Multi-scale extreme wind event AI identification method, device and medium integrating physical constraints

Abstract

The present invention discloses a multi-scale AI identification method, device and medium for extreme gale events that integrate physical constraints, and relates to the fields of atmospheric science and artificial intelligence technology. The method constructs a novel multi-scale extreme gale event identification model by integrating multi-scale meteorological data. The model effectively captures the characteristics of extreme gale at different scales through three modules: multi-scale feature extraction, feature fusion sharing, and multi-task output. In addition, the model introduces a physical constraint loss design to ensure that outputs such as wind speed and air pressure conform to the laws of meteorological dynamics. Through deep learning analysis of meteorological data, combined with the understanding of physical mechanisms, the present invention can effectively distinguish different types of extreme gale events such as hurricanes, thunderstorms and tornadoes, providing reliable data support and scientific basis for meteorological forecasting and emergency response.

Description

Technical Field

[0001] The present invention relates to the fields of atmospheric science and artificial intelligence technology, and more specifically, to a multi-scale AI identification method, device, and medium for extreme gale events that integrate physical constraints. Background Art

[0002] In the field of extreme wind event identification, current research is mainly divided into tasks such as small-scale thunderstorm wind identification and large-scale wind forecasting, and a series of technologies and methods have been developed based on different tasks.

[0003] Radar observation data, with its high temporal and spatial resolution of wind field information, provides an important basis for identifying small-scale, highly localized high-wind events. Many studies have developed identification techniques for small-scale high-wind events, such as thunderstorms and tornadoes, by leveraging radar echo characteristics, such as echo intensity and rotation, combined with deep learning methods. These deep learning-based methods, by mining the complex patterns inherent in radar data, can accurately capture the characteristics of small-scale high-wind events, providing timely information for early warning and emergency response. However, these methods are primarily limited to identifying small-scale, highly localized high-wind events and lack comprehensive coverage of extreme high-wind events at different scales.

[0004] Numerical weather forecast models play an important role in large-scale extreme wind events, such as those caused by tropical cyclones and strong cold fronts. These models, based on physical mechanisms, such as global climate models and regional climate models, can effectively simulate large-scale wind events caused by large weather systems by solving atmospheric dynamics equations. In recent years, a series of large-scale artificial intelligence meteorological models (such as Pangu) have also begun to be applied to the identification and prediction of large-scale weather systems. These models perform well in processing large-scale weather phenomena, but still face certain challenges in capturing refined small-scale features, which affects the accuracy of identification and prediction.

[0005] Overall, existing methods struggle to fully capture extreme wind events at different scales, and new technical solutions are urgently needed to fill this gap. Therefore, effectively integrating multi-source data, extracting multi-scale features, and incorporating physical mechanisms to achieve refined identification of extreme wind events at different scales has become an important research direction. Summary of the Invention

[0006] To solve the above technical problems, the present invention provides a multi-scale AI identification method, device and medium for extreme gale events that integrate physical constraints. In view of the diversity and complexity of extreme gale events, an identification scheme that integrates physical mechanisms is proposed to solve the shortcomings of traditional methods in event classification and identification accuracy, thereby improving the detection capability and identification accuracy of extreme gale events. First, by integrating multi-source meteorological data (such as satellite remote sensing data, meteorological reanalysis data and radar echo data), the characteristics of extreme gale events at different scales are comprehensively captured to ensure the consistency and high quality of the data, providing a solid foundation for model training; secondly, a multi-scale extreme gale event identification model MEWE-Net is constructed. Through multi-scale feature extraction and multi-task physical constraint loss function, the refined identification of various extreme gale events is achieved, the accuracy and stability of the model are improved, and reliable support is provided for meteorological warning and emergency response.

[0007] Specifically, the present invention aims to significantly improve the accuracy and stability of extreme wind identification by integrating multi-scale meteorological data and performing multi-task deep learning modeling, providing more reliable support for meteorological warning and emergency response.

[0008] This model processes meteorological data at different scales through three feature extraction pathways. Each pathway adjusts the network structure based on the input scale, extracting large-scale features (using ResNet-50), medium-scale features (using ResNet-34), and small-scale features (using MobileNetV2 + SE Block). A spatial attention mechanism is then used to fuse these multi-scale features to enhance focus on key spatial regions. Finally, a multi-task output layer is used to predict the areas of hurricanes, thunderstorms, and tornadoes, respectively. ResNet is a deep convolutional neural network that employs residual learning to effectively mitigate the vanishing gradient problem in deep network training. MobileNetV2 is a lightweight network designed for computational efficiency, making it suitable for resource-constrained environments. It also reduces computational complexity through depthwise separable convolutions. The spatial attention mechanism (SAM Block) automatically focuses on important regions of the input features, enhancing the model's ability to perceive key information. This design balances the characteristics of data at different scales and leverages the advantages of multi-task learning through shared layers, further improving the ability to identify extreme wind events.

[0009] In a first aspect, the present invention provides an AI-based identification method for multi-scale extreme wind events integrating physical constraints, the method comprising:

[0010] Acquiring meteorological data, and performing cleaning and standardization processing on the meteorological data to obtain preprocessed data;

[0011] Based on the pre-processed data, time-space alignment and interpolation are performed to obtain model input data; wherein, the model input data is represented as X={X sat ,X re ,X radar}∈R H×W×C , X sat represents satellite remote sensing data, X re is the meteorological reanalysis data, X radar is the radar echo data, H and W are the spatial dimensions, C is the number of channels of the data, and R is a set of real numbers;

[0012] Constructing a recognition model; wherein the recognition model includes a multi-scale feature extraction module, a feature fusion sharing layer, and a multi-task output layer. The multi-scale feature extraction module responds to the model input data and uses different convolutional structures for different data sources and different scales to perform high-level feature extraction, obtains multi-scale features, and feeds them to the feature fusion sharing layer; the feature fusion sharing layer uses a splicing layer and a spatial attention mechanism module to fuse the multi-scale features, obtains shared features, and feeds them to the multi-task output layer; the multi-task output layer is used to pass the shared features to the multi-task fully connected layer to output the probability of hurricanes, thunderstorms, and tornadoes at each location respectively;

[0013] Based on the three tasks of hurricane, thunderstorm and tornado, a multi-task physical constraint loss function is constructed, and a physical constraint loss term is designed for each task to obtain the hurricane physical constraint function, thunderstorm physical constraint function and tornado physical constraint function. According to the hurricane physical constraint function, thunderstorm physical constraint function and tornado physical constraint function, weight hyperparameters are added to obtain a comprehensive loss function;

[0014] The recognition model is trained and evaluated. During the training process, the comprehensive loss function is used as the optimization target, and the model parameters are continuously adjusted through back propagation to minimize the multi-task cross entropy loss.

[0015] Furthermore, the preprocessed data includes preprocessed satellite remote sensing data, meteorological reanalysis data, and radar echo data. Based on the preprocessed data, spatiotemporal alignment and interpolation are performed to obtain model input data, including:

[0016] The pre-processed satellite remote sensing data, meteorological reanalysis data and radar echo data are converted to the same geographic coordinate system to ensure that the meteorological characteristics can be consistently reflected at the same spatial scale;

[0017] The spatial resolution is adjusted by bilinear interpolation to obtain the same grid size H×W;

[0018] The time steps of different data sources are aligned through time interpolation, the preprocessed satellite remote sensing data are interpolated from a 10-minute frequency to a minute step of the preprocessed satellite radar echo data, and the missing time data are filled to obtain structured model input data.

[0019] Furthermore, the multi-scale feature extraction module includes a large-scale channel, a medium-scale channel and a small-scale channel;

[0020] In the large-scale channel, the model input data X first passes through the convolution layer Conv 7×7 Perform dimensionality reduction and obtain the first feature:

[0021]

[0022] Where, Represents the first feature, Conv 7×7 Represents a two-dimensional convolution layer with a convolution kernel of (7,7);

[0023] After batch normalization and maximum pooling layers, the second feature is extracted through the pre-trained ResNet-50 backbone network. The calculation formula is:

[0024]

[0025] Where, Represents the second feature, ResNet-50 represents the 50-layer residual neural network pre-trained in the torchvision library, BN represents batch normalization, and MaxPool represents maximum pooling;

[0026] The second feature is restored to its spatial resolution through the upsampling layer to obtain the large-scale feature F lar ;

[0027] In the mesoscale channel, the third characteristic is obtained by the following formula:

[0028]

[0029] Where, Represents the third feature, Conv 5×5 Represents a two-dimensional convolution layer with a convolution kernel of (5,5);

[0030] The fourth feature is extracted using batch normalization, average pooling layer and ResNet-34 network, and the calculation formula is:

[0031]

[0032] Where, Represents the fourth feature, ResNet-34 represents the 34-layer residual neural network pre-trained in the torchvision library, and AvgPool represents average pooling;

[0033] The fourth feature is passed through the upsampling layer to obtain the mid-scale feature F mid ;

[0034] In small-scale channels, the fifth characteristic is obtained by the following formula:

[0035]

[0036] Where, Represents the fifth feature, Conv 3×3 Represents a two-dimensional convolution layer with a convolution kernel of (5,5);

[0037] After further feature processing by batch normalization layer and average pooling layer, the sixth feature is obtained:

[0038]

[0039] Where, Represents the sixth feature; MobileNetV2 represents the lightweight convolutional backbone network pre-trained in the torchvision library;

[0040] Based on the sixth feature, the SE module is used to enhance the feature weights between channels to obtain small-scale features:

[0041]

[0042] Among them, F small Represents small-scale features, W1 and W2 represent the first weight matrix and the second weight matrix respectively, σ represents the Sigmoid function used to map the weights to the range of [0,1], ReLU is the activation function used for nonlinear mapping in the SE module, ⊙ represents element-by-element multiplication; i represents the row index of the feature map, and j represents the column index of the feature map.

[0043] Furthermore, in the feature fusion sharing layer, the received multi-scale feature large-scale feature F is converted into lar , mesoscale features F mid and small-scale features F small Splicing is performed on the channel dimension to obtain fusion features:

[0044] F concat =Concat(F lar ,F mid ,F small )

[0045] Where, Fconcat Represents fusion features, Concat represents concatenation operation;

[0046] The fusion feature F concat Input into the spatial attention mechanism module to obtain shared features:

[0047] F SAM =σ(Conv 7×7 (Concat(F avg ,F max )))⊙F concat

[0048] Where, F avg and F max Respectively represent F concat The feature map of average pooling and maximum pooling in the channel dimension, F SAM Indicates shared features.

[0049] Furthermore, in the multi-task output layer, the shared feature F SAM Adding the features of each scale and passing through the fully connected layer branches of each task, we get the prediction output of hurricane, thunderstorm, gale and tornado:

[0050]

[0051] Where, and They represent the prediction outputs of hurricanes, thunderstorms, and tornadoes, respectively. FC1, FC2, and FC3 represent the fully connected layer branches with weights and biases corresponding to the hurricane, thunderstorm, and tornado tasks, respectively.

[0052] Furthermore, the multi-task physical constraint loss function is expressed as:

[0053]

[0054] Where, L CE represents the multi-task physical constraint loss function, represents the model output of the i-th task, Y i represents the binary occurrence label of the true i-th task at each location, i = 1, 2, 3, corresponding to hurricane, thunderstorm and tornado respectively.

[0055] Furthermore, physical constraint loss terms are designed for each task to obtain hurricane physical constraint functions, thunderstorm and gale physical constraint functions, and tornado physical constraint functions. Based on the hurricane physical constraint functions, thunderstorm and gale physical constraint functions, and tornado physical constraint functions, weight hyperparameters are added to obtain a comprehensive loss function, including:

[0056] Based on the wind speed and air pressure information, the physical prior probability is defined to obtain the hurricane physical constraint function L h ; The thunderstorm wind prediction results are constrained by the local wind speed increase, and the thunderstorm wind physical constraint function L is obtained w ; Constrain the tornado prediction results by convection intensity, and obtain the tornado physical constraint function L t , calculated as:

[0057]

[0058] Where σ represents the Sigmoid function, V represents wind speed, P represents air pressure, ΔV represents the degree of wind speed increase, S represents convection intensity, and k1, k2, k3, and k4 all represent adjustment coefficients.

[0059] According to the hurricane physical constraint function, thunderstorm gale physical constraint function and tornado physical constraint function, weight hyperparameters are added to obtain a comprehensive loss function, which is expressed as:

[0060] L=L CE +λ1L h +λ2L w +λ3L t

[0061] Where λ1, λ2, and λ3 represent the first, second, and third weight hyperparameters, respectively, which are used to adjust the influence of hurricane physical constraints, thunderstorm wind physical constraints, and tornado physical constraints, respectively.

[0062] Furthermore, the recognition model is trained and evaluated, including:

[0063] During the model evaluation phase, precision, recall, and F1 scores were used to quantify the accuracy and robustness of the model in identifying hurricanes, thunderstorms, gales, and tornadoes. A physical consistency index was introduced to evaluate the physical relationship between wind speed and air pressure to test its rationality in model predictions. The error distribution of the model on different types of events was analyzed through a confusion matrix to identify and improve the probability of misclassification. Spatial distribution visualization was used to further verify the model's recognition ability in different geographic regions to ensure its reliability and applicability in actual meteorological environments.

[0064] In a second aspect, the present invention provides an AI-based identification device for multi-scale extreme wind events that integrates physical constraints, the device comprising:

[0065] a preprocessing unit configured to acquire meteorological data, and perform cleaning and standardization processing on the meteorological data to obtain preprocessed data;

[0066] The alignment and interpolation unit is configured to perform spatiotemporal alignment and interpolation based on the preprocessed data to obtain model input data; wherein the model input data is represented by X={X sat ,X re ,X radar}∈R H×W×C , X sat represents satellite remote sensing data, X re is the meteorological reanalysis data, X radar is the radar echo data, H and W are the spatial dimensions, C is the number of channels of the data, and R is a set of real numbers;

[0067] A model construction unit is configured to construct a recognition model; wherein the recognition model includes a multi-scale feature extraction module, a feature fusion sharing layer, and a multi-task output layer. The multi-scale feature extraction module, in response to the model input data, uses different convolutional structures for different data sources and different scales to perform high-level feature extraction, obtains multi-scale features, and feeds them to the feature fusion sharing layer; the feature fusion sharing layer uses a splicing layer and a spatial attention mechanism module to fuse the multi-scale features, obtains shared features, and feeds them to the multi-task output layer; the multi-task output layer is used to pass the shared features to the multi-task fully connected layer to output the probability of hurricanes, thunderstorms, and tornadoes at each location respectively;

[0068] The function construction unit is configured to construct a multi-task physical constraint loss function based on the three tasks of hurricane, thunderstorm and tornado, and design a physical constraint loss term for each task to obtain a hurricane physical constraint function, a thunderstorm physical constraint function, and a tornado physical constraint function; according to the hurricane physical constraint function, the thunderstorm physical constraint function, and the tornado physical constraint function, a weight hyperparameter is added to obtain a comprehensive loss function;

[0069] The training and evaluation unit is configured to train and evaluate the recognition model. During the training process, the comprehensive loss function is used as the optimization target, and the model parameters are continuously adjusted through back propagation to minimize the multi-task cross entropy loss.

[0070] In a third aspect, the present invention provides a readable storage medium, wherein the readable storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the method as described above.

[0071] The present invention has at least the following beneficial effects:

[0072] This paper constructs the MEWE-Net model to achieve high-precision identification of hurricanes, thunderstorms, and tornadoes. This method leverages multi-source data, including satellite remote sensing, meteorological reanalysis, and radar echo data, to ensure the temporal and spatial consistency of the input data. First, a data preprocessing phase cleans, standardizes, and spatially aligns the multi-source data to form structured input data. Next, the MEWE-Net model employs a multi-scale feature extraction module and a spatial attention mechanism to fuse these multi-scale features. Combined with a multi-task output layer, it accurately predicts the probability of extreme wind events.

[0073] 2) This invention utilizes a multi-task physical constraint loss design. This method leverages prior knowledge of meteorological dynamics to introduce physical constraints for hurricane, thunderstorm, and tornado tasks. Hurricanes are identified based on high wind speed and low pressure, thunderstorms are constrained by local wind speed variations, and tornadoes are identified by combining convective intensity constraints. Ultimately, a comprehensive loss function is constructed to balance classification accuracy and physical consistency. By optimizing this comprehensive loss function during model training, the model not only improves classification accuracy but also ensures the physical plausibility of the output.

[0074] 3) During the evaluation phase, the present invention comprehensively verifies the prediction performance of the model through classification indicators and physical consistency indicators, and further evaluates the recognition ability of the model in different regions by combining spatial distribution visualization, thereby ensuring its applicability and reliability in actual meteorological environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 A flowchart of an AI identification method for multi-scale extreme wind events integrating physical constraints according to an embodiment of the present invention is shown.

[0076] Figure 2 A sequence screening flow chart according to an embodiment of the present invention is shown.

[0077] Figure 3 A structural diagram of a multi-scale extreme wind event AI identification device integrating physical constraints according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0078] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments. The embodiments of the present invention are further described in detail below with reference to the accompanying drawings and specific embodiments, but are not intended to limit the present invention. For the various steps described herein, if there is no necessity for a contextual relationship between each other, the order in which they are described as examples herein should not be regarded as limiting, and those skilled in the art should know that they can be adjusted in order as long as the logic between them is not destroyed, resulting in the inability to implement the entire process.

[0079] In the fields of atmospheric science and weather forecasting, extreme wind events, as a meteorological phenomenon with significant environmental and social impacts, have garnered widespread attention. However, due to the diversity and complexity of extreme wind events, existing technologies struggle to comprehensively model multi-source data at different scales, resulting in certain limitations.

[0080] In current research, the identification of extreme high wind events often has analytical limitations for single-scale or specific types of events, and lacks a method for comprehensive identification of multi-scale extreme high wind events. Existing numerical weather forecast models and machine learning methods are often unable to effectively integrate multi-source data when capturing storm characteristics at different scales, thus limiting the comprehensive understanding and prediction of high wind events. The shortcomings of this method make it difficult to accurately identify local and large-scale high wind events when faced with complex meteorological phenomena. In addition, existing technologies lack flexibility and adaptability when dealing with the diversity of extreme weather events, and are unable to fully utilize the complementarity of various types of meteorological data, thus affecting the accuracy and reliability of the identification results. Therefore, developing a new method that can comprehensively identify extreme high wind events at different scales has become an urgent task in current research.

[0081] Based on this, an embodiment of the present invention provides an AI recognition method for multi-scale extreme wind events that integrates physical constraints. This method constructs a multi-scale extreme wind event recognition model MEWE-Net for multi-source data, designs a multi-task physical constraint loss, and achieves refined recognition of three types of events: hurricanes, thunderstorms, and tornadoes. Specifically, Figure 1 As shown, the method includes steps 1 to 5, which are described in detail as follows.

[0082] Step 1: Data collection and preprocessing: Collect satellite remote sensing, meteorological reanalysis and radar echo data, clean and standardize them, and handle missing and outlier values ​​to ensure data quality.

[0083] In step 1, meteorological data is obtained, and the meteorological data is cleaned and standardized to obtain pre-processed data.

[0084] For example, we first integrate meteorological data from different sources to build a high-quality input set with temporal and spatial consistency. Specifically, we collect three main data sources: satellite remote sensing data X sat , meteorological reanalysis data X re and radar echo data X radar, used to capture detailed information about multi-scale weather events. Satellite remote sensing imagery provides large-scale wind field characteristics, reanalysis data supplements atmospheric background parameters (such as temperature and pressure), and radar echo data provides high-resolution images of small-scale local weather characteristics. This data is cleaned and standardized, with missing values ​​filled in and outliers removed, eliminating dimensional differences and compressing the data to a uniform numerical range, ensuring consistency and high quality across all data types.

[0085] Step 2: Spatiotemporal alignment and interpolation: Unify the temporal frequency and spatial resolution of different data sources, complete projection conversion, spatial alignment and temporal interpolation, and ensure that the data has consistent spatiotemporal distribution.

[0086] In this embodiment, step 2 is a step of performing spatiotemporal alignment and interpolation on step 1. The preprocessed data can be subjected to spatiotemporal alignment and interpolation to obtain model input data; wherein the model input data is represented as X={X sat ,X re ,X radar}∈R H×W×C , X sat represents satellite remote sensing data, X re is the meteorological reanalysis data, X radar is the radar echo data, H and W are the spatial dimensions, C is the number of channels of the data, and R is a set of real numbers.

[0087] In an exemplary embodiment, in Step 2, the temporal frequency and spatial resolution of these data are further unified. First, the satellite, reanalysis, and radar data are projected into the same geographic coordinate system to ensure consistent meteorological characteristics at the same spatial scale. Bilinear interpolation is used to adjust the spatial resolution to obtain the same grid size H×W. Furthermore, temporal interpolation is used to align the time steps of the different data sources. Satellite remote sensing images are interpolated from a 10-minute frequency to a per-minute step of the radar data, and missing time data is filled in to ensure the integrity of the time series. After these steps, structured model input data is obtained.

[0088] Step 3: Build the recognition model MEWE-Net: Design multi-scale feature extraction, feature fusion sharing and multi-task output layers to form a complete deep neural network structure.

[0089] In this embodiment, the recognition model MEWE-Net includes a multi-scale feature extraction module, a feature fusion sharing layer, and a multi-task output layer. The multi-scale feature extraction module responds to the model input data and uses different convolutional structures for different data sources and different scales to extract high-level features, obtains multi-scale features, and feeds them to the feature fusion sharing layer; the feature fusion sharing layer uses a splicing layer and a spatial attention mechanism module to fuse the multi-scale features to obtain shared features and feed them to the multi-task output layer; the multi-task output layer is used to pass the shared features to the multi-task fully connected layer to output the probability of hurricanes, thunderstorms, and tornadoes at each location respectively;

[0090] In an exemplary embodiment, the recognition model MEWE-Net includes three parts: a multi-scale feature extraction module, a feature fusion sharing layer, and a multi-task output layer. It uses a multi-scale feature extraction and multi-task learning strategy to analyze different types of meteorological data to achieve accurate recognition of hurricanes, thunderstorms, and tornadoes. Figure 2 Specifically, for input data, the multi-scale feature extraction module first uses different convolutional structures (such as different convolution kernel sizes) for different data sources and different scales to extract high-level features. Then, in the feature fusion sharing layer, the splicing layer and the spatial attention mechanism module (SAM Block) are used to fuse multi-scale features to capture complex spatial relationships. Furthermore, the shared features are passed to the multi-task fully connected layer through the multi-task output layer to output the probability of hurricanes, thunderstorms, and tornadoes at each location.

[0091] The multi-scale feature extraction module is divided into three channels: large-scale, medium-scale and small-scale. In the large-scale channel, the input data X first passes through the convolution layer Conv 7×7 Perform dimensionality reduction to obtain features

[0092]

[0093] Where, Represents the first feature, Conv 7×7 Represents a two-dimensional convolution layer with a convolution kernel of (7,7);

[0094] After batch normalization and maximum pooling layers, the second feature is extracted through the pre-trained ResNet-50 backbone network. The calculation formula is:

[0095]

[0096] Where, Represents the second feature, ResNet-50 represents the 50-layer residual neural network pre-trained in the torchvision library, BN represents batch normalization, and MaxPool represents maximum pooling.

[0097] Then the spatial resolution is restored through the upsampling layer to obtain the final feature F lar .

[0098] In the medium-scale channel, the convolution kernel of the convolution layer is set to (5,5) size to adapt to the extraction of medium-scale features, resulting in:

[0099]

[0100] Where, Represents the third feature, Conv 5×5 Represents a two-dimensional convolutional layer with a convolution kernel of (5,5).

[0101] Batch normalization, average pooling layer and ResNet-34 network are further used to extract mid-scale features, i.e.

[0102]

[0103] Where, Represents the fourth feature, ResNet-34 represents the 34-layer residual neural network pre-trained in the torchvision library, and AvgPool represents average pooling.

[0104] Finally, after the upsampling layer, the mid-scale feature F is obtained mid .

[0105] In the small-scale channel, a convolution kernel of size (3,3) is used for feature mapping to obtain the feature

[0106]

[0107] Where, Represents the fifth feature, Conv 3×3 Represents a two-dimensional convolutional layer with a convolution kernel of (5,5).

[0108] Subsequently, the feature processing is further performed through the batch normalization layer and the average pooling layer to reduce the risk of overfitting. This channel uses the lightweight model MobileNetV2 to avoid the inefficiency caused by excessive small-scale information, namely:

[0109]

[0110] Where, Represents the sixth feature; MobileNetV2 represents the lightweight convolutional backbone network pre-trained in the torchvision library.

[0111] In this channel, SE Block (Squeeze-and-ExcitationBlock) is used to enhance the feature weights between channels to obtain the final small-scale feature F small :

[0112]

[0113] Among them, F small Represents small-scale features, W1 and W2 represent the first weight matrix and the second weight matrix respectively, σ represents the Sigmoid function used to map the weights to the range of [0,1], ReLU is the activation function used for nonlinear mapping in the SE module, ⊙ represents element-by-element multiplication; i represents the row index of the feature map, and j represents the column index of the feature map.

[0114] In the feature fusion sharing layer, the received multi-scale features F lar 、F mid and F small Splicing is performed on the channel dimension to obtain the following fusion features:

[0115] F concat =Concat(F lar ,F mid ,F small )

[0116] Then F concat Input into the spatial attention mechanism module (SAM Block) to further enhance the features of important spatial locations, namely:

[0117] F concat =Concat(F lar ,F mid ,F small )

[0118] Among them, σ represents the Sigmoid function, Conv 7×7 represents a convolution layer with a convolution kernel size of (7,7), F avg and F max Respectively represent F concat The feature map of average pooling and maximum pooling is performed in the channel dimension, and ⊙ represents element-by-element multiplication.

[0119] Finally, in the multi-task output layer, the fusion feature F SAM By adding the features of each scale and passing through the fully connected layer branches of each task, we can obtain the prediction output of three extreme wind events: hurricane, thunderstorm and tornado:

[0120]

[0121] Where, and They represent the prediction outputs of hurricanes, thunderstorms, and tornadoes, respectively. FC1, FC2, and FC3 represent the fully connected layer branches with weights and biases corresponding to the hurricane, thunderstorm, and tornado tasks, respectively.

[0122] Step 4: Multi-task physical constraint loss: Construct a multi-task loss function, introduce classification and regression losses, and combine physical constraints to ensure that the output results are physically reasonable.

[0123] In an exemplary embodiment, for the three tasks of hurricane, thunderstorm, and tornado, a multi-task physical constraint loss function is constructed to ensure that the output of the model for each task fits the true label value while ensuring compliance with the laws of meteorological dynamics. First, the cross entropy loss function for fitting the predicted value to the true value can be defined as:

[0124]

[0125] Where, L CE represents the multi-task physical constraint loss function, represents the model output of the i-th task, Y i represents the binary occurrence label of the true i-th task at each location, i = 1, 2, 3, corresponding to hurricane, thunderstorm and tornado respectively.

[0126] Based on the multi-task cross entropy loss function, a physical constraint loss term is designed for each extreme wind task. Since hurricanes usually occur when the wind speed is extremely high and the air pressure is extremely low, the physical prior probability is defined based on the wind speed and air pressure information to obtain the hurricane physical constraint function L h The probability of thunderstorms and gale is closely related to the increase of local wind speed. The prediction results are constrained by the increase of local wind speed, and the physical constraint function L of thunderstorms and gale is obtained. w Tornadoes are often accompanied by strong convective activities. The prediction results are constrained by the convective intensity in the input data, and the tornado physical constraint function L is obtained. t , calculated as:

[0127]

[0128] Where σ represents the Sigmoid function, V represents wind speed, P represents air pressure, ΔV represents the degree of wind speed increase, S represents convection intensity, and k1, k2, k3, and k4 represent adjustment coefficients. The above physical constraint loss terms are combined with the multi-task loss to obtain the final loss function:

[0129] L=L CE +λ1L h +λ2L w +λ3L t

[0130] Where λ1, λ2, and λ3 represent the first, second, and third weight hyperparameters, respectively, which are used to adjust the influence of hurricane physical constraints, thunderstorm wind physical constraints, and tornado physical constraints, respectively.

[0131] Step 5: Model training and evaluation: Jointly optimize multi-task losses to evaluate the accuracy and stability of model recognition.

[0132] In one exemplary embodiment, during model training, a comprehensive loss function L is used as the optimization objective. Model parameters are continuously adjusted through backpropagation to minimize the multi-task cross-entropy loss and improve the classification accuracy of hurricanes, thunderstorms, and tornadoes. Simultaneously, a physical constraint loss term ensures that the model output probabilities conform to meteorological dynamics, thereby maintaining classification performance while improving the physical consistency and credibility of the model's predictions. Ultimately, by optimizing the model's comprehensive loss function, the model will be able to accurately predict the probability of extreme weather events such as hurricanes, thunderstorms, and tornadoes.

[0133] During the model evaluation phase, a series of metrics are used to comprehensively assess the model's predictive performance and physical plausibility. First, classification evaluation metrics such as precision, recall, and F1 score are used to quantify the model's accuracy and robustness in identifying extreme events such as hurricanes, thunderstorms, and tornadoes. Furthermore, to ensure that the forecast results conform to the laws of meteorological physics, a physical consistency metric is introduced to evaluate the physical relationship between wind speed and air pressure and test its rationality in the model's predictions. Furthermore, a confusion matrix is ​​used to analyze the model's error distribution for different event types, identifying and correcting potential misclassifications. Finally, spatial distribution visualization is used to further validate the model's recognition capabilities across different geographic regions, ensuring its reliability and applicability in real-world meteorological environments. Through these multi-level evaluations, the model's performance and physical plausibility are fully verified, supporting its effectiveness in practical applications.

[0134] In summary, the innovation of this invention lies in the construction of MEWE-Net, a multi-scale extreme wind event identification model for multi-source data, and the introduction of a multi-task physical constraint loss function to achieve refined identification based on physical features. By integrating satellite remote sensing, meteorological reanalysis, and radar echo data, the model can effectively capture the spatiotemporal characteristics of events such as hurricanes, thunderstorms, and tornadoes at multiple scales, thereby achieving high-precision identification of extreme wind events. This technology integrates meteorological dynamics knowledge into model training through physical constraint loss, ensuring the physical consistency of model predictions and improving model reliability.

[0135] In terms of model architecture, the proposed MEWE-Net consists of a multi-scale feature extraction, a feature fusion sharing layer, and a multi-task output layer. It extracts meteorological features at different spatial scales through large, medium, and small-scale channels, and utilizes a spatial attention mechanism to enhance feature responses at important spatial locations. Furthermore, a multi-task physical constraint loss sets unique physical prior constraints for each high wind event. By factoring in physical conditions such as the relationship between hurricane wind speed and pressure, thunderstorm wind speed increments, and tornado convection intensity, the model's output for each task conforms to actual physical laws, further improving the model's prediction accuracy and physical consistency.

[0136] Furthermore, MEWE-Net employs multimodal data fusion and multi-task learning strategies. By aligning the spatiotemporal alignment and multi-scale processing of multi-source meteorological data, it establishes fine-grained prediction models for different types of high wind events. This design enables MEWE-Net to more accurately identify the location and probability of extreme high wind events, providing a more valuable forecast basis for extreme weather warnings.

[0137] This invention significantly improves the recognition accuracy and physical consistency of extreme wind events in meteorological disaster prediction through innovative data fusion, physical mechanism constraints and multi-task learning methods, and has good promotion and application prospects and industry value.

[0138] The embodiment of the present invention also provides a multi-scale extreme wind event AI recognition device integrating physical constraints, such as Figure 3 As shown, the device includes:

[0139] The preprocessing unit 301 is configured to obtain meteorological data, and clean and standardize the meteorological data to obtain preprocessed data;

[0140] The alignment and interpolation unit 302 is configured to perform spatiotemporal alignment and interpolation based on the pre-processed data to obtain model input data; wherein the model input data is represented by X={X sat ,X re ,X radar}∈R H×W×C , X sat represents satellite remote sensing data, X re is the meteorological reanalysis data, X radar is the radar echo data, H and W are the spatial dimensions, C is the number of channels of the data, and R is a set of real numbers;

[0141] The model construction unit 303 is configured to construct a recognition model; wherein the recognition model includes a multi-scale feature extraction module, a feature fusion sharing layer, and a multi-task output layer. The multi-scale feature extraction module, in response to the model input data, uses different convolutional structures for different data sources and different scales to perform high-level feature extraction, obtains multi-scale features, and feeds them to the feature fusion sharing layer; the feature fusion sharing layer uses a splicing layer and a spatial attention mechanism module to fuse the multi-scale features, obtains shared features, and feeds them to the multi-task output layer; the multi-task output layer is used to pass the shared features to the multi-task fully connected layer to output the probability of hurricanes, thunderstorms, and tornadoes at each location respectively;

[0142] The function construction unit 304 is configured to construct a multi-task physical constraint loss function based on the three tasks of hurricane, thunderstorm and tornado, and design a physical constraint loss term for each task to obtain a hurricane physical constraint function, a thunderstorm physical constraint function, and a tornado physical constraint function; and add weight hyperparameters based on the hurricane physical constraint function, the thunderstorm physical constraint function, and the tornado physical constraint function to obtain a comprehensive loss function;

[0143] The training evaluation unit 305 is configured to train and evaluate the recognition model. During the training process, the comprehensive loss function is used as the optimization target, and the model parameters are continuously adjusted through back propagation to minimize the multi-task cross entropy loss.

[0144] It should be noted that the various device structures described in this embodiment belong to the same technical concept as the method described previously, and achieve the same technical effect through the same principle, and will not be repeated here.

[0145] An embodiment of the present invention further provides a readable storage medium, which stores one or more programs. The one or more programs can be executed by one or more processors to implement the method described in any of the above embodiments.

[0146] Furthermore, although exemplary embodiments have been described herein, the scope includes any and all embodiments based on the present invention having equivalent elements, modifications, omissions, combinations (e.g., schemes where various embodiments intersect), adaptations, or changes. The elements in the claims are to be interpreted broadly based on the language employed in the claims and are not limited to the examples described in this specification or during the prosecution of this application, which examples are to be interpreted as non-exclusive. Therefore, this specification and examples are intended to be considered as examples only, with the true scope and spirit being indicated by the following claims and the full scope of their equivalents.

[0147] The above description is intended to be illustrative rather than restrictive. For example, the above examples (or one or more of their solutions) can be used in combination with each other. For example, those of ordinary skill in the art may use other embodiments when reading the above description. In addition, in the above-mentioned specific embodiments, various features can be grouped together to simplify the present invention. This should not be interpreted as an intention that a feature of an invention that is not claimed for protection is necessary for any claim. On the contrary, the subject matter of the present invention may be less than all the features of the embodiments of a particular invention. Thus, the following claims are incorporated into the specific embodiments as examples or embodiments, wherein each claim is independently a separate embodiment, and it is considered that these embodiments can be combined with each other in various combinations or arrangements. The scope of the present invention should be determined with reference to the appended claims and the full scope of equivalents to which these claims are entitled.

Claims

1. A multi-scale AI recognition method for extreme wind events integrating physical constraints, characterized by: The method comprises: Acquiring meteorological data, and performing cleaning and standardization processing on the meteorological data to obtain preprocessed data; Based on the pre-processed data, time-space alignment and interpolation are performed to obtain model input data; wherein, the model input data is represented as X={X sat ,X re ,X radar }∈R H×W×C , X sat represents satellite remote sensing data, X re is the meteorological reanalysis data, X radar is the radar echo data, H and W are the spatial dimensions, C is the number of channels of the data, and R is a set of real numbers; Constructing a recognition model; wherein the recognition model includes a multi-scale feature extraction module, a feature fusion sharing layer, and a multi-task output layer. The multi-scale feature extraction module responds to the model input data and uses different convolutional structures for different data sources and different scales to perform high-level feature extraction, obtains multi-scale features, and feeds them to the feature fusion sharing layer; the feature fusion sharing layer uses a splicing layer and a spatial attention mechanism module to fuse the multi-scale features, obtains shared features, and feeds them to the multi-task output layer; the multi-task output layer is used to pass the shared features to the multi-task fully connected layer to output the probability of hurricanes, thunderstorms, and tornadoes at each location respectively; Based on the three tasks of hurricane, thunderstorm and tornado, a multi-task physical constraint loss function is constructed, and a physical constraint loss term is designed for each task to obtain the hurricane physical constraint function, thunderstorm physical constraint function and tornado physical constraint function. According to the hurricane physical constraint function, thunderstorm physical constraint function and tornado physical constraint function, weight hyperparameters are added to obtain a comprehensive loss function; The recognition model is trained and evaluated. During the training process, the comprehensive loss function is used as the optimization target, and the model parameters are continuously adjusted through back propagation to minimize the multi-task cross entropy loss.

2. The method according to claim 1, characterized in that The preprocessed data includes preprocessed satellite remote sensing data, meteorological reanalysis data, and radar echo data. Based on the preprocessed data, spatiotemporal alignment and interpolation are performed to obtain model input data, including: The pre-processed satellite remote sensing data, meteorological reanalysis data and radar echo data are converted to the same geographic coordinate system to ensure that the meteorological characteristics can be consistently reflected at the same spatial scale; The spatial resolution is adjusted by bilinear interpolation to obtain the same grid size H×W; The time steps of different data sources are aligned through time interpolation, the preprocessed satellite remote sensing data are interpolated from a 10-minute frequency to a minute step of the preprocessed satellite radar echo data, and the missing time data are filled to obtain structured model input data.

3. The method according to claim 1, characterized in that The multi-scale feature extraction module includes a large-scale channel, a medium-scale channel and a small-scale channel; In the large-scale channel, the model input data X first passes through the convolution layer Conv 7×7 Perform dimensionality reduction and obtain the first feature: Where, Represents the first feature, Conv 7×7 Represents a two-dimensional convolution layer with a convolution kernel of (7,7); After batch normalization and maximum pooling layers, the second feature is extracted through the pre-trained ResNet-50 backbone network. The calculation formula is: Where, Represents the second feature, ResNet-50 represents the 50-layer residual neural network pre-trained in the torchvision library, BN represents batch normalization, and MaxPool represents maximum pooling; The second feature is restored to its spatial resolution through the upsampling layer to obtain the large-scale feature F lar ; In the mesoscale channel, the third characteristic is obtained by the following formula: Where, Represents the third feature, Conv 5×5 Represents a two-dimensional convolution layer with a convolution kernel of (5,5); The fourth feature is extracted using batch normalization, average pooling layer and ResNet-34 network. The calculation formula is: Where, Represents the fourth feature, ResNet-34 represents the 34-layer residual neural network pre-trained in the torchvision library, and AvgPool represents average pooling; The fourth feature is passed through the upsampling layer to obtain the mid-scale feature F mid ; In small-scale channels, the fifth characteristic is obtained by the following formula: Where, Represents the fifth feature, Conv 3×3 Represents a two-dimensional convolution layer with a convolution kernel of (5,5); After further feature processing by batch normalization layer and average pooling layer, the sixth feature is obtained: Where, Represents the sixth feature; MobileNetV2 represents the lightweight convolutional backbone network pre-trained in the torchvision library; Based on the sixth feature, the SE module is used to enhance the feature weights between channels to obtain small-scale features: Among them, F small Represents small-scale features, W1 and W2 represent the first weight matrix and the second weight matrix respectively, σ represents the Sigmoid function used to map the weights to the range of [0,1], ReLU is the activation function used for nonlinear mapping in the SE module, ⊙ represents element-by-element multiplication; i represents the row index of the feature map, and j represents the column index of the feature map.

4. The method according to claim 3, characterized in that In the feature fusion sharing layer, the received multi-scale feature large-scale feature F is converted into lar , mesoscale features F mid and small-scale features F small Splicing is performed on the channel dimension to obtain fusion features: F concat =Concat(F lar ,F mid ,F small ) Where, F concat Represents fusion features, Concat represents concatenation operation; The fusion feature F concat Input into the spatial attention mechanism module to obtain shared features: F SAM =σ(Conv 7×7 (Concat(F avg ,F max )))⊙F concat Where, F avg and F max Respectively represent F concat The feature map of average pooling and maximum pooling in the channel dimension, F SAM Indicates shared features.

5. The method according to claim 4, characterized in that In the multi-task output layer, the shared feature F SAM Adding the features of each scale and passing through the fully connected layer branches of each task, we get the prediction output of hurricane, thunderstorm, gale and tornado: Where, and They represent the prediction outputs of hurricanes, thunderstorms, and tornadoes, respectively. FC1, FC2, and FC3 represent the fully connected layer branches with weights and biases corresponding to the hurricane, thunderstorm, and tornado tasks, respectively.

6. The method according to claim 5, characterized in that The multi-task physical constraint loss function is expressed as: Where, L CE represents the multi-task physical constraint loss function, represents the model output of the i-th task, Y i represents the binary occurrence label of the true i-th task at each location, i = 1, 2, 3, corresponding to hurricane, thunderstorm and tornado respectively.

7. The method according to claim 6, characterized in that A physical constraint loss term is designed for each task to obtain a hurricane physical constraint function, a thunderstorm physical constraint function, and a tornado physical constraint function. Based on the hurricane physical constraint function, the thunderstorm physical constraint function, and the tornado physical constraint function, weight hyperparameters are added to obtain a comprehensive loss function, including: Based on the wind speed and air pressure information, the physical prior probability is defined to obtain the hurricane physical constraint function L h ; The thunderstorm wind prediction results are constrained by the local wind speed increase, and the thunderstorm wind physical constraint function L is obtained w ; Constrain the tornado prediction results by convection intensity, and obtain the tornado physical constraint function L t , calculated as: Where σ represents the Sigmoid function, V represents wind speed, P represents air pressure, ΔV represents the degree of wind speed increase, S represents convection intensity, and k1, k2, k3, and k4 all represent adjustment coefficients. According to the hurricane physical constraint function, thunderstorm gale physical constraint function and tornado physical constraint function, weight hyperparameters are added to obtain a comprehensive loss function, which is expressed as: L=L CE +λ1L h +λ2L w +λ3L t Where λ1, λ2, and λ3 represent the first, second, and third weight hyperparameters, respectively, which are used to adjust the influence of hurricane physical constraints, thunderstorm wind physical constraints, and tornado physical constraints, respectively.

8. The method according to claim 1, characterized in that The recognition model is trained and evaluated, including: During the model evaluation phase, precision, recall, and F1 scores were used to quantify the accuracy and robustness of the model in identifying hurricanes, thunderstorms, gales, and tornadoes. A physical consistency index was introduced to evaluate the physical relationship between wind speed and air pressure to test its rationality in model predictions. The error distribution of the model on different types of events was analyzed through a confusion matrix to identify and improve the probability of misclassification. Spatial distribution visualization was used to further verify the model's recognition ability in different geographic regions to ensure its reliability and applicability in actual meteorological environments.

9. A multi-scale AI recognition device for extreme wind events integrating physical constraints, characterized by: The device comprises: a preprocessing unit configured to acquire meteorological data, and perform cleaning and standardization processing on the meteorological data to obtain preprocessed data; The alignment and interpolation unit is configured to perform spatiotemporal alignment and interpolation based on the preprocessed data to obtain model input data; wherein the model input data is represented by X={X sat ,X re ,X radar }∈R H×W×C , X sat represents satellite remote sensing data, X re is the meteorological reanalysis data, X radar is the radar echo data, H and W are the spatial dimensions, C is the number of channels of the data, and R is a set of real numbers; A model construction unit is configured to construct a recognition model; wherein the recognition model includes a multi-scale feature extraction module, a feature fusion sharing layer, and a multi-task output layer. The multi-scale feature extraction module, in response to the model input data, uses different convolutional structures for different data sources and different scales to perform high-level feature extraction, obtains multi-scale features, and feeds them to the feature fusion sharing layer; the feature fusion sharing layer uses a splicing layer and a spatial attention mechanism module to fuse the multi-scale features, obtains shared features, and feeds them to the multi-task output layer; the multi-task output layer is used to pass the shared features to the multi-task fully connected layer to output the probability of hurricanes, thunderstorms, and tornadoes at each location respectively; The function construction unit is configured to construct a multi-task physical constraint loss function based on the three tasks of hurricane, thunderstorm and tornado, and design a physical constraint loss term for each task to obtain a hurricane physical constraint function, a thunderstorm physical constraint function, and a tornado physical constraint function; according to the hurricane physical constraint function, the thunderstorm physical constraint function, and the tornado physical constraint function, a weight hyperparameter is added to obtain a comprehensive loss function; The training and evaluation unit is configured to train and evaluate the recognition model. During the training process, the comprehensive loss function is used as the optimization target, and the model parameters are continuously adjusted through back propagation to minimize the multi-task cross entropy loss. 10 . A non-transitory computer-readable storage medium storing instructions, which, when executed by a processor, perform the method according to claim 1 .

Images