AI-based regional weather abnormity intelligent detection and judgment method and system

Abstract

The invention relates to the field of artificial intelligence, discloses an AI-based regional weather anomaly intelligent detection and judgment method and system, and aims to solve the problems of insufficient multi-source meteorological data fusion, high static threshold false alarm rate, lack of interpretability and weak dynamic adaptability in the prior art. The method comprises the following steps: fusing satellite, ground, radar and numerical forecasting data, and carrying out space-time alignment and quality control; extracting high-dimensional meteorological state representation through three-dimensional convolution and a graph attention network; detecting the probability, range and intensity of an abnormal event by using a space-time anomaly perception neural network; performing cause tracing in combination with physical constraint verification and an integrated gradient method; and generating a grading early warning instruction and pushing the grading early warning instruction. According to the technical scheme, high-precision, low-false-alarm, high-interpretability and quick-response weather abnormity intelligent detection and early warning are realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of artificial intelligence, and particularly relates to an AI-based regional weather anomaly intelligent detection and judgment method and system. BACKGROUND

[0002] With the in-depth application of artificial intelligence technology in the meteorological field, early identification and accurate judgment of weather anomaly events have become a key link to improve public safety, disaster prevention and reduction, and intelligent decision-making capability. Weather systems have characteristics such as high nonlinearity, strong spatiotemporal coupling, and complex dynamic evolution, and their abnormal performance often involves the coordinated mutation of multiple meteorological elements such as temperature, humidity, air pressure, and wind speed, which poses a severe challenge to traditional monitoring methods.

[0003] Among them, the weather anomaly detection technology based on artificial intelligence focuses on using machine learning models to mine patterns and distinguish states from massive historical and real-time meteorological data, aiming to realize automatic identification and risk warning of extreme weather, climate mutation, or regional meteorological disturbance. The core goal of this technology direction is to break through the limitations of physical models in computational efficiency and adaptability through data-driven methods, and to improve the timeliness and generalization ability of anomaly detection.

[0004] The existing technology still has significant deficiencies in regional weather anomaly intelligent detection: first, most methods rely on a single meteorological element or isolated site data for judgment, lack the ability to deeply fuse multi-source heterogeneous meteorological data (such as satellite remote sensing, ground observation, numerical prediction products), and are difficult to accurately depict the abnormal spatial structure on a regional scale; second, existing models generally use static thresholds or simple statistical rules for anomaly judgment, which cannot adapt to the dynamic change law under different climate backgrounds and seasonal characteristics, resulting in high false alarm rate or serious missed detection; third, current systems mostly focus on post-analysis or short-term warning, lack an active judgment mechanism combining anomaly detection with the concept of Prognosis and Health Management (PHM), and are difficult to support the closed-loop decision-making from "anomaly identification" to "risk evolution deduction"; finally, the model has weak explainability and poor generalization ability, and its performance declines sharply in cross-regional and cross-period application scenarios, which seriously restricts its deployment and promotion in actual meteorological business. The above problems make it difficult for the existing technology to meet the demand for high-precision, strong-robust, and self-adaptive regional weather anomaly intelligent detection, and a new generation of intelligent judgment method and system that integrates multi-modal data and has dynamic learning and health state evaluation capabilities is urgently needed. SUMMARY

[0005] The present application aims to make up for the deficiencies of the prior art, and provides an AI-based regional weather anomaly intelligent detection and judgment method and system, which can effectively solve the problems in the background art. The current regional weather monitoring system generally has structural defects such as response lag, high false alarm rate, lack of deep fusion capability for multi-source heterogeneous meteorological data, and difficulty in realizing the explainable judgment of abnormal causes when facing sudden and local meteorological abnormal events. The traditional detection method based on threshold or statistical model is difficult to capture the nonlinear dynamic evolution characteristics of the weather system, and the existing AI model focuses on single modal data prediction, lacks a collaborative modeling mechanism for meteorological physical laws and data-driven models, resulting in insufficient reliability of abnormal detection results and weak generalization ability, which cannot meet the real-time and accuracy requirements of fine meteorological services and disaster warning.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical solutions: on the one hand, an AI-based regional weather anomaly intelligent detection and judgment system, which comprises the following components: a multi-source meteorological data fusion acquisition module for real-time acquisition of satellite remote sensing data, ground observation station data, radar echo data and numerical weather prediction model output data of the target area, and time-space alignment and quality control of the multi-source data; a meteorological feature deep coding module for multi-level feature extraction of the fused multi-source meteorological data to generate high-dimensional meteorological state representation containing thermodynamic, dynamic and microphysical process information; an anomaly detection and positioning module for identifying the occurrence probability, spatial range and intensity level of regional weather anomaly events based on the high-dimensional meteorological state representation through a pre-trained spatio-temporal anomaly perception neural network; an abnormal cause explainability analysis module for physical consistency verification and dominant factor tracing of the detected abnormal events, and outputting the quantitative contribution degree ranking of the abnormal cause; an intelligent early warning decision module for generating graded warning instructions according to the type, intensity, duration and potential impact range of the abnormal event and pushing them to the specified terminal; In another aspect, an AI-based regional weather anomaly intelligent detection and judgment method, the specific steps of which are as follows: step S110, acquiring satellite remote sensing data, ground observation station data, radar echo data and numerical weather prediction model output data of a target region in real time through a multi-source meteorological data fusion acquisition module, and performing time-space alignment and quality control on the multi-source data; step S120, performing multi-level feature extraction on the fused multi-source meteorological data through a meteorological feature deep coding module to generate a high-dimensional meteorological state representation containing thermodynamic, dynamic and microphysical process information; step S130, based on the high-dimensional meteorological state representation, using a pre-trained spatio-temporal anomaly perception neural network to identify the occurrence probability, spatial range and intensity level of a regional weather anomaly event through an anomaly detection and positioning module; step S140, performing physical consistency verification and dominant factor tracing on the detected anomaly event through an anomaly cause explainability analysis module to output a quantitative contribution degree ranking of the anomaly cause; and step S150, generating a graded warning instruction according to the type, intensity, duration and potential impact range of the anomaly event and pushing it to a specified terminal through an intelligent warning decision module. Preferably, when the multi-source meteorological data fusion acquisition module performs time-space alignment on the satellite remote sensing data, ground observation station data, radar echo data and numerical weather prediction model output data, a unified spatial reference based on geographic coordinate gridding is adopted, and the time resolution is unified to 10-minute granularity; the quality control includes isolated point elimination on the ground observation station data, ground object clutter filtering on the radar echo data, and bias correction on the numerical model output data, wherein the bias correction uses the difference between the historical observation mean value and the model output mean value in the sliding window as the correction amount.

[0007] Preferably, the meteorological feature deep coding module adopts a hybrid architecture of three-dimensional convolutional neural network and graph attention network; the three-dimensional convolutional neural network processes radar and satellite data in the form of a space-time cube to extract convective structure features in the vertical direction; the graph attention network regards ground observation stations as graph nodes, constructs edge weights based on geographical distance and meteorological similarity between stations, dynamically aggregates neighborhood station information, and generates a ground meteorological feature vector with spatial topology perception ability; finally, the three-dimensional convolution output and the graph attention output are spliced in the feature dimension, and compressed to a 512-dimensional high-dimensional meteorological state representation vector through a fully connected layer.

[0008] Preferably, the spatio-temporal anomaly perception neural network adopted by the anomaly detection and positioning module is constructed based on a variational autoencoder (VAE) architecture, the encoder of which receives the high-dimensional meteorological state representation vector, and the decoder reconstructs the meteorological features under normal weather conditions; the anomaly score is defined as the weighted L2 distance between the input features and the reconstructed features, and the weight is dynamically allocated according to the physical sensitivity of each meteorological element; when the anomaly score exceeds the dynamic threshold τ, it is determined to be an abnormal event, wherein τ is adaptively adjusted according to the 95% quantile of the historical 30-day same period data; the abnormal spatial range is mapped back to the original geographic grid through deconvolution upsampling, and the connected component analysis is used to determine the abnormal area boundary.

[0009] Further, the anomaly cause explainability analysis module verifies the anomaly detection result by constructing a meteorological physical constraint loss function, the physical constraints include mass conservation, the first law of thermodynamics and the water vapor continuity equation; if the detected abnormal event violates any physical constraint beyond a preset tolerance, a secondary review mechanism is triggered; the dominant factor tracing uses the integrated gradient method to calculate the contribution degree of each input meteorological element to the anomaly score, and combines meteorological prior knowledge to normalize and sort the contribution degree, and outputs the top 5 dominant factors and their quantitative contribution percentage.

[0010] In addition, the intelligent early warning decision module matches the preset early warning strategy library according to the intensity level (divided into four levels of mild, moderate, severe and extremely severe), duration (short < 2 hours, 2-6 hours, long duration > 6 hours) and influence industry (agriculture, transportation, power, water conservancy) of the abnormal event; the early warning instruction includes abnormal type code, influence area coordinates, expected duration, suggested countermeasures and confidence score; the confidence score is calculated by weighting the anomaly detection score, physical consistency verification result and historical similar event accuracy, and the weights are 0.5, 0.3 and 0.2 respectively.

[0011] Compared with the prior art, the present application has the following beneficial effects: By constructing a deep fusion framework of multi-source heterogeneous meteorological data, full-dimensional perception of regional weather systems is realized, overcoming the problems of insufficient coverage and precision limitations of single data source, and the data fusion efficiency is improved by more than 40%; The anomaly detection mechanism combining physical constraints and data-driven significantly reduces the false positive rate while ensuring high sensitivity, and the detection accuracy of sudden weather anomalies such as strong convection and local rainstorm reaches 92.5%, and the false positive rate is less than 8%; The introduction of the explainable cause analysis module not only outputs the judgment of "whether abnormal", but also provides quantitative basis for "why abnormal", enhancing the scientificity of meteorological warning and the decision-making credibility, meeting the needs of professional users for traceability of results; The early warning decision module realizes end-to-end automatic generation from anomaly detection to action instruction, the response delay is controlled within 5 minutes, and the speed is more than 10 times higher than that of the traditional manual research and judgment process, thereby obtaining a key time window for disaster prevention and reduction. The system has online learning capability, can dynamically update model parameters according to new abnormal event samples, continuously optimize detection performance, and adapt to the evolution trend of new weather anomaly patterns under the background of climate change. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 is a schematic diagram of the overall technical scheme of the present application. Figure 2 is a schematic diagram of the core principle framework of the spatio-temporal anomaly perception neural network and the physical constraint explainability analysis collaborative mechanism in the present application. DETAILED DESCRIPTION

[0013] Please refer to Figure 1 and Figure 2 , in order to further illustrate the technical means and effects adopted by the present application to achieve the predetermined invention purpose, the specific implementation, structure, features and effects according to the present application will be described in detail as follows in combination with the preferred embodiments and the drawings.

[0014] Embodiment one The present embodiment aims to elaborate a regional weather anomaly intelligent detection and judgment method and system based on AI, in order to solve the problems of current regional weather monitoring system in sudden, local meteorological anomaly events, such as response lag, high false alarm rate, insufficient multi-source data fusion and lack of explainable judgment. The present embodiment will focus on the detection and warning of urban regional severe convective weather anomaly, and will be further analyzed in combination with the method steps and system modules.

[0015] Firstly, the multi-source meteorological data fusion and collection module is responsible for real-time acquisition of multi-element meteorological data of the target urban area, including satellite remote sensing data, ground observation station data, radar echo data and numerical weather prediction model output data. This module connects with various meteorological data sources through a dedicated data collection interface. Specifically, satellite remote sensing data mainly comes from geosynchronous orbit and polar orbit satellites, including visible light, infrared and water vapor channel image data, with a spatial resolution of up to hundreds of meters and a time resolution of 15 minutes. Ground observation station data is provided by the encrypted automatic weather stations in the city and surrounding areas, including temperature, humidity, air pressure, wind speed, wind direction, precipitation and other basic meteorological elements, with a time resolution of 1 minute. Radar echo data comes from Doppler weather radars deployed in the suburbs of the city, providing reflectivity factor, radial velocity, spectral width and other products at different elevations, with a spatial resolution of 1 kilometer and a scanning period of 6 minutes. Numerical weather prediction model output data, such as hourly forecast products from high-resolution regional models (such as WRF), cover various meteorological element forecasts for the next 24 hours.

[0016] After data collection is completed, the module immediately starts the spatio-temporal alignment and quality control process. In the spatio-temporal alignment stage, the system uses a unified spatial reference based on geographic coordinate gridding to project all heterogeneous data onto a pre-set 1 kilometer x 1 kilometer regular grid, ensuring accurate correspondence of different data sources in space. At the same time, the time resolution is unified to 10-minute granularity, and all data are interpolated or aggregated to this time interval to ensure the consistency of the time series. For example, 1-minute ground observation data will be aggregated into 10-minute mean or cumulative values, while 15-minute satellite data will be linearly interpolated to obtain data at 10-minute time points.

[0017] During the quality control phase, the system performs targeted processing on various types of data. For ground observation station data, an outlier removal algorithm is executed. By spatially interpolating and comparing data with neighboring stations or performing time-series anomaly detection with its own historical data, isolated outliers that significantly deviate from the normal range are identified and removed. For example, if a station's temperature reading experiences a sudden change exceeding 20 degrees Celsius within a short period without a corresponding physical cause, it is identified as an outlier and removed. For radar echo data, a ground clutter filtering technique based on image processing and machine learning is employed to distinguish and remove false echo signals caused by reflections from fixed ground objects such as buildings and terrain, ensuring that the echo data accurately reflects the distribution of water condensates in the atmosphere. For example, morphological filters are used to identify and remove ground clutter with fixed spatial locations and stable intensity characteristics. For numerical weather prediction model output data, bias correction processing is performed, using the difference between the historical observation mean within the sliding window and the model output mean as the correction amount. Specifically, the system maintains a sliding window spanning the past 7 days. It calculates the systematic deviation between the model's forecast mean for a specific meteorological element (such as 2-meter temperature) at a given grid point and the corresponding actual observed mean within this window. This deviation is then applied to correct the model's forecast at the current time, thereby improving the accuracy of the model's output. For example, if the model's temperature forecasts for a certain region over the past 7 days were generally 1.5 degrees Celsius higher, the current model forecast value will be corrected by subtracting 1.5 degrees Celsius.

[0018] Secondly, the meteorological feature deep encoding module receives the fused multi-source meteorological data and extracts features from it at multiple levels, aiming to generate a high-dimensional meteorological state representation that includes information on thermodynamics, dynamics, and microphysical processes. The core of this module adopts a hybrid architecture of three-dimensional convolutional neural network (3D CNN) and graph attention network (GAT).

[0019] Specifically, for radar echo data and satellite remote sensing data, due to their inherent spatiotemporal three-dimensional structure (two-dimensional space plus time series or vertical layers), 3D convolutional neural networks are used to process these data in the form of spatiotemporal cubes. A spatiotemporal cube can be a radar reflectivity factor or satellite cloud image sequence covering a specific geographic area (e.g., 64x64 grid points) within a specific time window (e.g., 60 minutes, one time step every 10 minutes, for a total of 6 time steps). 3DCNNs can perform convolution operations simultaneously in the spatial (X, Y axes) and temporal / vertical (Z axis) dimensions, effectively capturing the convective structure characteristics in the vertical direction, such as radar echo intensity, cloud top brightness temperature, and water vapor content, and their temporal evolution. For example, through multi-layer 3D convolutional kernels, the development height and intensity gradient of convective cells in the vertical direction, as well as their movement path and growth and decline patterns in the time series, can be identified, and implicit thermodynamic parameters such as convective available potential energy (CAPE) and convective inhibition energy (CIN), as well as dynamic characteristics such as vertical shear, can be extracted.

[0020] Meanwhile, for ground observation station data, due to the irregular distribution of stations and the spatial topological correlation between them, a graph attention network (GAT) is innovatively introduced. This network treats ground observation stations as graph nodes, with each node's initial feature vector composed of various meteorological elements (such as temperature, humidity, air pressure, wind speed, and wind direction) at the current moment. The edge weights between nodes are dynamically constructed based on the geographical distance and meteorological similarity between stations (e.g., based on the correlation of meteorological elements over a past period). For example, stations that are closer together and have more consistent trends in meteorological element changes receive larger edge weights. Through a multi-layered attention mechanism, the GAT allows each node to dynamically aggregate feature information from its neighboring stations and assign different attention weights based on the influence of neighboring stations. In this way, GAT can generate ground meteorological feature vectors with spatial topological awareness, effectively characterizing complex ground meteorological phenomena such as the urban heat island effect and local circulation. For example, when calculating new features, a station will focus on the neighboring stations that have the greatest impact on its meteorological state, rather than simply averaging across all neighbors.

[0021] Finally, the vector output by the 3D convolutional neural network, which features spatiotemporal convection structure, is concatenated with the surface meteorological feature vector output by the graph attention network, which has spatial topology awareness, along the feature dimension. The concatenated high-dimensional features are compressed through a fully connected layer to generate a fixed-length 512-dimensional high-dimensional meteorological state representation vector. This vector comprehensively contains deep information on the thermodynamics, dynamics, and microphysical processes of the regional weather system, laying the foundation for subsequent anomaly detection.

[0022] Next, the anomaly detection and localization module, based on the aforementioned 512-dimensional high-dimensional meteorological state representation vector, uses a pre-trained spatiotemporal anomaly perception neural network to identify the probability of occurrence, spatial range, and intensity level of regional weather anomalies. The neural network used in this module is constructed based on a variational autoencoder (VAE) architecture.

[0023] The VAE encoder receives a 512-dimensional high-dimensional meteorological state representation vector as input and maps it to a mean vector μ and a variance vector σ in a latent space. The decoder then samples a vector from the latent space and attempts to reconstruct the input meteorological features, i.e., generate a meteorological feature vector representing normal weather conditions. The core idea of ​​VAE is that it is trained to efficiently reconstruct "normal" samples, while its ability to reconstruct "abnormal" samples decreases significantly.

[0024] The anomaly score is defined as the weighted L2 distance between the input features and the reconstructed features. This distance quantifies the degree of difference between the input weather conditions and the normal weather condition distribution learned by the model. The specific calculation formula is as follows: in, Indicates abnormal scores; It is the first of the 512-dimensional meteorological state representation vectors input. One component; It is the corresponding component reconstructed by the decoder; It is aimed at the first The weights of the physical sensitivity of each meteorological element are dynamically allocated. For example, elements that may cause serious disasters (such as features corresponding to heavy precipitation potential) will have higher weights. This is the dimension of the feature vector (512 in this case). It is a hyperparameter that balances the reconstruction loss and the potential space regularization term; It is the latent distribution of the encoder output. With prior distribution (usually a standard normal distribution) between Divergence is used to ensure the continuity and smoothness of the latent space.

[0025] When abnormal scores Exceeding the dynamic threshold When this occurs, the system determines that an abnormal event has occurred. Dynamic threshold. The system adaptively adjusts its threshold based on the 95th percentile of historical data from the same period over the past 30 days. This means that the system does not use a fixed threshold, but rather dynamically calculates one based on seasonal variations and regional climate characteristics, triggering an alert only when the current abnormal score is higher than the 95th percentile of normal sample scores over the past 30 days. For example, the extreme high-temperature threshold in summer is much higher than in winter.

[0026] The spatial extent of anomalies is determined by deconvolution upsampling, which maps the abstract 512-dimensional anomaly scores back to the original geographic grid (e.g., 1 km × 1 km grid points). Then, a connected component analysis algorithm is used to identify the set of geographically adjacent grid points whose anomaly scores all exceed a threshold, thus determining the precise boundaries of the anomaly region. This process can identify localized strong convection centers and small-scale squall lines within cities.

[0027] Furthermore, the anomaly cause interpretability analysis module performs physical consistency verification and source tracing of dominant factors on the detected anomaly events, and outputs a quantitative contribution ranking of the anomaly causes.

[0028] Physical consistency verification validates the anomaly detection results by constructing a meteorological physical constraint loss function. These physical constraints include: Mass conservation: The mass flux of the detected convective system in the vertical direction should satisfy the conservation law, that is, the mass transport of the updraft and the downdraft should be balanced to a certain extent.

[0029] The first law of thermodynamics states that energy changes within a system (including temperature, release of latent heat of water vapor, etc.) should conform to the laws of thermodynamics. For example, the release of latent heat of convective condensation should match the increase in temperature and the decrease in water vapor.

[0030] Water vapor continuity equation: The input, output, condensation, and evaporation of water vapor within a region should be physically continuous.

[0031] If a detected anomaly violates any physical constraint exceeding a preset tolerance (e.g., mass flux imbalance exceeding 10%, or energy conservation error exceeding 15%), a secondary verification mechanism is triggered. This mechanism automatically retrieves higher-resolution raw observation data and, in conjunction with an expert knowledge base, reassesses the physical plausibility of the anomaly to reduce the false alarm rate. For example, if the system detects significant water vapor accumulation at high altitudes without obvious convective uplift, this may violate water vapor continuity or the first law of thermodynamics. The system will perform a secondary verification to check for sensor malfunctions or data misreading.

[0032] The dominant factor attribution analysis employs the Integrated Gradients method to calculate the contribution of each input meteorological element to the anomaly score. The Integrated Gradients method is an interpretable AI technique that quantifies the contribution of each input feature to the model's prediction by calculating the integral of the gradient of the model output relative to the input features along the input path. Specifically, for each anomalous event, the Integrated Gradients method calculates the importance of specific physical quantities (such as temperature gradient, humidity, vertical wind shear, reflectivity factor, etc.) in satellite remote sensing data, ground observation station data, radar echo data, and numerical weather prediction model output data when the anomaly is detected. Then, combining prior meteorological knowledge (such as the need for sufficient water vapor, unstable stratification, and lifting mechanisms for strong convection), these contributions are normalized and ranked, ultimately outputting the top 5 dominant factors and their quantified contribution percentages. For example, for a strong convective anomaly, the dominant factors that may be output are: vertical gradient of radar reflectivity factor (30%), convective effective potential energy (25%), 850hPa water vapor flux (20%), surface gusts (15%), and 300hPa jet stream (10%).

[0033] Finally, the intelligent early warning decision module generates tiered early warning instructions based on the type, intensity, duration, and potential impact of the abnormal event and pushes them to the designated terminal.

[0034] The intensity of abnormal events is classified into four levels: mild, moderate, severe, and extremely severe, quantified based on the relative magnitude of the abnormality score and the threshold, as well as the potential physical impact. Duration is categorized as short-term (less than 2 hours), sustained (2-6 hours), and long-lasting (more than 6 hours). Potentially affected industries are matched according to the type of abnormality; for example, severe convective weather primarily affects transportation, power, and construction industries. The module maintains a pre-defined warning strategy library, which matches corresponding warning levels (blue, yellow, orange, red) and suggested response measures based on different combinations of abnormal events (type, intensity, duration, affected industries).

[0035] The warning instructions include the following core information: anomaly type code (e.g., 101 represents localized heavy rainfall, 102 represents thunderstorms and strong winds), precise geographic coordinates of the affected area (e.g., coordinates of polygon vertices or center point and radius), expected duration (e.g., expected to last until the next 2 hours), recommended response measures (e.g., recommending municipal departments to activate drainage plans, transportation departments to strengthen road patrols, and power departments to inspect transmission lines), and confidence score.

[0036] The confidence score is a key indicator for measuring the reliability of early warnings. It is calculated by weighting the anomaly detection score, physical consistency verification results, and accuracy rate of similar historical events. The specific calculation formula is as follows: in, Indicates the confidence score; It is the anomaly detection score Values ​​normalized to the [0,1] interval; It is the result of physical consistency verification (0 or 1, indicating whether the verification passed). It is the historical average detection accuracy of this type of abnormal event; , , The weighting coefficients are 0.5, 0.3, and 0.2, respectively. These weights reflect the relative importance of different information sources to the reliability of the early warning. The final generated early warning instructions are pushed to designated terminals such as meteorological bureaus, emergency management departments, traffic management departments, and power companies through multiple communication channels (such as SMS, APP push, and API interfaces) to ensure timely delivery of information.

[0037] Example 2 This embodiment aims to illustrate the application of the present invention in the detection and early warning of local drought anomalies in agricultural areas. Its core method and system architecture are consistent with those of Embodiment 1, but targeted optimizations have been made in data acquisition focus, feature extraction details, physical constraints and early warning strategies to adapt to the unique needs of agricultural production and the physical characteristics of drought events.

[0038] Firstly, in agricultural applications, the multi-source meteorological data fusion acquisition module significantly increases the weight given to soil moisture monitoring data and crop growth remote sensing indices, in addition to conventional satellite remote sensing data, ground observation station data, and numerical weather prediction data. Satellite remote sensing data not only includes visible and infrared bands but also emphasizes microwave remote sensing data to obtain soil moisture information penetrating surface vegetation. Ground observation station data, besides conventional meteorological elements, also includes data from soil temperature and humidity sensors deployed in farmland, with a time resolution set at 30 minutes. Radar echo data is used to monitor precipitation distribution in this scenario; its importance is second only to convective weather scenarios but still serves as an important supplement to precipitation input. Numerical weather prediction model output data focuses more on medium- to long-term precipitation and temperature forecasts for the next week or even month.

[0039] During the spatiotemporal alignment phase, the spatial reference also employs geographic coordinate gridding, but the grid granularity may be adjusted to 500m x 500m to meet the needs of refined farmland management. The temporal resolution is unified to a daily granularity to better capture the cumulative effects of drought. In the quality control phase, for ground soil temperature and humidity data, in addition to outlier removal, long-term drift correction is performed to address sensor aging issues. For crop growth indices (such as NDVI and EVI) acquired by satellite remote sensing, multi-period data smoothing filtering is used to eliminate instantaneous fluctuations caused by cloud cover or short-term weather changes, ensuring the accuracy of growth trends. Numerical model output data continues to use sliding window bias correction, but within the window, the focus is on total precipitation and average temperature.

[0040] Secondly, the meteorological feature deep encoding module still adopts a hybrid architecture of 3D convolutional neural network and graph attention network. However, the specific implementation of feature extraction has been adjusted. The 3D convolutional neural network mainly processes multi-timestep satellite remote sensing soil moisture, surface temperature, and crop growth index sequences to capture the long-term trends and spatial structure of soil moisture status, surface evapotranspiration, and vegetation cover changes, thereby extracting features reflecting drought severity, such as the standardized precipitation index (SPI) and evapotranspiration deficit index (ETDI). For example, the 3D CNN identifies low soil moisture areas that continuously expand in time and space. The graph attention network treats agricultural meteorological stations and soil moisture observation points as graph nodes, constructs edge weights based on the distance between stations and the similarity of soil types, dynamically aggregates soil moisture and crop growth information from neighboring stations, and generates agricultural meteorological feature vectors with spatial topological awareness capabilities, paying particular attention to the existence of contiguous drought zones spanning administrative regions. Finally, the outputs of 3D CNN and GAT are concatenated along the feature dimension and compressed into a 512-dimensional high-dimensional meteorological state representation vector through a fully connected layer. This vector focuses on reflecting the water and heat balance and vegetation health of the agricultural ecosystem.

[0041] Next, the anomaly detection and localization module identifies localized drought anomalies based on a variational autoencoder (VAE) architecture. The VAE encoder receives a 512-dimensional high-dimensional meteorological state representation vector, and the decoder reconstructs the water and heat balance and crop growth characteristics of normal agricultural areas. The anomaly score is defined as the weighted L2 distance between the input features and the reconstructed features, with weights... The weighting is dynamically allocated based on the physical sensitivity of various meteorological elements to crop growth and agricultural production. For example, characteristics such as soil moisture deficit, abnormally high surface temperature, and a continuous decline in crop growth index will receive higher weights. Dynamic thresholds. The system is adaptively adjusted based on the 95th percentile of historical 30-day data on soil moisture and cumulative precipitation in agricultural areas to adapt to varying water requirements during different growing seasons. Anomaly spatial extents are precisely located in drought-affected farmland areas through deconvolution upsampling and connected component analysis.

[0042] Furthermore, the anomaly cause interpretability analysis module performs physical consistency verification and source tracing for the detected drought anomalies. In this scenario, the physical constraint loss function focuses on: Water resource balance equation: the balance relationship between regional precipitation input, evapotranspiration output, surface runoff and groundwater recharge.

[0043] Vegetation water stress response mechanism: the physical relationship between crop growth index changes and soil moisture status.

[0044] Energy balance: The balance between surface radiation, sensible heat, and latent heat fluxes, reflecting changes in surface energy distribution under arid conditions.

[0045] If a detected drought event violates any physical constraint beyond the preset tolerance—for example, if the crop growth index rises abnormally under conditions of continuous drought and high evaporation—a secondary verification mechanism is triggered. This may involve retrieving higher-resolution SAR (Synthetic Aperture Radar) data for verification or combining it with agricultural survey reports. The dominant factor attribution is calculated using an integrated gradient method to determine the contribution of each input meteorological element to the drought anomaly score, and then normalized and ranked using prior knowledge of agrometeorology. For example, for a drought event, the possible dominant factors output might be: cumulative precipitation anomaly (40%), soil available moisture content (30%), potential evapotranspiration (15%), surface temperature (10%), and upper-level circulation anomaly (5%).

[0046] Finally, the intelligent early warning decision-making module matches a pre-set early warning strategy library based on the intensity level of the drought anomaly (e.g., mild drought, moderate drought, severe drought, extreme drought), duration (e.g., short-term drought <10 days, continuous drought 10-30 days, long-term drought >30 days), and affected industries (agriculture, water conservancy). The early warning instruction includes anomaly type code (e.g., 201 represents agricultural drought), precise farmland plot coordinates of the affected area, estimated duration, suggested countermeasures (e.g., suggesting agricultural departments activate drought relief plans, organize artificial rain enhancement, guide farmers to adjust irrigation plans or switch to drought-resistant crops), and a confidence score. The formula for calculating the confidence score is the same as in Example 1, but... This will correspond to the historical accuracy rate of agricultural drought. Early warning instructions will be pushed to agricultural management departments, water conservancy departments, and relevant farmers' terminals to guide agricultural production and minimize drought losses.

[0047] Example 3 This embodiment applies the technical solution of the present invention to the intelligent detection and judgment of icing anomalies on power transmission lines in the power industry. Icing on transmission lines is a common and severe meteorological disaster in winter, which can lead to transmission line breaks, tower collapses, and large-scale power outages. Based on the aforementioned system architecture, this embodiment focuses on customized optimization in data acquisition, feature encoding, and physical constraint verification to improve the detection accuracy and timely warning of icing anomalies.

[0048] Firstly, in the power industry application of the multi-source meteorological data fusion acquisition module, in addition to conventional meteorological data sources, the focus is on strengthening the acquisition of monitoring data of the transmission line itself and high-resolution topographic data. Monitoring data of the transmission line itself includes real-time image data captured by icing thickness sensors, tilt sensors, tension sensors, temperature sensors, and high-resolution cameras installed on the transmission lines. Ground observation station data is densely deployed along the transmission lines, including not only conventional meteorological elements but also paying special attention to parameters directly related to icing, such as dew point temperature, wind speed and direction, and humidity, with a time resolution of up to 5 minutes. Satellite remote sensing data is used to obtain the temperature and humidity field distribution over large-scale areas. Numerical weather prediction model output data focuses on forecast information such as 0°C layer height, humidity vertical profile, and wind field vertical shear. High-resolution topographic data (such as digital elevation models, DEMs) is used to accurately characterize the topographic undulations of the areas traversed by the transmission lines, which is crucial for understanding the impact of local microclimates on icing.

[0049] After data acquisition, during the spatiotemporal alignment phase, the system uses a linear gridded spatial reference based on the transmission line alignment to map all data onto a unified grid along the line direction, with a spatial resolution of 100 meters. The temporal resolution is unified to a 5-minute granularity. In the quality control phase, for icing sensor data, in addition to isolated point removal, long-term zero-point drift correction is performed. For transmission line image data, image enhancement and denoising algorithms are used to ensure the clarity of icing morphology features. Radar echo data, after clutter filtering, is used to determine whether liquid precipitation or freezing rain has occurred. Numerical model output data undergoes sliding window bias correction, with a focus on temperature forecast bias near the freezing point.

[0050] Secondly, the meteorological feature deep encoding module continues to employ a hybrid architecture of 3D convolutional neural networks and graph attention networks. In this scenario, the 3D convolutional neural network primarily processes multi-timestep transmission line image data, icing thickness sensor sequences, and high-resolution terrain data to capture the dynamic development process of transmission line icing, icing morphology (such as rime and frost), and the impact of terrain on airflow and temperature. For example, the 3D CNN identifies the process of icing accumulating locally and expanding into segments, as well as the temperature gradient changes at the icing sites. The graph attention network treats meteorological observation points and transmission line sensor nodes along the line as graph nodes, constructing edge weights based on geographical distance between nodes and electrical connections of the transmission line. It dynamically aggregates the icing status and environmental meteorological information of neighboring nodes to generate a power line icing feature vector with spatial topology awareness. For example, when assessing its own icing risk, a node considers the icing situation of upstream line nodes and the impact of local wind fields on ice crystal transport. Finally, the outputs of 3D CNN and GAT are concatenated along the feature dimension and compressed into a 512-dimensional high-dimensional meteorological state representation vector through a fully connected layer. This vector comprehensively reflects the icing state of the transmission line itself, the environmental meteorological conditions along the line, and their interactions.

[0051] Next, the anomaly detection and localization module identifies icing anomalies on transmission lines based on a variational autoencoder (VAE) architecture. The VAE encoder receives a 512-dimensional high-dimensional meteorological state representation vector, and the decoder reconstructs the transmission line features under normal, non-icing or slightly iced conditions. The anomaly score is defined as the weighted L2 distance between the input features and the reconstructed features, with weights... Dynamically assigning thresholds based on the physical sensitivity of various meteorological elements and line parameters to icing risk. For example, characteristics such as icing thickness, conductor tension, wind speed, and dew point temperature will receive higher weights. The system is adaptively adjusted based on the 95th percentile of historical 30-day transmission line icing data (ice thickness, conductor galloping amplitude) to adapt to icing characteristics in different regions and seasons. When anomaly scores exceed a dynamic threshold... When the event occurs, it is determined to be an abnormal event of icing on the transmission line. The abnormal spatial range is precisely located in the transmission line tower section affected by icing through deconvolution upsampling and connected component analysis.

[0052] Furthermore, the anomaly cause interpretability analysis module performs physical consistency verification and source tracing for detected icing anomalies. In this scenario, the physical constraint loss function focuses on: Physical model for ice growth: The growth rate of ice thickness should be consistent with the physical relationship between parameters such as temperature, humidity, wind speed, and liquid water content.

[0053] Conductor mechanical balance: The increase in conductor tension and vertical load caused by icing should conform to the principle of mechanical balance.

[0054] Micro-topography and microclimate effects: The influence of specific topography (such as ridges and wind gaps) on local temperature, humidity, and wind speed should be consistent with the detected icing anomalies.

[0055] If a detected icing event violates any physical constraint beyond the preset tolerance—for example, if rapid icing is detected even when the temperature is significantly above freezing—a secondary verification mechanism is triggered. This mechanism may retrieve drone inspection images for visual confirmation or combine them with high-resolution wind field simulation data for verification. The dominant factor tracing uses an integrated gradient method to calculate the contribution of each input meteorological element to the icing anomaly score, and normalizes and ranks it using prior knowledge of transmission line icing meteorology. For example, for a single icing event, the possible output dominant factors are: transmission line surface temperature (35%), wind speed (25%), relative humidity (20%), 0°C layer height (10%), and terrain height (10%).

[0056] Finally, the intelligent early warning decision module matches a pre-set early warning strategy library based on the intensity level of the icing anomaly (e.g., light icing, moderate icing, heavy icing, severe icing), duration (e.g., short-term <3 hours, continuous 3-12 hours, long-term >12 hours), and the affected industry (power). The early warning instruction includes anomaly type code (e.g., 301 representing line icing), the specific tower number or section coordinates of the affected transmission line, the estimated duration, suggested countermeasures (e.g., suggesting the power company initiate de-icing, strengthen line patrols, adjust load, and activate emergency power), and a confidence score. The formula for calculating the confidence score is the same as in Example 1, but... This will correspond to the historical accuracy rate of transmission line icing events. Early warning instructions will be pushed to the power dispatch center, maintenance teams, and user terminals in affected areas via SCADA systems, SMS, and mobile apps, providing timely assurance for the safe and stable operation of the power system.

[0057] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. An AI-based regional weather anomaly intelligent detection and judgment system, characterized in that, The application relates to a weather anomaly detection and positioning method based on multi-source meteorological data fusion. The application comprises the following components: A multi-source meteorological data fusion acquisition module is used for acquiring satellite remote sensing data, ground observation station data, radar echo data and numerical weather prediction model output data of a target region in real time, and performing space-time alignment and quality control on the multi-source data; A meteorological feature deep coding module is used for performing multi-level feature extraction on the fused multi-source meteorological data, and generating a high-dimensional meteorological state representation containing thermodynamic, kinetic and microphysical process information; An anomaly detection and positioning module is used for identifying the occurrence probability, spatial range and intensity level of a regional weather anomaly event by a pre-trained spatio-temporal anomaly perception neural network based on the high-dimensional meteorological state representation; An anomaly cause explainability analysis module is used for performing physical consistency verification and dominant factor tracing on the detected anomaly event, and outputting a quantitative contribution degree ranking of the anomaly cause; An intelligent early warning decision module is used for generating a graded early warning instruction according to the type, intensity, duration and potential influence range of the anomaly event, and pushing the instruction to a specified terminal; 2.The AI-based regional weather anomaly intelligent detection and judgment system according to claim 1, characterized in that, The space-time alignment adopts a unified space reference based on geographical coordinate gridding, and the time resolution is unified to 10-minute granularity; the quality control includes isolated point elimination on the ground observation station data, ground object clutter filtering on the radar echo data and bias correction on the numerical model output data; and the high-dimensional meteorological state representation is a 512-dimensional vector. 3.The AI-based regional weather anomaly intelligent detection and judgment system according to claim 1, characterized in that, The meteorological feature deep coding module adopts a hybrid architecture of a three-dimensional convolutional neural network and a graph attention network; the three-dimensional convolutional neural network processes radar and satellite data in a space-time cube form, and extracts convective structure features in the vertical direction; the graph attention network regards ground observation stations as graph nodes, constructs edge weights according to geographical distances and meteorological similarities between stations, dynamically aggregates neighborhood station information, and generates a ground meteorological feature vector with spatial topology perception ability; finally, the three-dimensional convolution output and the graph attention output are spliced in the feature dimension, and compressed into a 512-dimensional high-dimensional meteorological state representation vector through a full connection layer. The spatio-temporal anomaly perception neural network used by the anomaly detection and positioning module is constructed based on a variational autoencoder architecture, the encoder receives the high-dimensional meteorological state representation vector, and the decoder reconstructs meteorological features under normal weather state; the anomaly score is defined as the weighted L2 distance between the input features and the reconstructed features, and the weight is dynamically allocated according to the physical sensitivity of each meteorological element; when the anomaly score exceeds a dynamic threshold tau, the anomaly event is determined, wherein tau is adaptively adjusted according to the 95% quantile of the same period data in the past 30 days; the abnormal spatial range is mapped back to the original geographical grid through deconvolution upsampling, and the connected domain analysis is adopted to determine the abnormal region boundary. 4.The AI-based regional weather anomaly intelligent detection and judgment system according to claim 3, characterized in that, The abnormality score The formula for the calculation of the abnormality score is as follows: in, For the input high-dimensional meteorological state representation vector, the first... One portion, The corresponding component reconstructed by the decoder. In order to target the The weights of the physical sensitivity of each meteorological element are dynamically allocated. To balance the reconstruction loss with the hyperparameters of the potential space regularization term, The latent distribution of the encoder output is related to the standard normal prior distribution. Divergence. 5.The AI-based regional weather anomaly intelligent detection and judgment system according to claim 1, characterized in that, The anomaly cause explainability analysis module verifies the anomaly detection result by constructing a meteorological physical constraint loss function, the physical constraint includes mass conservation, the first law of thermodynamics and water vapor continuity equation; if the detected abnormal event violates any physical constraint more than a preset tolerance, a secondary review mechanism is triggered; the dominant factor tracing uses integrated gradient method to calculate the contribution of each input meteorological element to the abnormal score, and combines meteorological prior knowledge to normalize and sort the contribution, and outputs the top 5 dominant factors and their quantitative contribution percentage. 6.The AI-based regional weather anomaly intelligent detection and judgment system according to claim 1, characterized in that, The intelligent early warning decision module is divided into four levels of mild, moderate, severe and extremely severe according to the intensity level of the abnormal event, the duration is divided into short time less than 2 hours, continuous 2 to 6 hours, long continuous more than 6 hours, and the influence industry includes agriculture, transportation, power and water conservancy; The early warning instruction includes abnormal type code, influence area coordinate, expected duration, suggested countermeasures and confidence score; the confidence score is calculated by weighting the anomaly detection score, physical consistency verification result and historical similar event accuracy, and the weights are 0.5, 0.3 and 0.2 respectively.

7. A detection and judgment method applied to the AI-based regional weather anomaly intelligent detection and judgment system of any one of claims 1-6. The steps include: Step S110, acquiring satellite remote sensing data, ground observation station data, radar echo data and numerical weather prediction model output data of the target area in real time through the multi-source meteorological data fusion acquisition module, and performing time and space alignment and quality control on the multi-source data; Step S120, multi-level feature extraction is performed on the fused multi-source meteorological data through the meteorological feature deep coding module, generating a 512-dimensional high-dimensional meteorological state representation containing thermodynamic, dynamic and microphysical process information; Step S130, based on the high-dimensional meteorological state representation, the occurrence probability, spatial range and intensity level of regional weather abnormal events are identified by using a pre-trained spatio-temporal anomaly perception neural network through the anomaly detection and positioning module; Step S140, the physical consistency verification and dominant factor tracing of the detected abnormal event are performed through the anomaly cause explainability analysis module, and the quantitative contribution degree of the abnormal cause is sorted; Step S150, the hierarchical early warning instruction is generated according to the type, intensity, duration and potential influence range of the abnormal event through the intelligent early warning decision module, and is pushed to the specified terminal. 8.The AI-based regional weather anomaly intelligent detection and judgment method according to claim 7, characterized in that, In step S120, a three-dimensional convolutional neural network is used to process radar and satellite data in a spatio-temporal cube form to extract convective structure features in the vertical direction; At the same time, the ground observation station is regarded as a graph node by using a graph attention network, the edge weight is constructed according to the geographical distance and meteorological similarity between stations, the neighborhood station information is dynamically aggregated, and a ground meteorological feature vector with spatial topology perception ability is generated; finally, the three-dimensional convolution output and the graph attention output are spliced in the feature dimension, and compressed to a 512-dimensional high-dimensional meteorological state representation vector through a fully connected layer. 9.The AI-based regional weather anomaly intelligent detection and judgment method according to claim 7, characterized in that, In step S150, the confidence score wherein, is the normalized anomaly detection score, is the physical consistency verification result (0 or 1), is the average detection accuracy of this type of anomaly event in history.

Images