A method and apparatus for predicting the intensity of a cyclone

Abstract

The application discloses a cyclone intensity prediction method and device, and the method comprises the following steps: acquiring physical factor data, satellite cloud image data and cyclone intensity data in a historical period; performing pretreatment on the physical factor data, the satellite cloud image data and the cyclone intensity data to obtain a physical factor image, a satellite image image and historical cyclone intensity; inputting the physical factor image, the satellite image image and the historical cyclone intensity into a tropical cyclone intensity prediction model to obtain a 24-hour cyclone intensity prediction result; wherein the tropical cyclone intensity prediction model comprises a hierarchical fusion subnetwork, a deep convolutional neural network, a prior knowledge module and an uncertainty estimation module. The fusion of the uncertainty estimation and the long-term prior knowledge improves the prediction ability of the cyclone intensity change trend and the accuracy of the intensity prediction.

Description

Technical Field

[0001] This application relates to a method and apparatus for predicting cyclone intensity, belonging to the field of meteorological forecasting technology. Background Technology

[0002] Tropical cyclones are among the most destructive natural phenomena globally, triggering catastrophic consequences such as strong winds, storm surges, and landslides caused by heavy rainfall. The severity of their impact depends primarily on the intensity of the tropical cyclone; therefore, accurate and timely intensity forecasting is crucial for disaster prevention and mitigation.

[0003] Currently, the main methods for predicting tropical cyclone intensity include numerical weather prediction models and deep learning models. Numerical weather prediction models are the mainstream method in operational forecasting, but they require enormous computational resources. In recent years, with the advancement of artificial intelligence technology, deep learning methods have been widely used and have achieved significant results in tropical cyclone intensity prediction due to their advantages of effectively simulating high-order nonlinear systems and relatively low computational cost. However, there are still shortcomings. On the one hand, most existing deep learning models are deterministic systems, which cannot quantify the inherent uncertainties caused by observational input noise and the model's own abstraction. On the other hand, most of these models are based only on short-term trajectory data of a single cyclone, ignoring broader long-term and universal evolutionary patterns, resulting in "shortsightedness" and limited generalization ability. Summary of the Invention

[0004] Objective: In view of at least one of the above technical problems, this application provides a method and apparatus for predicting cyclone intensity, which improves the accuracy of cyclone intensity prediction based on uncertainty estimation and long-term priors.

[0005] The technical solution adopted in this application is:

[0006] In a first aspect, this application provides a method for predicting cyclone intensity, including:

[0007] Acquire physical factor data, satellite cloud image data, and cyclone intensity data over a historical period;

[0008] Preprocessing physical factor data, satellite cloud image data, and cyclone intensity data yields physical factor images, satellite imagery, and historical cyclone intensity data.

[0009] By inputting physical factor images, satellite imagery, and historical cyclone intensity into the tropical cyclone intensity prediction model, the 24-hour cyclone intensity prediction results are obtained.

[0010] The tropical cyclone intensity prediction model includes a hierarchical fusion sub-network, a deep convolutional neural network, a prior knowledge module, and an uncertainty estimation module. The training method includes: acquiring labeled multi-year physical factor data, satellite cloud image data, and cyclone intensity data; preprocessing the physical factor data, satellite cloud image data, and cyclone intensity data to construct a training dataset; and using the training dataset to train and optimize the tropical cyclone intensity prediction model to obtain a trained tropical cyclone intensity prediction model.

[0011] Secondly, this application provides a cyclone intensity prediction device, including a processor and a storage medium;

[0012] The storage medium is used to store instructions;

[0013] The processor is configured to operate according to the instructions to execute the method according to the first aspect.

[0014] Beneficial effects: The cyclone intensity prediction method and apparatus provided in this application have the following advantages: In view of the problems that existing deep learning methods cannot quantify prediction uncertainty and have insufficient generalization ability due to ignoring long-term evolution laws in tropical cyclone intensity prediction, this application proposes a tropical cyclone intensity prediction method that integrates uncertainty estimation and long-term prior. Compared with existing methods, the embodiments of this application improve the prediction ability of cyclone intensity change trend and the accuracy of intensity prediction. Attached Figure Description

[0015] Figure 1 This is a schematic flowchart of a cyclone intensity prediction method according to an embodiment of this application;

[0016] Figure 2 This is a schematic diagram of a cyclone intensity prediction model according to an embodiment of this application;

[0017] Figure 3 This is a schematic diagram of an uncertainty estimation module according to an embodiment of this application;

[0018] Figure 4 This is a comparative schematic diagram of the first cyclone intensity forecast case provided in the embodiments of this application;

[0019] Figure 5 This is a comparative schematic diagram of a second cyclone intensity forecast case provided in the embodiments of this application;

[0020] Figure 6 This is a comparative schematic diagram of a third cyclone intensity forecast case provided in the embodiments of this application. Detailed Implementation

[0021] The present application will be further described below with reference to the accompanying drawings and embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present application, and should not be used to limit the scope of protection of the present application.

[0022] In the description of this application, "several" means one or more, "multiple" means two or more, "greater than," "less than," and "exceeding" are understood to exclude the stated number, while "above," "below," and "within" are understood to include the stated number. The use of "first" and "second" in the description is merely for distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the order of the indicated technical features.

[0023] In the description of this application, the terms "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0024] The term "and / or" simply describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0025] Example 1: This example provides a method for predicting cyclone intensity, such as... Figure 1 As shown, it includes:

[0026] Acquire physical factor data, satellite cloud image data, and cyclone intensity data over a historical period;

[0027] Preprocessing of physical factor data, satellite cloud image data, and cyclone intensity data yields physical factor images, satellite imagery, and historical cyclone intensity data (physical factor images, satellite imagery, and cyclone intensity data at times t-24h, t-18h, t-12h, t-6h, and t).

[0028] By inputting physical factor images, satellite imagery, and historical cyclone intensity into the tropical cyclone intensity prediction model, the 24-hour cyclone intensity prediction results are obtained.

[0029] The tropical cyclone intensity prediction model includes a hierarchical fusion sub-network, a deep convolutional neural network, a prior knowledge module (LEPE), and an uncertainty estimation module. The training method includes: acquiring labeled multi-year physical factor data, satellite cloud image data, and cyclone intensity data; preprocessing the physical factor data, satellite cloud image data, and cyclone intensity data to construct a training dataset; and using the training dataset to train and optimize the tropical cyclone intensity prediction model to obtain a trained tropical cyclone intensity prediction model.

[0030] It should be noted that the training dataset can be constructed using the multi-year Northwest Pacific Best Track Dataset (CMA BST), which includes cyclone names, latitude and longitude of the cyclones at various times, and actual cyclone intensities. Based on the time range of each cyclone, corresponding satellite cloud imagery and physical factor data are obtained. The physical factor data is sourced from the ERA5 reanalysis data of the European Centre for Medium-Range Weather Forecasts (ECMWF), selecting three key factors: sea level temperature (SST), 600 hPa relative humidity (RH), and vertical wind shear (VWS) between the 850 hPa and 200 hPa pressure layers. The satellite cloud imagery data is sourced from the Gridsat-b1 dataset, primarily using the infrared and water vapor channels, which effectively depict water vapor and cloud top characteristics.

[0031] In some embodiments, a training dataset is constructed by preprocessing physical factor data, satellite cloud image data, and cyclone intensity data, including:

[0032] Based on the latitude and longitude of the cyclone center in the optimal path dataset, the physical factor center and satellite cloud image data center are located. The physical factor data and satellite cloud image data are then cropped to obtain physical factor images and satellite imagery; (e.g., cropping an image centered on the latitude and longitude of the tropical cyclone at a specific time from each satellite image). Pixel region, and extract one from each physical factor data (grid area)

[0033] For each time t within a cyclone cycle, locate the physical factor images, satellite imagery, and cyclone intensity data for t-24h, t-18h, t-12h, t-6h, t, and t+24h;

[0034] Using each cyclone cycle as a unit, plot the physical factors at t-24h, t-18h, t-12h, t-6h, t, and t+24h. Satellite imagery Cyclone intensity data The data is integrated (i.e., the physical factor images, satellite imagery, and cyclone intensity data are aligned according to each time point) to form a training dataset.

[0035] In some embodiments, such as Figure 2 As shown, the data processing procedure of the tropical cyclone intensity prediction model includes:

[0036] S1. Using a hierarchical fusion subnetwork to process physical factor images Feature extraction is performed to obtain intermediate features of physical factors. .

[0037] In some embodiments, in step S1, the physical factor image is processed using a hierarchical fusion subnetwork. Feature extraction is performed to obtain intermediate features of physical factors. ,include:

[0038] Physical factor image Sea level temperature images were obtained by decomposing the data according to physical factors. Relative humidity image and vertical wind shear images ; sea level temperature image Relative humidity image and vertical wind shear images The initial fusion features are obtained by weighting. ;

[0039] Sea level temperature images were processed using convolutional layers respectively. Relative humidity image and vertical wind shear images Initial fusion features Feature extraction was performed to obtain the first sea level temperature feature. First relative humidity characteristic First vertical wind shear characteristics Initial fusion and feature extraction ,Will , , , Weighted fusion yields the first fusion feature ;

[0040] Using convolutional layers respectively , , , Feature extraction was performed to obtain the second sea level temperature feature. Second relative humidity characteristics Second vertical wind shear characteristics First fusion feature extraction ,Will , , , Weighted fusion yields the second fusion feature ;

[0041] Using convolutional layers respectively , , , Feature extraction was performed to obtain the third sea level temperature feature. Third relative humidity characteristics Third vertical wind shear characteristics Second fusion feature extraction ,Will , , , Weighted fusion yields the third fusion feature As an intermediate feature of physical factors .

[0042] The hierarchical fusion subnetwork consists of three layers of fusion subnetworks, and the data processing procedure for each layer of fusion subnetwork is represented as follows:

[0043]

[0044] in, Indicates the first Layered fusion subnetwork, , Indicates the first The input and output of the layer fusion subnetwork combine sea level temperature images. Relative humidity image and vertical wind shear images Considered , , , All of these are learnable weight parameters, and conv represents convolution.

[0045] S2. Utilizing deep convolutional neural networks to process satellite imagery. Feature extraction is performed to obtain intermediate features from satellite images. .

[0046] In some embodiments, in step S2, a deep convolutional neural network is used to process the satellite imagery. Feature extraction is performed to obtain intermediate features from satellite images. ,include:

[0047] The deep convolutional neural network comprises a first convolutional module, a second convolutional module, and a third convolutional module connected in sequence. Each of the first, second, and third convolutional modules includes an initial convolutional layer and a 3x3 convolutional layer. The max pooling layer has a size of 3; the kernel size of the initial convolutional layer in the first convolutional module is 7. 7. With a stride of 2, the initial convolutional layer of the second convolutional module uses a kernel size of 3. 3. With a stride of 1, the initial convolutional layer of the third convolutional module uses a kernel size of 13×13 with a stride of 1.

[0048] In this embodiment, satellite imagery The input has 10 channels, with multispectral channels (water vapor and cloud top features) and temporal channels integrated together. In the first convolutional module, it first passes through an initial convolutional layer, which uses a size of... A convolutional kernel with a stride of 2 is used to quickly reduce the input image resolution and expand the number of channels from 15 to 64, thereby capturing a wide range of cloud structure and texture features. Then, a... The max-pooling layer further downsamples the features and enhances their translation invariance. Then, the convolution and pooling operations are repeated twice via a second and third convolutional module, aiming to learn more complex feature representations.

[0049] S3. Utilize prior knowledge modules to analyze historical cyclone intensity. Feature extraction was performed to obtain intermediate features of historical cyclone intensity. ;

[0050] In some embodiments, in step S3, the prior knowledge module is used to analyze the historical cyclone intensity. Feature extraction was performed to obtain intermediate features of historical cyclone intensity. ,include:

[0051] Historical cyclone intensity Each time step is treated as a node in the graph, and through successive first graph convolutional layers and second graph convolutional layers, a pre-constructed adjacency matrix is ​​utilized. The feature dimensions of each node are increased, and the increased cyclone intensity is then used for feature extraction through a convolutional layer and a fully connected layer to obtain intermediate features of historical cyclone intensity. .

[0052] In this embodiment, the first convolutional layer increases the feature dimension of each node from 5 to 8 dimensions, and the second convolutional layer further increases it to 16 dimensions. The cyclone intensity after dimension increase is used to extract the intermediate features of historical cyclone intensity through a convolutional layer and a fully connected layer. (20 dimensions).

[0053] Furthermore, by utilizing a pre-constructed adjacency matrix The feature dimension of each node is increased, represented as:

[0054] Using a pre-built adjacency matrix The feature dimension of each node is increased, represented as:

[0055] ;

[0056] in, Indicates the first The output of the convolutional layer is shown. Indicates the convolutional layer of the graph. Represents a non-linear activation function. Indicates the first The input to the convolutional layer is shown in the image. Indicates the first The learnable weight matrix in a graph convolutional layer.

[0057] Furthermore, in some embodiments, the adjacency matrix The construction methods include:

[0058] Maximum sustained wind speed (MSW) is divided into 9 intensity categories (0, 1, 2, 3, 4, 5, 6, 7, 8), as shown in Table 1;

[0059] Table 1: Intensity Category of Maximum Sustained Wind Speed

[0060]

[0061] Based on the Best Paths dataset (BST) from 1949 to 2014, the intensity of the maximum sustained wind speed within each adjacent time step is statistically analyzed from category... Switch to Category Number of times According to all Construct a frequency matrix ;in Let C represent a matrix where all elements are real numbers. K×K represents the dimension of the matrix. It means that matrix C is a square matrix with K rows and K columns (K is 9 in this example).

[0062] For the frequency matrix Each row is normalized to obtain the adjacency matrix. .

[0063] S4. Intermediate features of physical factors Intermediate features of satellite imagery Intermediate characteristics of historical cyclone intensity After flattening, the mixture is cascaded and the average value of the cyclone intensity prediction results is obtained through a fully connected layer. and variance .

[0064] S5. Utilize intermediate features of satellite imagery Intermediate characteristics of historical cyclone intensity Input uncertainty estimation module to obtain satellite image reconstruction features Historical Cyclone Intensity Reconstruction Characteristics .

[0065] Further, in step S5, as Figure 3 As shown, the data processing procedure of the uncertainty estimation module specifically includes:

[0066] S51, respectively, the intermediate features of satellite images are processed by the first multilayer perceptron. Intermediate characteristics of historical cyclone intensity Mapping is performed to obtain satellite image mapping features Historical cyclone intensity mapping characteristics ;

[0067] In step S51, intermediate features of satellite imagery Intermediate characteristics of historical cyclone intensity It is mapped into a higher-dimensional shared latent space with a unified dimension, preparing for subsequent cross-modal interactions.

[0068] S52, Satellite Image Mapping Features Historical cyclone intensity mapping characteristics Image features are obtained by performing cross-attention mechanism operations. Historical characteristics ;

[0069] In step S52, the data from the two modalities undergo a cross-attention operation, wherein the first decoder uses satellite imagery mapping features. For querying, use historical cyclone intensity mapping features. For memory, generate an updated image feature that incorporates historical information. The second decoder, on the contrary, maps features based on historical cyclone intensity. For querying, based on satellite image mapping features To aid memory, generate an updated historical feature that incorporates visual information. .

[0070] S53, respectively, image features Historical characteristics The first mean is obtained by performing self-attention mechanism operations (each through a self-attention module and a fully connected layer). and first variance Second mean Second variance ;

[0071] S54, Based on the first mean and first variance The first latent variable is obtained by sampling. According to the second mean Second variance The second latent variable is obtained by sampling. ;

[0072] In step S54, a reparameterization technique is employed. A latent variable with random perturbation is obtained by sampling from the posterior distribution using the following formula:

[0073] ;

[0074] ;

[0075] in ), Let represent a random variable that follows a probability distribution. This represents a normal distribution with a mean of 0 and a variance of 1.

[0076] S55, respectively, the first latent variable is processed through the second multilayer perceptron. Second hidden variable Decoding and reconstruction yields satellite image reconstruction features Historical Cyclone Intensity Reconstruction Characteristics .

[0077] During training, the uncertainty estimation module ultimately outputs the uncertainty loss function of the satellite cloud image data. Uncertainty loss function for cyclone intensity data , represented as:

[0078] ;

[0079] ;

[0080] in, This indicates L1 normalization.

[0081] Constructing the model uncertainty estimation loss function , represented as:

[0082] ;

[0083] in, Indicates L2 normalization, Indicates the true label, , This represents the mean and variance of the cyclone intensity prediction results output by the model.

[0084] Construct the total loss function of the tropical cyclone intensity prediction model , represented as:

[0085] ;

[0086] in, This represents the loss function for estimating model uncertainty. , These represent the uncertainty loss functions for satellite cloud image data and cyclone intensity data, respectively. and These are the weight hyperparameters; in this embodiment, they are all set to 0.1. During model training and optimization, the total loss function is minimized as much as possible to obtain the optimal model, which serves as the trained tropical cyclone intensity prediction model.

[0087] Simulation example: See Figures 4 to 6 As shown in the figure, this embodiment provides a comparative diagram of three cyclone intensity forecast cases to verify the experimental results. Figure 4 , Figure 5 , Figure 6 This paper presents three representative tropical cyclones (TCs) as case studies to evaluate the performance of the TCIF-LPUE model proposed in this embodiment against several operational forecasting models in real-world scenarios. These diverse cases were specifically selected for this application embodiment, including a long-lived typhoon, a Category 5 typhoon, two rapid intensification processes, and a typhoon that maintained high intensity for an extended period, thus providing a comprehensive assessment of the model's capabilities. The focus here is on the model's performance in two key aspects of tropical cyclone intensity forecasting: capturing long-term evolution trends, i.e., accurately characterizing rapid intensification or deterioration processes. The model in this application embodiment demonstrates excellent ability to track the overall intensity evolution of long-lived typhoons throughout their entire lifecycle. For powerful and long-lasting typhoons such as Typhoon Shureki (2021), Typhoon Noru (2017), and Typhoon Dandelion (2021), the predicted intensity curves (shown by the blue line) and the optimal path data (shown by the black line) show a high degree of agreement from formation to dissipation. For example, in the case of Typhoon Noru-2017, which underwent a complex re-intensification process, the model in this embodiment achieved an impressive mean absolute error (MAE) of only 1.86 m / s and a root mean square error (RMSE) of 2.80 m / s. This demonstrates that the model in this embodiment not only provides accurate point-to-point forecasts but also comprehensively captures the entire process of long-term development, peak performance, and decline of tropical cyclones, showcasing its stability and profound understanding of the intrinsic physical processes governing tropical cyclone evolution.

[0088] A key challenge in tropical cyclone intensity forecasting is predicting rapid intensification (RI) changes. The embodiments of this application demonstrate a strong ability to identify and forecast these challenging events. The model excels at capturing rapid intensification. For Super Typhoon Shulik-2021, optimal track data showed an explosive increase in intensity of 30 m / s within the 24 hours leading up to 12:00 UTC on April 17. TCIF-LPUE predicted a significant intensification of 23 m / s, while models from the China Meteorological Administration (CMA) and the Japan Meteorological Agency (JMA) showed much slower intensification rates during this critical period. The model's accuracy is also evident in the case of Typhoon Dandelion-2021. Between 00:00 UTC on September 25 and 00:00 UTC on September 26, it predicted an intensity increment of 21.6 m / s, which closely matches the observed intensification of 25 m / s. In contrast, the CMA forecast was conservative and failed to capture the true rate of intensification. The model also skillfully handled multiple rapid intensification phases. During Typhoon Noru's second rapid intensification phase, from 18:00 UTC on July 29th to 18:00 UTC on July 31st, TCIF-LPUE correctly identified the start of this explosive event, while the official forecasts from CMA and JMA showed a significant lag.

[0089] Furthermore, the embodiments of this application also demonstrate excellent performance in predicting peak intensity and subsequent attenuation phases. A prominent example is "Olu-2017," for which TCIF-LPUE predicted a peak intensity of 52.40 m / s at 12:00 UTC on July 31, which is remarkably consistent with the observed peak intensity of 52 m / s. After reaching the peak, TCIF-LPUE's forecasts of the lengthy attenuation phases in all five cases, including the sharp decline observed in "Dandelion-2021," closely tracked the observed intensity decrease. However, the forecasts of CMA and JMA deviated significantly from the optimal path data. This indicates that the model of this embodiment can effectively respond to environmental factors that lead to attenuation.

[0090] all in all, Figures 4 to 6 The case studies highlight the advantages of the intensity forecasting model proposed in this application. It not only accurately tracks the macroscopic life history trends of various tropical cyclones but also excels in the critical and challenging task of predicting rapid intensification changes. Its ability to capture the onset and rate of rapid intensification, accurately predict peak intensity, and closely follow subsequent weakening trends signifies its enormous potential as a reliable and valuable operational tropical cyclone intensity forecasting tool.

[0091] This embodiment optimizes and tests the cyclone intensity forecasting model TCIF-LPUE based on uncertainty estimation and long-term prior cyclone intensity prediction methods.

[0092] The original dataset was divided into training and validation sets, covering the period from 2000 to 2014, containing 21,474 samples from 352 tropical cyclone events. 80% of these samples were randomly assigned for training, and the remaining 20% ​​were used for validation. The test set spanned from 2015 to 2022, containing 5,822 samples from 183 tropical cyclone events. The spatiotemporal features of tropical cyclones were used as input features for tropical cyclone intensity prediction, and tropical cyclone intensity data was used as the model's output data, constructing a feature combination dataset for model training.

[0093] The feature combination dataset is input into the TCIF-LPUE tropical cyclone intensity prediction model for training and optimization. Model parameter optimization includes, but is not limited to, adjustments to hyperparameters such as input feature dimension, number of hidden layers, network structure parameters, connection layer parameters, number of training batches, number of training epochs, and model learning rate.

[0094] After the hyperparameters of the TCIF-LPUE tropical cyclone intensity prediction model were optimized, the model parameters were trained again using all the training set data. An optimization algorithm was employed, using forward propagation to calculate the loss function and backpropagation to update the gradient to obtain the model's weight parameters, until the training set loss function converged to its minimum value. Once the model was trained, training was terminated, and the optimal model was saved. The predictive performance (evaluation metrics) of the TCIF-LPUE tropical cyclone intensity prediction model was then evaluated.

[0095] The accuracy of the model is evaluated by calculating the mean absolute error (MAE) and root mean square error (RMSE) based on the predicted and actual values ​​of the test set samples. The results show that the MAE of the TCIF-LPUE model for tropical cyclone intensity prediction in this embodiment is 3.35 m / s and the RMSE is 4.37 m / s under non-real-time settings; and the MAE is 4.00 m / s and the RMSE is 5.26 m / s under simulated real-time settings.

[0096] Other deep learning models and business forecasting models yielded the following results on the test set:

[0097] The MAE of the SAF-Net model is 4.30 m / s;

[0098] The MAE of the TC-Pred model is 3.98 m / s;

[0099] The MAE of the SWP-TCI model is 5.68 m / s;

[0100] The MAE of the MSCAR model is 4.29 m / s;

[0101] The MAE of the ECMWF-IFS business model is 7.09 m / s;

[0102] The MAE of the NCEP-GFS service model is 5.70 m / s;

[0103] The MAE for the CMA business model is 4.46 m / s;

[0104] The MAE for the JMA business model is 4.59 m / s.

[0105] This demonstrates that when evaluating different models using the same dataset, a smaller error index indicates higher model prediction accuracy. Specifically, the model based on the TCIF-LPUE tropical cyclone intensity prediction method provided in this application outperforms other deep learning methods and established business benchmark models in terms of tropical cyclone intensity prediction accuracy.

[0106] Furthermore, the model of this application demonstrated excellent and consistent performance during the testing period from 2015 to 2022. Even under more demanding real-time forecasting conditions, the model of this application maintained a low and stable error distribution, ultimately achieving an average MAE of 4.00 m / s. This stable performance provides a unique competitive advantage; the model of this application not only has an overall MAE lower than all operational benchmark models, but also achieved performance equal to or better than the best operational model in six of the eight-year testing period. This demonstrates that the model of this application has strong generalization ability and robustness in tropical cyclone seasons with various characteristics.

[0107] Example 2: Based on Example 1, this example provides a cyclone intensity prediction device, including a processor and a storage medium;

[0108] The storage medium is used to store instructions;

[0109] The processor is configured to operate according to the instructions to execute the method according to Embodiment 1.

[0110] Example 3: Based on Example 1, this example provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in Example 1.

[0111] Example 4: Based on Example 1, this example provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described in Example 1.

[0112] Example 5: Based on Example 1, this example provides a computer program product, including a computer program that, when executed by a processor, implements the method described in Example 1.

[0113] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0114] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0115] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0116] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0117] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for predicting cyclone intensity, characterized in that, include: Acquire physical factor data, satellite cloud image data, and cyclone intensity data over a historical period; Preprocessing physical factor data, satellite cloud image data, and cyclone intensity data yields physical factor images, satellite imagery, and historical cyclone intensity data. By inputting physical factor images, satellite imagery, and historical cyclone intensity into the tropical cyclone intensity prediction model, the 24-hour cyclone intensity prediction results are obtained. The tropical cyclone intensity prediction model includes a hierarchical fusion sub-network, a deep convolutional neural network, a prior knowledge module, and an uncertainty estimation module. The training method includes: acquiring labeled multi-year physical factor data, satellite cloud image data, and cyclone intensity data; preprocessing the physical factor data, satellite cloud image data, and cyclone intensity data to construct a training dataset; and using the training dataset to train and optimize the tropical cyclone intensity prediction model to obtain a trained tropical cyclone intensity prediction model. The data processing procedure of the tropical cyclone intensity prediction model includes: using a hierarchical fusion subnetwork to process physical factor images. Feature extraction is performed to obtain intermediate features of physical factors. Using deep convolutional neural networks to analyze satellite imagery Feature extraction is performed to obtain intermediate features from satellite images. Using prior knowledge modules to analyze historical cyclone intensity Feature extraction was performed to obtain intermediate features of historical cyclone intensity. ; intermediate features of physical factors Intermediate features of satellite imagery Intermediate characteristics of historical cyclone intensity After flattening, the mixture is cascaded and the average value of the cyclone intensity prediction results is obtained through a fully connected layer. and variance Intermediate features of satellite imagery Intermediate characteristics of historical cyclone intensity Input uncertainty estimation module to obtain satellite image reconstruction features Historical Cyclone Intensity Reconstruction Characteristics ; The data processing of the uncertainty estimation module includes: processing intermediate features of satellite images through the first multilayer perceptron. Intermediate characteristics of historical cyclone intensity Mapping is performed to obtain satellite image mapping features Historical cyclone intensity mapping characteristics ; Mapping features of satellite images Historical cyclone intensity mapping characteristics Image features are obtained by performing cross-attention mechanism operations. Historical characteristics ; respectively for image features Historical characteristics The first mean was obtained by performing self-attention mechanism operations. and first variance Second mean Second variance According to the first mean and first variance The first latent variable is obtained by sampling. According to the second mean Second variance The second latent variable is obtained by sampling. The first latent variable was processed by a second multilayer perceptron. Second hidden variable Decoding and reconstruction yields satellite image reconstruction features Historical Cyclone Intensity Reconstruction Characteristics .

2. The method according to claim 1, characterized in that, The total loss function of the tropical cyclone intensity prediction model , represented as: ； ； ； ； in, This represents the loss function for estimating model uncertainty. and For weight hyperparameters, , These represent the uncertainty loss functions for satellite cloud image data and cyclone intensity data, respectively. Indicates L2 normalization, Indicates the true label, , This represents the mean and variance of the cyclone intensity prediction results output by the model; This indicates L1 normalization.

3. The method according to claim 1, characterized in that, Using hierarchical fusion subnetworks for physical factor images Feature extraction is performed to obtain intermediate features of physical factors. ,include: Physical factor image Sea level temperature images were obtained by decomposing the data according to physical factors. Relative humidity image and vertical wind shear images ; sea level temperature image Relative humidity image and vertical wind shear images The initial fusion features are obtained by weighting. ; Sea level temperature images were processed using convolutional layers respectively. Relative humidity image and vertical wind shear images Initial fusion features Feature extraction was performed to obtain the first sea level temperature feature. First relative humidity characteristic First vertical wind shear characteristics Initial fusion and feature extraction ,Will , , , Weighted fusion yields the first fusion feature ; Using convolutional layers respectively , , , Feature extraction was performed to obtain the second sea level temperature feature. Second relative humidity characteristics Second vertical wind shear characteristics First fusion feature extraction ,Will , , , Weighted fusion yields the second fusion feature ; Using convolutional layers respectively , , , Feature extraction was performed to obtain the third sea level temperature feature. Third relative humidity characteristics Third vertical wind shear characteristics Second fusion feature extraction ,Will , , , Weighted fusion yields the third fusion feature As an intermediate feature of physical factors .

4. The method according to claim 1, characterized in that, Using deep convolutional neural networks to analyze satellite imagery Feature extraction is performed to obtain intermediate features from satellite images. ,include: The deep convolutional neural network comprises a first convolutional module, a second convolutional module, and a third convolutional module connected in sequence. Each of the first, second, and third convolutional modules includes an initial convolutional layer and a 3x3 convolutional layer. The max pooling layer has a size of 3; the kernel size of the initial convolutional layer in the first convolutional module is 7.

7. With a stride of 2, the initial convolutional layer of the second convolutional module uses a kernel size of 3.

3. With a stride of 1, the initial convolutional layer of the third convolutional module uses a kernel size of 13×13 with a stride of 1.

5. The method according to claim 1, characterized in that, Using prior knowledge modules to analyze historical cyclone intensity Feature extraction was performed to obtain intermediate features of historical cyclone intensity. ,include: Historical cyclone intensity Each time step is treated as a node in the graph, and through successive first graph convolutional layers and second graph convolutional layers, a pre-constructed adjacency matrix is ​​utilized. The feature dimensions of each node are increased, and the increased cyclone intensity is then used for feature extraction through a convolutional layer and a fully connected layer to obtain intermediate features of historical cyclone intensity. ; Using a pre-built adjacency matrix The feature dimension of each node is increased, represented as: ； in, Indicates the first The output of the convolutional layer is shown. Indicates the convolutional layer of the graph. Represents a non-linear activation function. Indicates the first The input to the convolutional layer is shown in the image. Indicates the first The learnable weight matrix in a graph convolutional layer.

6. The method according to claim 5, characterized in that, Adjacency matrix The construction methods include: Maximum sustained wind speeds are divided into 9 intensity categories; Based on the best path dataset from 1949 to 2014, the intensity of the maximum sustained wind speed within each adjacent time step is statistically analyzed from the category... Switch to Category Number of times According to all Construct a frequency matrix ;in Let C represent a matrix where all elements are real numbers, and K×K represent the dimension of the matrix, meaning that matrix C has K rows and K columns. For the frequency matrix Each row is normalized to obtain the adjacency matrix. .

7. The method according to claim 1, characterized in that, The training dataset was constructed by preprocessing physical factor data, satellite cloud imagery data, and cyclone intensity data, including: Based on the latitude and longitude of the cyclone center in the optimal path data, the physical factor center and satellite cloud image data center are located. The physical factor data and satellite cloud image data are then cropped to obtain physical factor images and satellite imagery. For each time t within a cyclone cycle, locate the physical factor images, satellite imagery, and cyclone intensity data for t-24h, t-18h, t-12h, t-6h, t, and t+24h; Using each cyclone cycle as a unit, physical factor images, satellite imagery, and cyclone intensity data at t-24h, t-18h, t-12h, t-6h, t, and t+24h are integrated to form a training dataset.

8. A cyclone intensity prediction device, characterized in that, Including processor and storage media; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the method according to any one of claims 1 to 7.

Images