling frequencies of different nodes in the same area, which may cause data redundancy or waste of resources.

[0104] This application further proposes adjusting the sampling frequency of monitoring nodes based on risk scores, second confidence values, and comprehensive confidence scores, including: adjusting the sampling frequency of monitoring nodes based on the joint result of the risk score, second confidence value, and comprehensive confidence score. When the risk score exceeds the dynamic alarm threshold and the second confidence value is less than or equal to the second confidence value threshold, the sampling frequency of the corresponding monitoring node is set to a high-frequency sampling value. When the risk score is lower than the dynamic alarm threshold and the second confidence value is higher than the second confidence value threshold, the sampling frequency of the corresponding monitoring node is set to a low-frequency sampling value. When the risk score is lower than the dynamic alarm threshold and remains stable within a continuous fixed time window, the sampling frequency of the monitoring node is restored to a normal sampling value. The sampling frequency of monitoring nodes in spatially correlated areas that share observation regions is kept consistent.

[0105] High-frequency sampling values ​​can be set to be collected once per second, low-frequency sampling values ​​to be collected once every five minutes, and regular sampling values ​​to be collected once per minute. Monitoring nodes in a shared observation area are associated through geographical proximity or wind propagation paths. When adjusting the sampling frequency, it must be ensured that the sampling frequency of all nodes in the same area is updated synchronously. For example, if there are three shared observation nodes in an area, when the risk score of one node triggers high-frequency sampling, the sampling frequency of the other two nodes is synchronously increased to once per second.

[0106] Specifically, when the risk score exceeds the dynamic alarm threshold and the second confidence value is low, high-frequency sampling values ​​are used to enhance data collection density and improve the accuracy of risk identification. When the risk score is below the threshold and the second confidence value is high, low-frequency sampling values ​​reduce resource consumption. A continuous fixed time window is set to 30 minutes; if the risk score does not fluctuate within this window, the regular sampling frequency is restored. Consistency of node sampling frequency in the shared observation area is achieved through unified scheduling on the cloud platform. For example, by issuing commands to synchronously adjust the collection cycle parameters of multiple nodes, data conflicts or resource waste caused by differences in node sampling frequencies in the same area are avoided. Thus, while ensuring the accuracy of risk perception, the allocation efficiency of monitoring resources is optimized.

[0107] As a preferred embodiment, the solution of this application is implemented as follows: During the alarm information transmission stage, an alarm payload containing alarm level, alarm object, spatial correlation, validity period, sampling adjustment parameters, and timestamp is generated. Each alarm payload generates a 16-bit unique identifier using a hash algorithm. For alarm information with the same alarm level and overlapping valid time windows within the same monitoring area, spatial correlation mapping is used for merging to form a single alarm payload with a merged node list. Sampling adjustment instructions are synchronously integrated into a single instruction. When sending, red alarm levels are given priority. A priority queue is established using the MQTT protocol. If no confirmation is received from the cloud platform within ten seconds, retransmission is performed via a TCP long connection. After three failed retransmissions, the system switches to the BeiDou short message channel for transmission. On the terminal display device, alarm information is mapped to a four-color interface (red, orange, yellow, blue) according to its level, simultaneously triggering a tiered alarm sound from a buzzer and an LED flashing signal. The spatial correlation in the alarm payload is displayed by overlaying a GIS map.

[0108] Through the above technical solutions, this application achieves efficient transmission and reliable delivery of extreme weather warning information, solving the problems of information loss caused by repeated alarm transmission and single transmission channels in traditional early warning systems. Unique identifiers and merging mechanisms avoid resource waste caused by duplicate alarms from multiple nodes in the same area. Multi-channel redundant transmission and priority control ensure timely delivery of high-level alarms even in harsh communication environments. Sound and light coordinated prompts and spatial visualization enhance the identifiability of alarm information and the efficiency of emergency response.

[0109] The above embodiments construct a meteorological AI early warning system based on edge recognition, cloud-based spatiotemporal inference, and dynamic alarm control, achieving highly reliable, low-latency, full-cycle dynamic perception and tiered response to extreme weather risks. By introducing a reliable scoring mechanism based on data quality, proximity consistency, and historical stability, adaptive correction and node replacement are performed on multi-source meteorological data, improving the reliability of input data. Deploying a spatiotemporal anomaly detection model at the edge enables real-time local feature extraction and preliminary risk identification, shortening the early warning response time. In the cloud, by integrating spatiotemporal graph models with physical and semantic constraints such as meteorological energy conservation, water vapor balance, and cross-modal consistency, physical consistency and global correction of risk identification results are achieved, improving the accuracy and stability of early warnings. Based on a dynamic threshold determination and sampling frequency control mechanism using risk scores, second confidence values, and comprehensive reliable scores, monitoring density and alarm levels can be adjusted according to risk levels, ultimately achieving intelligent, collaborative, and tiered dynamic early warning for extreme weather events, enhancing the real-time and intelligent response capabilities in the field of meteorological disaster monitoring.

[0110] In another preferred embodiment based on the above embodiments, see [reference] Figure 2As shown, this embodiment provides a meteorological AI anomaly early warning system for extreme weather risk perception, used to apply the above-mentioned meteorological AI anomaly early warning method for extreme weather risk perception, including: The acquisition unit is configured to acquire multi-source meteorological data and perform preprocessing to obtain standardized multi-source meteorological data. The preprocessing includes normalization and screening.

[0111] The identification unit is configured to run a spacetime anomaly detection model in an edge computing node, extract features and identify anomalies from standardized multi-source meteorological data over a continuous time period, and obtain preliminary risk information and a first confidence value.

[0112] The prediction unit is configured to upload preliminary risk information to the cloud, construct a spatiotemporal graph model, and apply consistency constraints during the training and inference of the spatiotemporal graph model to obtain risk identification results.

[0113] The judgment and adjustment unit is configured to determine the dynamic alarm threshold based on the risk identification result. When the risk score exceeds the dynamic alarm threshold and the comprehensive confidence score meets the comprehensive confidence score threshold, a graded meteorological anomaly alarm message is generated, and the sampling frequency of the monitoring node is adjusted according to the risk score, the second confidence value and the comprehensive confidence score.

[0114] The early warning unit is configured to send meteorological anomaly alarm information and sampling adjustment instructions to the cloud platform and terminal display device to conduct risk perception and early warning issuance for extreme weather.

[0115] Specifically, the acquisition unit receives data from ground monitoring stations, radar, satellites, and video. It generates a confidence score by verifying data integrity and comparing it with neighboring nodes, and performs interpolation or node replacement for low-confidence data. The identification unit divides standardized data into spatiotemporal windows, integrates multimodal features, and outputs the anomaly type and intensity, combining this with the confidence score to generate a first confidence value. The prediction unit constructs a dynamic spatial correlation graph in the cloud, optimizes model training through energy conservation, water vapor balance, and cross-modal consistency constraints, and generates a risk score including a second confidence value. The judgment and adjustment unit dynamically adjusts alarm thresholds based on historical archived data and real-time confidence values. An alarm is triggered when the risk score continuously exceeds the limit, and the sampling frequency of monitoring nodes is adjusted to high-frequency, low-frequency, or conventional modes according to the risk level. The early warning unit encapsulates alarm information and sampling instructions into a payload, transmits it to terminal devices in priority order, avoids duplicate alarms through merging processing, and switches to a backup channel in case of transmission failure. This system, through edge computing and cloud collaboration, achieves data quality assessment, dynamic threshold adjustment, and closed-loop control, improving the accuracy and timeliness of extreme weather warnings.

[0116] As a preferred embodiment, the solution of this application is implemented as follows: The meteorological AI anomaly early warning system for extreme weather risk perception includes a data acquisition unit, an identification unit, a prediction unit, a judgment and adjustment unit, and an early warning unit. The data acquisition unit acquires raw meteorological data through ground monitoring stations, radar, satellites, and video surveillance terminals, performs data format unification and normalization processing, calculates the credibility score of each monitoring node, and corrects the score by interpolating neighboring nodes or reconstructing historical data when the credibility score is below a threshold, outputting standardized multi-source meteorological data. The identification unit is deployed on edge computing nodes, divides the standardized data into spatiotemporal windows, extracts motion features from video surveillance, radar echo morphology features, and satellite cloud structure features, fuses them, and inputs them into a spatiotemporal anomaly detection model to generate preliminary risk information and a first confidence value. The prediction unit receives the preliminary risk information, constructs a spatiotemporal map model in the cloud that includes geographical proximity, wind direction propagation paths, and terrain influence relationships, introduces meteorological energy conservation constraints, water vapor balance constraints, and cross-modal consistency constraints for training and inference, and outputs a risk score and a second confidence value. The judgment and adjustment unit dynamically adjusts the alarm threshold based on historical risk scores and the current second confidence value. When the risk score continuously exceeds the threshold and the comprehensive confidence score meets the standard, a graded alarm message is generated and the sampling frequency of the monitoring node is adjusted. The sampling frequency switches between high frequency, low frequency, or normal mode based on the joint state of the risk score, second confidence value, and comprehensive confidence score. The early warning unit encapsulates the alarm information into a payload containing the level, spatial correlation, and timestamp. After merging duplicate alarms, it sends them to the cloud platform and terminal devices in priority order. The terminal devices trigger graded audio-visual prompts based on the payload.

[0117] Through the above technical solutions, this application solves the problem of false alarms caused by the uncontrollable quality of multi-source meteorological data, and improves the reliability of data input through a reliable scoring mechanism and data correction strategy. It overcomes the deficiency of lack of global correction in local identification results of edge nodes, and enhances the accuracy of risk identification by utilizing physical constraints and multimodal consistency constraints in the spatiotemporal graph model. It optimizes the resource waste and response delay caused by fixed sampling frequency, and achieves adaptive adjustment of the sampling frequency based on dynamic risk scoring and confidence values. It avoids duplicate alarms and transmission redundancy, and improves the efficiency of early warning issuance through alarm payload merging and priority sorting mechanisms.

[0118] 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. A weather AI anomaly early warning method for extreme weather risk perception, characterized in that, The method comprises the following steps: Collecting and preprocessing multi-source meteorological data to obtain standardized multi-source meteorological data, the preprocessing including normalization processing and screening processing, the screening processing including calculating a trust score according to the data quality, adjacent consistency and historical stability of each monitoring node, and performing data correction or node replacement when the trust score is lower than a trust score threshold, the multi-source meteorological data including ground monitoring station data, radar observation data, satellite remote sensing data and video monitoring data; Running a spatio-temporal anomaly detection model in an edge computing node to perform feature extraction and anomaly identification on the standardized multi-source meteorological data in a continuous time period to obtain preliminary risk information and a first confidence value; Uploading the preliminary risk information to the cloud to build a spatio-temporal graph model and impose consistency constraints in the training and reasoning process of the spatio-temporal graph model to obtain a risk identification result, wherein the consistency constraints participate in the training in the form of model constraint terms and are used to correct the risk identification result in the reasoning stage, the spatio-temporal graph model contains meteorological monitoring nodes and their spatial correlation, the consistency constraints include meteorological energy conservation constraints, water vapor balance constraints and cross-modal consistency constraints, and the risk identification result includes a risk score and a second confidence value; Determining a dynamic alarm threshold according to the risk identification result, generating a hierarchical meteorological anomaly alarm information when the risk score exceeds the dynamic alarm threshold and the comprehensive trust score meets a comprehensive trust score threshold, and adjusting the sampling frequency of the monitoring node according to the risk score, the second confidence value and the comprehensive trust score, wherein the comprehensive trust score is obtained by weighting the trust scores of the monitoring nodes covered by the alarm object; Sending the meteorological anomaly alarm information and the sampling adjustment instruction to the cloud platform and the terminal display device to perform risk perception and early warning of extreme weather.

2. The extreme weather risk-aware meteorological AI anomaly alerting method of claim 1, wherein, The collecting and preprocessing of multi-source meteorological data comprises: Collecting raw meteorological data of ground monitoring stations, radars, satellite remote sensing and video monitoring terminals respectively; Performing unified formatting and normalization processing on data of different sources, different resolutions and different sampling periods to compare physical quantities such as temperature, humidity, wind speed, rainfall and reflectivity on a unified numerical scale; Calculating the trust score of each monitoring node, the trust score being determined according to the data integrity of the monitoring node, the consistency of adjacent node observation results and the historical fluctuation stability of the node, and when the trust score is lower than a preset threshold, abnormal data is corrected by interpolation replacement of adjacent high-trust nodes or backtracking reconstruction of the historical stable interval, and if a node continuously falls below the trust score threshold, a standby node takes over the corresponding monitoring task; The corrected multi-source meteorological data is the standardized multi-source meteorological data.

3. The extreme weather risk-aware meteorological AI anomaly alerting method of claim 1, wherein, When running the spatio-temporal anomaly detection model in the edge computing node, the method comprises the following steps: Dividing the standardized multi-source meteorological data by time sliding window and spatial grid, and performing time alignment and spatial registration on the standardized multi-source meteorological data in the same window; respectively acquire motion change features of the video monitoring data, echo shape features of the radar observation data, and cloud system structure features of the satellite remote sensing data, and perform weighted fusion on different source features according to a trust score of a monitoring node, and perform occlusion detection, low-illumination enhancement, and echo interference elimination during the fusion process; discriminate the fused features to obtain preliminary risk information including an abnormal type and an abnormal intensity, and generate the first confidence value according to a score output by the spatio-temporal anomaly detection model in combination with a consistency of adjacent nodes and the trust score of the monitoring node.

4. The extreme weather risk-aware meteorological AI anomaly alerting method of claim 3, wherein, When constructing the spatio-temporal graph model, the following steps are included: taking the monitoring nodes as graph nodes, taking geographical adjacent relationships, wind direction propagation paths, and terrain influence relationships between the nodes as edges, and forming a dynamic spatial correlation relationship; periodically updating connection strengths between the nodes according to edge weight changes in the model training stage, so that the spatial correlation relationship is adjusted according to changes in airflow direction and geographical conditions.

5. The extreme weather risk-aware meteorological AI anomaly alerting method of claim 4, wherein, The consistency constraint includes: The meteorological energy conservation constraint and the water vapor balance constraint jointly constrain the prediction results of temperature, air pressure, humidity, and precipitation in the spatio-temporal graph model by introducing physical boundary conditions of energy and humidity in the model training stage, the joint constraint includes balance detection of energy consumption and water vapor flux on the prediction results generated in each iteration in the spatio-temporal graph model, and feedback of deviation information to adjust the weights of the spatio-temporal graph model; in the inference stage, when it is detected that the prediction result does not match the physical boundary condition, the weight distribution of the corresponding region in the risk score is corrected; the cross-modal consistency constraint aligns the echo distribution of the radar observation data, the cloud system shape of the satellite remote sensing data, and the real-time meteorological parameters of the ground monitoring station data in time and space in the model training stage, and extracts common features between different modalities based on a feature mapping network; in the inference stage, when any modal data drifts or is missing, the weight of this modal in the risk identification result is adjusted according to the high-confidence results of the remaining modal data.

6. The extreme weather risk-aware meteorological AI anomaly alerting method of claim 5, wherein, When obtaining the risk identification result, the following steps are included: receive the edge identification result and the output result of the spatio-temporal graph model, and perform weighted fusion on the node-level identification result and the regional-level identification result according to the comprehensive trust score to obtain a candidate risk score; perform deviation detection on the candidate risk score according to the meteorological energy conservation constraint and the water vapor balance constraint, and when the deviation exceeds the allowed deviation range, adjust the risk score of the corresponding region, so that the risk score is consistent with the change trend of meteorological elements; compare different data modalities, and correct the data modalities that exist drift or missing through the joint results of the radar observation data, the satellite remote sensing data, and the ground monitoring station data; verify the risk changes of spatially adjacent monitoring nodes, and when the risk changes of adjacent nodes show a continuous or synchronous trend, increase the second confidence value of the risk identification result, and finally obtain the risk identification result.

7. The extreme weather risk-aware meteorological AI anomaly alerting method of claim 6, wherein, When determining the dynamic alarm threshold, the following steps are included: archive historical risk scores according to monitoring regions and seasons, and extract reference thresholds based on similar weather processes and similar underlying conditions. The second confidence value and the comprehensive confidence score are used to adjust the reference threshold, so that the trigger threshold is increased when the second confidence value is lower than the second confidence value threshold, and the trigger threshold is decreased when the second confidence value is higher than the second confidence value threshold; the risk score in a continuous time period is continuously checked and hysteresis controlled, and the hierarchical meteorological anomaly alarm information is generated when the risk score continuously exceeds the dynamic alarm threshold and the duration exceeds the duration threshold, and the alarm level is determined according to the exceeding amplitude and the duration of the risk score, and the adjacent monitoring nodes in the spatial correlation relationship are combined to avoid repeated alarms.

8. The extreme weather risk-aware meteorological AI anomaly alerting method of claim 7, wherein, When the risk score, the second confidence value and the comprehensive confidence score are used to adjust the sampling frequency of the monitoring node, the following is included: The risk score, the second confidence value and the comprehensive confidence score are used to adjust the sampling frequency of the monitoring node; when the risk score exceeds the dynamic alarm threshold and the second confidence value is less than or equal to the second confidence value threshold, the sampling frequency of the corresponding monitoring node is set to a high-frequency sampling value; when the risk score is lower than the dynamic alarm threshold and the second confidence value is higher than the second confidence value threshold, the sampling frequency of the corresponding monitoring node is set to a low-frequency sampling value; When the risk score is lower than the dynamic alarm threshold and stable in a continuous fixed time window, the sampling frequency of the monitoring node is restored to a regular sampling value; The sampling frequencies of the monitoring nodes in the spatial correlation relationship that share the observation area are kept consistent.

9. The extreme weather risk-aware meteorological AI anomaly alerting method of claim 1, wherein, When the meteorological anomaly alarm information and the sampling adjustment instruction are sent to the cloud platform and the terminal display device, the following is included: generating an alarm payload containing the alarm level, the alarm object, the spatial correlation relationship, the validity period, the sampling adjustment parameter and the timestamp, and assigning a unique alarm identifier to each alarm payload; The same alarm level within the same spatial correlation relationship within the validity period is combined to form a single alarm payload and a single sampling adjustment instruction; The sending order is established from high to low according to the alarm level, and the message is retransmitted when no reply is received within the confirmation time limit, and the backup transmission channel is switched to when the retransmission is still unsuccessful; The alarm payload is displayed and acousto-optic prompt is performed on the terminal display device side.

10. An extreme weather risk-aware meteorological AI anomaly alert system for applying the extreme weather risk-aware meteorological AI anomaly alert method of any one of claims 1-9, characterized in that, It includes: The acquisition unit is configured to collect multi-source meteorological data and perform preprocessing to obtain standardized multi-source meteorological data, and the preprocessing includes normalization processing and screening processing; The identification unit is configured to run the spatio-temporal anomaly detection model in the edge computing node, extract features and identify anomalies from the standardized multi-source meteorological data in a continuous time period, and obtain preliminary risk information and a first confidence value; The prediction unit is configured to upload the preliminary risk information to the cloud, build a spatio-temporal graph model and impose consistency constraints in the training and inference process of the spatio-temporal graph model, and obtain a risk identification result; A judgment adjustment unit is configured to determine a dynamic alarm threshold according to the risk identification result, generate a graded meteorological anomaly alarm information when the risk score exceeds the dynamic alarm threshold and the comprehensive confidence score meets a comprehensive confidence score threshold, and adjust the sampling frequency of the monitoring node according to the risk score, the second confidence value and the comprehensive confidence score. A warning unit is configured to send the meteorological anomaly alarm information and the sampling adjustment instruction to a cloud platform and a terminal display device for risk perception and early warning publication of extreme weather.

Images