Intelligent generation method of grid-level short-term and imminent forecast for severe convective weather

Abstract

The invention relates to the technical field of weather forecasting, in particular to an intelligent generation method for grid-level short and temporary forecasting for severe convective weather, which comprises the following steps: step 1, data acquisition: acquiring meteorological radar, satellite remote sensing, Internet of Things sensors, social media texts and urban infrastructure monitoring data in real time; step 2, model construction: taking a space-time diagram convolutional network as a backbone network, inputting a space-time feature tensor, and introducing a physical constraint term; step 3, generating a result: outputting a grid-level severe convection weather element distribution diagram in the next 0-2 hours by utilizing the prediction model; 4, feedback and optimization: dynamically calculating a correction factor based on the difference between the real-time monitoring data of the user side and the forecast result; and dynamically deploying the updated model to an edge computing node. According to the method, through physical constraint space-time diagram convolution modeling and user feedback driven edge adaptive optimization, an efficient, robust and landing intelligent solution is provided for severe convection weather short-term and imminent forecasting.

Description

Technical Field

[0001] This invention relates to the field of meteorological forecasting technology, specifically to a grid-level short-term forecasting intelligent generation method for severe convective weather. Background Technology

[0002] Severe convective weather is a sudden and destructive weather phenomenon triggered by small- to medium-scale weather systems. It is often accompanied by extreme weather events such as thunderstorms, strong winds, short-duration heavy rainfall, and hail. It is characterized by strong spatial locality, rapid evolution, and significant destructive potential. Traditional numerical weather prediction models are limited by computational timeliness and parameterization schemes, making it difficult to meet the requirements for refined early warning at the 0-2 hour, 1km grid level. Meanwhile, short-term forecasting methods based on radar extrapolation are insufficient in capturing the evolution patterns of weather systems under complex terrain, resulting in insufficient early warning accuracy for disasters such as urban flooding and lightning strikes on power transmission networks.

[0003] Existing technologies suffer from the following main shortcomings: First, insufficient dynamic fusion capability of heterogeneous data sources. Existing methods typically rely on traditional observational data such as weather radar and satellite remote sensing, but lack effective mechanisms for multimodal fusion of data from IoT sensors, social media text, and urban infrastructure monitoring. For example, urban flood risk assessment relies solely on fixed rain gauges, failing to integrate multi-dimensional real-time data such as water depth identification from traffic monitoring videos and underground pipe network water level sensors, resulting in dynamic assessment errors exceeding 40% for localized flood risk. Second, insufficient integration of physical mechanisms and data-driven approaches. Existing deep learning models often employ extrapolation of single elements (such as radar echoes) or single-task prediction (such as precipitation), without explicitly modeling the physical coupling relationship between wind, rain, and thunder. Research indicates that neglecting the driving effect of wind field dynamics on the movement of rain clouds increases the mean square error of 2-hour precipitation prediction by 15-20%. Third, lack of closed-loop optimization mechanisms for operational scenarios. Existing forecasting systems are mostly one-way generation models, lacking a dynamic adjustment mechanism driven by user-end measured data feedback. In practical deployments, it has been found that the forecast error of traditional models in complex urban environments accumulates by more than 30% over time, and they cannot adaptively adjust accuracy based on the resource constraints of edge computing nodes. These shortcomings severely restrict the accuracy and real-time response capability of severe convective weather warnings. Therefore, there is an urgent need to construct a new intelligent forecasting method that integrates physical constraints, dynamic fusion of multi-source data, and edge adaptive optimization. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides a grid-level intelligent generation method for short-term forecasts of severe convective weather. Through physically constrained spatiotemporal graph convolutional modeling and user feedback-driven edge adaptive optimization, it offers an efficient, robust, and practical intelligent solution for short-term forecasting of severe convective weather.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows: A grid-level short-term forecast intelligent generation method for severe convective weather, comprising the following steps:

[0006] Step 1, Data Acquisition: Real-time acquisition of meteorological radar, satellite remote sensing, IoT sensors, social media text and urban infrastructure monitoring data; normalization of multi-source data based on spatiotemporal alignment module; calculation of credibility weight of each data source through dynamic weight allocation mechanism; and filtering of abnormal data based on spatiotemporal continuity test and physical feasibility verification to generate fused spatiotemporal feature tensor.

[0007] Step 2, Model Construction: Using a spatiotemporal graph convolutional network as the backbone network, the spatiotemporal feature tensor is input. By introducing the wind field dynamics equation, rainfall conservation equation and lightning energy transfer equation as physical constraints, the loss function of the deep learning model is jointly optimized to explicitly model the spatiotemporal synergistic effect of the three elements of wind, rain and lightning.

[0008] Step 3, Generate Results: Utilize the prediction model to output a grid-level distribution map of severe convective weather elements for the next 0-2 hours, including wind speed, precipitation, and lightning intensity. Then, use an autoencoder to perform dimensionality reduction and reconstruction of the predicted features, and combine residual connections to retain detailed information of the original radar echoes.

[0009] Step 4, Feedback and Optimization: Based on the difference between real-time monitoring data and forecast results from the user end, dynamically calculate the correction factor, update the model parameters through backpropagation, and trigger dynamic adjustments to data fusion weights and physical constraint terms; dynamically deploy the updated model to edge computing nodes, and select the accuracy mode for model inference based on the real-time computing resources of the user end.

[0010] Furthermore, in step one, the dynamic weight allocation mechanism dynamically adjusts the weights of each data source based on the spatiotemporal resolution, historical accuracy, and real-time anomaly detection results of the data source using an adaptive Kalman filter algorithm. Among them, social media text data participates in the fusion after extracting geotags and weather keywords through natural language processing.

[0011] Furthermore, in step one, the urban infrastructure monitoring data includes: water depth identification results from traffic monitoring videos, underground pipe network water level sensor data, and urban heat island effect infrared remote sensing data, which are then fused after spatial distribution features are extracted through a convolutional neural network.

[0012] Furthermore, in step one, the abnormal data filtering method is as follows: for noisy data in IoT sensors and false information in social media, a threshold method is used for preliminary screening, and then the spatiotemporal continuity constraint is modeled by Markov random field to jointly determine and remove abnormal data.

[0013] Furthermore, in step two, the spatiotemporal convolutional network is composed of a stack of forward temporal convolutional layers, graph convolutional layers, and backward temporal convolutional layers. The graph convolutional layers use Chebyshev polynomial approximation to achieve nonlocal spatial dependency modeling, while the temporal convolutional layers use causal convolution and gated linear units to capture multi-scale temporal features.

[0014] Furthermore, in step two, in the spatiotemporal map convolutional layer, channel attention weights are generated through global max pooling and average pooling, and spatial attention weights are generated using a 7×7 convolutional layer. The channel attention weights and spatial attention weights are multiplied element-wise and then weighted and fused with the original feature map to enhance the extraction of spatiotemporal features in the strong echo region.

[0015] Furthermore, in step two, during the model training phase, the physical constraints include the divergence error of the wind field dynamics equation, the mass balance error of the rainfall conservation equation, and the electric field intensity error of the lightning energy transfer equation as additional loss terms, which are jointly optimized with the prediction error to constrain the physical consistency of the model output.

[0016] Furthermore, in step three, the autoencoder consists of fully connected layers that compress the predicted features into a low-dimensional latent space; the decoder reconstructs a high-resolution feature map through residual connections, retaining the detailed information of the original radar echo, and outputs a prediction grid with a resolution of no less than 1km×1km.

[0017] Furthermore, in step four, the correction factor is calculated as follows: based on the mean square error between the measured data and the forecast results fed back by the user, combined with the time decay function and the spatial correlation matrix, a grid-by-grid correction coefficient is generated, and the model parameters are updated online through the lightweight fine-tuning module.

[0018] Furthermore, in step four, the precision mode includes the full-parameter mode, the quantization compression mode, and the lightweight mode generated by knowledge distillation. The precision mode uses the full-parameter mode to perform high-precision inference with the complete STGCN model, the quantization compression mode reduces the computational load of the model by using 8-bit fixed-point numbers, and the lightweight mode simplifies the network structure based on knowledge distillation. Each mode is dynamically switched according to the user's terminal resources.

[0019] The above approach has the following beneficial effects:

[0020] 1. This scheme employs a dynamic weight allocation mechanism based on the spatiotemporal resolution of the data source, historical accuracy, and real-time anomaly detection results. It uses an adaptive Kalman filter algorithm to dynamically adjust the weights, effectively addressing the problem of insufficient utilization of emerging data sources (such as social media) by traditional fixed weight methods, thereby enhancing the flexibility and reliability of data fusion. Anomaly filtering is jointly determined by a threshold method and Markov random field spatiotemporal continuity constraints, significantly reducing the impact of IoT sensor noise and social media misinformation, ensuring the physical rationality and spatiotemporal consistency of the input data, and providing a high-quality data foundation for model construction.

[0021] 2. This scheme employs physical constraint modeling by introducing error terms from the wind field dynamics equation, precipitation conservation equation, and lightning energy transfer equation as additional losses. This explicitly constrains the physical consistency of the model output, overcoming the "false prediction" problem of traditional pure data-driven models in the evolution of severe convective systems, and improving the predictive scientific accuracy of the coordinated evolution of wind speed, precipitation, and lightning. The spatiotemporal graph convolutional network, combined with a dual attention mechanism, uses nonlocal graph convolution to model global spatial dependencies and causal convolution to capture multi-scale temporal features. Furthermore, it dynamically focuses on strong echo regions through a channel-space dual attention mechanism, significantly improving the ability to extract spatiotemporal features from the core areas of severe convective weather.

[0022] 3. This scheme preserves the detailed information of radar echoes (such as echo boundaries and intensity gradients) during the dimensionality reduction process by compressing the low-dimensional potential space and reconstructing high-resolution feature maps. This avoids the forecast ambiguity caused by feature loss in traditional methods and achieves refined output of 1km×1km grid-level forecasts.

[0023] 4. In this solution, user feedback-driven dynamic correction generates a grid-by-grid correction factor based on the difference between measured data and forecast results. A lightweight fine-tuning module enables online updates of model parameters, ensuring continuous model optimization in business scenarios. Edge computing multi-mode deployment dynamically switches between full-parameter mode, quantization compression mode, and lightweight knowledge distillation mode, balancing high-precision inference with low latency requirements. It is adaptable to various application scenarios, including meteorological bureau servers, mobile terminals, and drones, enhancing the system's practicality and flexibility.

[0024] 5. This solution provides reliable input to the physical constraint model through high-quality data fusion. The model output dynamically optimizes data weights and physical constraint terms based on user feedback, forming an adaptive technical closed loop. The lightweight mode and quantization compression technology reduce the computing load at the edge. Combined with the feedback mechanism, it achieves efficient collaboration between "cloud training and edge inference" to meet the real-time requirements of severe convective weather warnings.

[0025] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0026] Figure 1 This is a schematic diagram illustrating the method steps of an embodiment of the grid-level short-term forecast intelligent generation method for severe convective weather according to the present invention. Detailed Implementation

[0027] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0029] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0030] The following detailed description illustrates the specific implementation method:

[0031] Example:

[0032] As attached Figure 1 As shown: A grid-level short-term forecast intelligent generation method for severe convective weather includes the following steps:

[0033] Step 1, Data Acquisition: Real-time acquisition of data from meteorological radar, satellite remote sensing, IoT sensors, social media texts, and urban infrastructure monitoring data. Urban infrastructure monitoring data includes: water depth identification results from traffic monitoring videos, underground pipe network water level sensor data, and urban heat island effect infrared remote sensing data. Spatial distribution features are extracted through convolutional neural networks and then fused.

[0034] The multi-source data is normalized based on the spatiotemporal alignment module. The credibility weight of each data source is calculated through a dynamic weight allocation mechanism. The dynamic weight allocation mechanism dynamically adjusts the weight of each data source based on the spatiotemporal resolution, historical accuracy, and real-time anomaly detection results of the data source, using an adaptive Kalman filter algorithm.

[0035]

[0036] in, This represents the weight of data source k at time t, used to dynamically reflect the reliability of the data source; K t Represented as a Kalman gain matrix, the weight update magnitude is dynamically adjusted. When the data source fluctuates greatly, K... t Increase to respond quickly to changes; when the data source tends to stabilize, K t Reduce to smooth out noise; Let H represent the real-time confidence score of data source k at time t, which is calculated by weighting the spatiotemporal resolution, historical accuracy, and real-time anomaly rate; H represents the observation matrix, which is used to map the weights to the confidence score space (H=1 in the simplified scenario).

[0037] The social media text data was fused after extracting geotags and weather keywords through natural language processing. Abnormal data was filtered based on spatiotemporal continuity checks and physical feasibility verification to generate a fused spatiotemporal feature tensor. The abnormal data filtering method was as follows: noise data from IoT sensors and false information from social media were initially screened using a threshold method, and then spatiotemporal continuity constraints were modeled using Markov random fields to jointly determine and remove abnormal data.

[0038] Step 2, Model Construction: The spatiotemporal graph convolutional network is used as the backbone network. The spatiotemporal convolutional network is composed of forward temporal convolutional layers, graph convolutional layers, and backward temporal convolutional layers stacked together. The graph convolutional layers use Chebyshev polynomial approximation to achieve nonlocal spatial dependency modeling. The temporal convolutional layers use causal convolution and gated linear units to capture multi-scale temporal features. In the spatiotemporal graph convolutional layers, channel attention weights are generated through global max pooling and average pooling, and spatial attention weights are generated using 7×7 convolutional layers. The channel attention weights and spatial attention weights are multiplied element-wise and then weighted and fused with the original feature map to enhance the extraction of spatiotemporal features in strong echo regions.

[0039] The input spatiotemporal feature tensor is used, and the wind field dynamics equation, rainfall conservation equation, and lightning energy transfer equation are introduced as physical constraints. During the model training phase, the divergence error of the wind field dynamics equation, the mass balance error of the rainfall conservation equation, and the electric field intensity error of the lightning energy transfer equation are used as additional loss terms, which are jointly optimized with the prediction error to constrain the physical consistency of the model output. The loss function of the deep learning model is jointly optimized to explicitly model the spatiotemporal synergistic effect of wind, rain, and lightning, specifically:

[0040] L total =L pred +λ1L div +λ1L mass +λ3L energy

[0041] Among them, L pred To predict losses, the model output (wind, rain, thunder) is used to calculate the difference between the actual observations and the model output, ensuring that the model prediction results match the actual observation data. λ1, λ2, and λ3 represent the physical constraint weights of the three elements of wind, rain, and thunder, respectively. Specifically, in this embodiment, the values ​​are 0.2, 0.15, and 0.1, respectively.

[0042] The formula for calculating wind field divergence loss is:

[0043]

[0044] In the formula, For a three-dimensional wind speed field, Let the wind field divergence in three-dimensional space be the incompressible flow assumption based on the Navier-Stokes equations:

[0045]

[0046] The rainfall mass conservation loss equation is:

[0047]

[0048] Where ρ represents the rainwater density (kg / m³) 3 ).

[0049] The lightning energy loss equation is:

[0050] L energy =||Ek·CAPE|| 2

[0051] Where E represents the model prediction of the electric field strength, CAPE represents the convective effective potential energy calculated from meteorological sounding data, and k is the scaling factor fitted using historical data.

[0052] Step 3, Generate Results: Utilize the prediction model to output a grid-level distribution map of severe convective weather elements for the next 0-2 hours, including wind speed, precipitation, and lightning intensity. Then, use an autoencoder to perform dimensionality reduction and reconstruction of the predicted features, and combine residual connections to retain detailed information of the original radar echoes.

[0053] The autoencoder consists of fully connected layers that compress the predicted features into a low-dimensional latent space; the decoder reconstructs a high-resolution feature map through residual connections, preserving the detailed information of the original radar echo, and outputs a prediction grid with a resolution of not less than 1km.

[0054] Step 4, Feedback and Optimization: Based on the difference between the real-time monitoring data and the forecast results from the user terminal, the correction factor is dynamically calculated. The correction factor is calculated as follows: based on the mean square error between the measured data and the forecast results fed back from the user terminal, combined with the time decay function and the spatial correlation matrix, the correction coefficients for each grid point are generated, and the model parameters are updated online through the lightweight fine-tuning module.

[0055] The model parameters are updated through backpropagation, triggering dynamic adjustments to data fusion weights and physical constraints. The updated model is dynamically deployed to edge computing nodes, and the model inference precision mode is selected based on the real-time computing resources of the user end. The precision modes include full parameter mode, quantization compression mode, and lightweight mode generated by knowledge distillation. The precision mode uses the full parameter mode to perform high-precision inference with the complete STGCN model, the quantization compression mode reduces the model's computational load by using 8-bit fixed-point numbers, and the lightweight mode simplifies the network structure based on knowledge distillation. Each mode is dynamically switched according to the user end resources.

[0056] The specific implementation process is as follows: Taking Shanghai meteorological data collection as an example, data is collected and weights are assigned to each data point:

[0057] Weather radar: S-band radar data is collected by the meteorological bureau, with a time resolution of 6 minutes, a spatial resolution of 1 km, a reflectivity range of 0–70 dBZ, and an initial weight of 0.85.

[0058] Satellite remote sensing: acquire infrared cloud images from Fengyun-4 and identify convective cloud clusters with cloud top brightness temperatures below -60℃;

[0059] IoT Sensors: Integrates 100 underground pipe network water level sensors in Shanghai (uploading data once per second) to detect local water accumulation risk, with a weight of 0.7;

[0060] Social media text: Real-time capture of geotagged posts (e.g., "Severe flooding on Pudong XX Road") containing keywords "Shanghai rainstorm" and "thunderstorm" from social media platforms, with an initial weight of 0.15;

[0061] Urban infrastructure data: Traffic surveillance video was analyzed, and the depth of water accumulation areas was detected using YOLOv5 (accuracy ±5cm). The water accumulation depth features were extracted from the traffic surveillance video using CNN (ResNet-18), and the heat island effect data was segmented using U-Net to divide the heat distribution.

[0062] When weather radar data is missing due to obstruction, the sensor weight is automatically increased to 0.8. When a sensor suddenly reports a water depth of 3.2m (physically infeasible), it is directly rejected using the threshold method (water depth threshold is 0-2m). Dynamic weight allocation reduces the fusion error to 8.2%.

[0063] The spatiotemporal graph convolutional layer employs causal convolution and gating mechanisms to capture the frame-to-frame dependencies of radar echo sequences; the graph convolutional layer models nonlocal spatial dependencies based on Chebyshev multinomials (covering a radius of 15km), breaking through the local field-of-view limitations of traditional CNNs; channel attention focuses on key meteorological elements (such as strong echoes), and spatial attention enhances the extraction of details in the core convective region.

[0064] In model training, in addition to prediction error, wind field divergence error, rainfall mass conservation error, and lightning energy transfer error are introduced as constraint terms. For example, the physical constraint term is expressed as:

[0065] Wind field divergence loss: The velocity field divergence error is calculated based on the Navier-Stokes equations, with a weight λ1 = 0.2;

[0066] Rainfall conservation loss: The mass balance error is calculated by the difference between the cumulative rainfall intensity and the evaporation rate, λ2 = 0.15;

[0067] Lightning energy loss: correlation constraint between electric field strength and CAPE exponent, λ3=0.1;

[0068] The forced model predicts rainfall distribution that conforms to the law of conservation of mass, thus avoiding the generation of "false rain clusters".

[0069] Based on the discrepancy between the actual precipitation data from the user's end and the forecast results, and combining time decay and spatial correlation, a grid-by-grid correction coefficient is generated. For example, if the predicted precipitation in the Pudong area is 50 mm and the actual measured precipitation is 60 mm, the weight of subsequent forecasts for that area will be dynamically increased.

[0070] The model set up in this scheme is used as the experimental group and compared with the traditional optical flow method in the existing technology as the control group. The dataset is the severe convective radar data (40,000 samples) of Shanghai from 2017 to 2020. The evaluation indicators include: rainfall CSI (critical success index), wind speed MAE (mean absolute error), lightning time error, GPU memory usage and edge inference latency.

[0071]

[0072]

[0073] Conclusion: Experimental results show that this method significantly improves the accuracy and efficiency of short-term forecasts for severe convective weather through the synergistic effect of dynamic fusion of multi-source heterogeneous data, physical constraint modeling, and closed-loop feedback optimization. Compared with traditional methods, the critical success index (CSI) for precipitation prediction increased from 0.62 to 0.83, the mean absolute error of wind speed decreased by 24%, and the time error of lightning occurrence decreased by 51.7%. Simultaneously, through a lightweight edge computing strategy, inference latency was reduced to 68 milliseconds (a 45.6% reduction compared to traditional methods), and memory usage was optimized by 17.3%. In a pilot project in Shanghai, severe convective weather warnings were issued 58 minutes in advance, verifying the significant advantages of this method in improving forecast precision, physical rationality, and real-time response capabilities.

[0074] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A grid-level short-term forecast intelligent generation method for severe convective weather, characterized in that, Includes the following steps: Step 1, Data Acquisition: Real-time acquisition of meteorological radar, satellite remote sensing, IoT sensors, social media text and urban infrastructure monitoring data; normalization of multi-source data based on spatiotemporal alignment module; calculation of credibility weight of each data source through dynamic weight allocation mechanism; and filtering of abnormal data based on spatiotemporal continuity test and physical feasibility verification to generate fused spatiotemporal feature tensor. Step 2, Model Construction: Using a spatiotemporal graph convolutional network as the backbone network, the spatiotemporal feature tensor is input. By introducing the wind field dynamics equation, rainfall conservation equation and lightning energy transfer equation as physical constraints, the loss function of the deep learning model is jointly optimized to explicitly model the spatiotemporal synergistic effect of the three elements of wind, rain and lightning. Step 3, Generate Results: Utilize the prediction model to output a grid-level distribution map of severe convective weather elements for the next 0-2 hours, including wind speed, precipitation, and lightning intensity. Then, use an autoencoder to perform dimensionality reduction and reconstruction of the predicted features, and combine residual connections to retain detailed information of the original radar echoes. Step 4, Feedback and Optimization: Based on the difference between real-time monitoring data and forecast results from the user end, dynamically calculate the correction factor, update the model parameters through backpropagation, and trigger dynamic adjustments to data fusion weights and physical constraint terms; dynamically deploy the updated model to edge computing nodes, and select the accuracy mode for model inference based on the real-time computing resources of the user end.

2. The intelligent generation method for gridded short-term forecasts of severe convective weather according to claim 1, characterized in that, In step one, the dynamic weight allocation mechanism dynamically adjusts the weights of each data source based on the spatiotemporal resolution, historical accuracy, and real-time anomaly detection results of the data source using an adaptive Kalman filter algorithm. Among them, social media text data participates in the fusion after extracting geotags and weather keywords through natural language processing.

3. The intelligent generation method for gridded short-term forecasts of severe convective weather according to claim 2, characterized in that, In step one, the urban infrastructure monitoring data includes: water depth identification results from traffic monitoring videos, underground pipe network water level sensor data, and urban heat island effect infrared remote sensing data. These data are then fused together after spatial distribution features are extracted using a convolutional neural network.

4. The intelligent generation method for gridded short-term forecasts of severe convective weather according to claim 3, characterized in that, In step one, the abnormal data filtering method is as follows: for noisy data in IoT sensors and false information in social media, a threshold method is used for preliminary screening, and then the spatiotemporal continuity constraint is modeled by Markov random field to jointly determine and remove abnormal data.

5. The intelligent generation method for gridded short-term forecasts of severe convective weather according to claim 4, characterized in that, In step two, the spatiotemporal convolutional network is composed of a stack of forward temporal convolutional layers, graph convolutional layers, and backward temporal convolutional layers. The graph convolutional layers use Chebyshev polynomial approximation to achieve nonlocal spatial dependency modeling, while the temporal convolutional layers use causal convolution and gated linear units to capture multi-scale temporal features.

6. The intelligent generation method for gridded short-term forecasts of severe convective weather according to claim 5, characterized in that, In step two, in the spatiotemporal map convolutional layer, channel attention weights are generated through global max pooling and average pooling, and spatial attention weights are generated using a 7×7 convolutional layer. The channel attention weights and spatial attention weights are multiplied element-wise and then weighted and fused with the original feature map to enhance the extraction of spatiotemporal features in the strong echo region.

7. The intelligent generation method for gridded short-term forecasts of severe convective weather according to claim 6, characterized in that, In step two, during the model training phase, the physical constraints include the divergence error of the wind field dynamics equation, the mass balance error of the rainfall conservation equation, and the electric field intensity error of the lightning energy transfer equation as additional loss terms, which are jointly optimized with the prediction error to constrain the physical consistency of the model output.

8. The intelligent generation method for grid-level short-term forecasts of severe convective weather according to claim 7, characterized in that, In step three, the autoencoder consists of fully connected layers that compress the predicted features into a low-dimensional latent space. The decoder reconstructs a high-resolution feature map through residual connections, preserving the detailed information of the original radar echo, and outputs a forecast grid with a resolution of no less than 1km×1km.

9. The intelligent generation method for gridded short-term forecasts of severe convective weather according to claim 8, characterized in that, In step four, the correction factor is calculated as follows: based on the mean square error between the measured data and the forecast results fed back by the user, combined with the time decay function and the spatial correlation matrix, a grid-by-grid correction coefficient is generated, and the model parameters are updated online through the lightweight fine-tuning module.

10. The intelligent generation method for gridded short-term forecasts of severe convective weather according to claim 9, characterized in that, In step four, the precision mode includes full-parameter mode, quantization compression mode, and lightweight mode generated by knowledge distillation. The precision mode uses the full-parameter mode to perform high-precision inference with the complete STGCN model. The quantization compression mode reduces the computational load of the model by using 8-bit fixed-point numbers. The lightweight mode simplifies the network structure based on knowledge distillation. Each mode is dynamically switched according to the user's terminal resources.

Images