Tropical cyclone strength prediction method and device based on large meteorological model

Abstract

The invention relates to a tropical cyclone strength prediction method and device based on a large meteorological model, and the method comprises the following stages: 1, mapping tropical cyclone strength information to a discrete vector implicit space through an encoder under the conditions of a satellite image and reanalysis of atmospheric field data, a mapping result is compared with all vectors of a learnable discrete codebook, the closest vector is taken and restored to original tropical cyclone strength information through a decoder, and iterative training is carried out on the encoder, the discrete codebook and the decoder; and a second stage: inputting historical meteorological data into the trained encoder and discrete codebook, converting the historical meteorological data into discrete hidden variables, carrying out autoregression iteration to obtain forecast hidden variables, and restoring tropical cyclone strength information through the decoder by taking wind-wind forecast field data as physical constraints to obtain a tropical cyclone strength prediction result. Compared with the prior art, the method has the advantages that long-term forecasting can be achieved, good performance is preserved, and the calculation cost is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of tropical cyclone intensity prediction, and in particular to a tropical cyclone intensity prediction method and device based on a large meteorological model. Background Art

[0002] Tropical cyclone (TC) is a very destructive natural disaster. In the past few decades, the path forecast of tropical cyclones has made great progress, but its intensity forecast is still difficult. Traditional tropical cyclone intensity forecast is usually divided into dynamic model and statistical model. The dynamic model is mainly based on fluid dynamics equations, and solves the tropical cyclone intensity from the atmospheric field, such as the China Meteorological Administration Global Assimilation Forecast System (CMA-GFS), the National Center for Environmental Prediction Global Forecast System (NCEP-GFS), and the European Center for Medium-Range Weather Forecasts Integrated Forecast System (ECMWF-IFS). The dynamic model often consumes a lot of computing resources, and because the physical process of tropical cyclone intensity is not clear and the influencing factors are complex, the accuracy of its intensity forecast still needs to be improved. The statistical model mainly relies on historical observation data modeling, and uses statistical techniques such as regression analysis, probability statistics, and time series analysis to achieve forecasts, such as the famous Hurricane Intensity Statistical Forecast Scheme (SHIPS). The statistical model is simple and easy to implement and deploy, but it cannot effectively model the nonlinear processes of tropical cyclones, has poor prediction accuracy, and is difficult to predict extreme situations where tropical cyclones suddenly intensify.

[0003] In recent years, deep learning methods have shown great advantages in various prediction tasks. They are good at handling nonlinear processes and can use multi-source data to capture potential physical relationships. At present, many researchers have made various attempts and made certain progress in short-term forecasting in the northwest Pacific Ocean. The Multi-Modal Spatial-Temporal Network (MMSTN) in the document "A Multi-Modal Spatial-Temporal Network for Tropical Cyclone Short-Term Prediction" (Huang C, Bai C, Chan S, et al. Geophysical Research Letters, 2022, 49 (4): e2021GL096898.) proposes the use of generative adversarial networks for ensemble forecasting. Later, the Multi-Generator Tropical Cyclone Forecasting (MGTCF) model proposed in the paper "MGTCF: multi-generator tropical cyclone forecasting with heterogeneous meteorological data" (Huang C, Bai C, Chan S, et al. Proceedings of the AAAI Conference on Artificial Intelligence. 2023, 37 (4): 5096-5104.) introduced multi-source data fusion on this basis, further improving the performance. However, the performance of these two methods in actual deterministic forecasting is poor and cannot meet actual needs.

[0004] There are also some methods that focus on global oceans, such as the Multi-modal Multi-scale Causal Autoregressive Model (MSCAR) published in the paper "Global Tropical Cyclone Intensity Forecasting with Multi-modal Multi-scale Causal Autoregressive Model" (Wang X, Chen K, Liu L, et al. arxiv preprint arxiv:2402.13270, 2024.). The MSCAR model is based on multi-source data, including satellite image data, ECMWF fifth-generation global climate atmospheric reanalysis data (ECMWFReanalysis v5, ERA5) and tropical cyclone intensity information. The process can be roughly divided into the following three steps:

[0005] 1) Feature Pyramid Networks (FPN) are used to extract multi-scale spatial feature information from satellite images and ERA5.

[0006] 2) The spatial feature information is cross-attended with the cyclone intensity information. The calculation method is changed according to the natural temporal causal relationship. That is, the cyclone intensity information at the current moment will only be calculated with the spatial feature information at the current moment and the historical moment to obtain the initial latent space vector representing the intensity.

[0007] 3) The tropical cyclone intensity information at the last moment of input is fused with the initial latent space vector, and the latent space vector of the next forecast moment is obtained through cross-attention calculation. Then, it is added to the tropical cyclone intensity information at the last moment through a feedforward neural network to obtain the final forecast tropical cyclone intensity.

[0008] 4) Perform the same operation as step 3) on the latent space vector of the prediction time obtained in step 3) and repeat T steps to obtain the prediction result after T steps.

[0009] This scheme mainly calculates multi-source data directly through cross-attention to obtain the latent space vector calculation results, and repeatedly iterates to obtain the forecast results after T steps. The real-time calculation of this scheme requires a lot of computing resources, and the long-term forecast performance is poor. The final official forecast often requires manual correction, which tests the quality and ability of the forecaster.

[0010] In summary, the existing solutions have the following defects:

[0011] 1) The numerical forecast models currently used by various meteorological agencies consume a large amount of computing resources, and the final official forecast often requires manual correction, which tests the quality and ability of forecasters.

[0012] 2) Most AI methods focus on short-term forecasts of 24 hours, and have poor performance in long-term forecasts

[0013] 3) Most methods are limited to local sea areas and have poor adaptability to other sea areas.

[0014] 4) Most artificial intelligence methods have poor robustness, rely on reanalysis data, and cannot provide real-time forecasts. Summary of the invention

[0015] The purpose of the present invention is to provide a method and device for predicting tropical cyclone intensity based on a large meteorological model in order to overcome the defects of the above-mentioned prior art that a large amount of computing resources are consumed and the long-term forecasting performance is poor.

[0016] The purpose of the present invention can be achieved by the following technical solutions:

[0017] A tropical cyclone intensity prediction method based on a large meteorological model includes the following stages:

[0018] Phase 1: Obtain satellite images, reanalysis atmospheric field data, and tropical cyclone intensity information of the area to be tested in advance. Using satellite images and reanalysis atmospheric field data as conditions, use the encoder to map the tropical cyclone intensity information to a discrete vector latent space. Compare the mapping result with all vectors of the learnable discrete codebook, take the closest vector and restore it back to the original tropical cyclone intensity information through the decoder, and iteratively train the encoder, discrete codebook, and decoder.

[0019] The second stage: obtain the historical meteorological data and wind and rain forecast field data of the test area, input the historical meteorological data into the trained encoder and discrete codebook, convert it into discrete latent variables, and perform autoregressive iteration to obtain the forecast latent variables. Use the wind and rain forecast field data as physical constraints, and restore the tropical cyclone intensity information through the trained decoder to obtain the tropical cyclone intensity prediction result.

[0020] Furthermore, in the first stage, the encoder uses a convolutional neural network to downsample satellite images and reanalysis atmospheric field data, outputs feature maps of different scales, and then adopts a top-down feature fusion mechanism to gradually upsample from deep feature maps using linear interpolation, and fuses with the shallow feature maps of the same scale output in the previous downsampling process, and finally obtains multi-scale features, decomposes the feature maps of each scale into blocks, and then flattens them according to the feature dimension to obtain basic units of spatial features; linear layer mapping is used to represent tropical cyclone intensity information as basic units of intensity information;

[0021] The basic unit of intensity information is used as the query matrix, and the basic unit of spatial features is used as the key matrix and value matrix. Cross-attention calculation is performed to obtain latent spatial features, which are then matched with the vectors in the learnable discrete codebook. The closest vector is taken to obtain the quantized feature vector, which is input into the decoder to restore the original tropical cyclone intensity.

[0022] Furthermore, the decoder uses the quantized feature vector as a query matrix, and operates the reanalyzed atmospheric field data through the downsampling and feature fusion mechanism of the encoder to obtain the basic unit of spatial features as the key matrix and value matrix for cross-attention calculation, and finally uses a feedforward neural network to map back to the original intensity information value.

[0023] Furthermore, for satellite images and reanalysis atmospheric field data with inconsistent spatial scales, when the encoder performs downsampling, two different convolutional neural networks are used to downsample the satellite images and the reanalysis atmospheric field data to the same scale, and then splice and fuse them, and then upsample them according to the scale of the reanalysis atmospheric field data.

[0024] Furthermore, in the second stage, the historical meteorological data includes historical satellite image data, historical reanalysis atmospheric field data and historical tropical cyclone intensity information.

[0025] Furthermore, in the second stage, the latent space vector iteration model is used to iterate the discrete latent variables for T steps to obtain the latent variables for the forecast step T, which are used to predict the tropical cyclone intensity in the future step T.

[0026] Furthermore, the latent space vector iteration model is a causal cross-attention calculation module in MSCAR, which performs cross-attention calculation on the information at the current moment only with the information at the current moment and the historical moments.

[0027] Furthermore, in the second stage, the wind and rain forecast field data are input into the decoder together with the forecast latent variables to obtain preliminary results; the wind and rain forecast field data are processed through the physical constraint model to obtain the prediction results, and are superimposed with the preliminary results to obtain the final tropical cyclone intensity prediction results.

[0028] Furthermore, the physical constraint model sets basis functions that conform to the physical laws of tropical cyclones and selects basis functions using a learnable sparse matrix to obtain the nonlinear characteristics of tropical cyclone intensity, thereby improving the accuracy of tropical cyclone intensity prediction.

[0029] The present invention also provides a tropical cyclone intensity prediction device based on a large meteorological model, comprising a memory and a processor, wherein the memory stores a computer program, and the processor calls the computer program to execute the steps of the method described above.

[0030] Compared with the prior art, the present invention has the following advantages:

[0031] (1) In the pre-training stage, the present invention uses satellite image data and ERA5 reanalysis data as conditions to map tropical cyclone intensity information to a latent space through an encoder, compares the output with all vectors of the learnable codebook, outputs the closest vector to the decoder, and restores the original tropical cyclone intensity through the decoder. The pre-training performed in this way can obtain a discrete codebook that represents the tropical cyclone intensity information in the latent space, that is, obtain a discrete potential representation of the tropical cyclone intensity;

[0032] In the real-time prediction process, the trained discrete codebook can quickly calculate the hidden variables for the prediction of tropical cyclone intensity, and introduce the wind and rain forecast field data to restore the tropical cyclone intensity information through the decoder, which can achieve high-precision long-term forecast of tropical cyclone intensity with lower inference cost and faster inference speed.

[0033] (2) The present invention also uses the wind and rain forecast field data as the physical constraint for tropical cyclone intensity prediction, so that the final forecast is more consistent with actual physical laws.

[0034] (3) The present invention overcomes the problem that most deep learning methods are limited to short-term predictions, and can achieve long-term predictions of arbitrary step lengths while maintaining good performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 A schematic diagram of the first stage training process of a tropical cyclone intensity prediction method based on a large meteorological model provided in an embodiment of the present invention;

[0036] Figure 2 A schematic diagram of the second stage processing process of a tropical cyclone intensity prediction method based on a large meteorological model provided in an embodiment of the present invention;

[0037] Figure 3 It is a schematic diagram of prediction error results of a tropical cyclone intensity prediction method based on a large meteorological model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0038] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0039] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0040] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0041] Example 1

[0042] The Fengwu large model proposed in the document "Fengwu: Pushing the skillful global medium-range weather forecast beyond 10 days lead" (Chen K, Han T, Gong J, et al. arxiv preprint arxiv: 2304.02948, 2023) successfully achieved an effective forecast of 10.75 days. This application uses the prediction results of the Fengwu large model in the tropical cyclone intensity prediction to achieve high-precision long-term tropical cyclone intensity prediction. Specifically, a tropical cyclone intensity prediction method based on a large meteorological model is proposed, which includes the following stages:

[0043] In the first stage, a conditional vector quantization variational autoencoder needs to be pre-trained, including: obtaining satellite images, reanalysis atmospheric field data and tropical cyclone intensity information of the test area in advance, using satellite images and reanalysis atmospheric field data as conditions, mapping tropical cyclone intensity information to discrete vector latent space through the encoder, comparing the mapping result with all vectors of the learnable discrete codebook, taking the closest vector and restoring it to the original tropical cyclone intensity information through the decoder, and iteratively training the encoder, discrete codebook and decoder;

[0044] The second stage: obtain the historical meteorological data and wind and rain forecast field data of the test area, input the historical meteorological data into the trained encoder and discrete codebook, convert it into discrete latent variables, and perform autoregressive iteration to obtain the forecast latent variables. Use the wind and rain forecast field data as physical constraints, and restore the tropical cyclone intensity information through the trained decoder to obtain the tropical cyclone intensity prediction result.

[0045] Specifically, in the first stage, the encoder uses a convolutional neural network to downsample satellite images and reanalysis atmospheric field data, outputs feature maps of different scales, and then adopts a top-down feature fusion mechanism to gradually upsample from deep feature maps using linear interpolation, and fuses them with the shallow feature maps of the same scale output in the previous downsampling process, and finally obtains multi-scale features. The feature maps of each scale are decomposed into blocks, and then flattened according to the feature dimension to obtain the basic unit of spatial features; the tropical cyclone intensity information is represented as the basic unit of intensity information using linear layer mapping;

[0046] The basic unit of intensity information is used as the query matrix, and the basic unit of spatial features is used as the key matrix and value matrix. Cross-attention calculation is performed to obtain latent spatial features, which are then matched with the vectors in the learnable discrete codebook. The closest vector is taken to obtain the quantized feature vector, which is input into the decoder to restore the original tropical cyclone intensity.

[0047] In this embodiment, the encoder adopts a feature pyramid network (FPN) design similar to MSCAR, uses a convolutional neural network to downsample the reanalysis field data and satellite image data, outputs feature maps of different scales, and then adopts a top-down feature fusion mechanism to gradually upsample from the deep feature map using linear interpolation, and fuses it with the shallow feature map of the same scale output in the previous downsampling process, and finally obtains rich multi-scale features. Through this multi-scale extraction method, the structural characteristics of different scales during the entire life cycle of the tropical cyclone can be effectively captured, such as the cloud band change characteristics and typhoon eye in different periods.

[0048] It is worth noting that the original FPN is different in that due to the inconsistent spatial scale of the input source, two different convolutional neural networks are used to sample the two data samples to the same scale and then splice and fuse them when downsampling, and then upsample according to the scale of the reanalysis field data. The feature maps of different scales are finally obtained using a segmentation method similar to the Vision Transformer (ViT) (Dosovitskiy A, Beyer L, Kolesnikov A, et al. An image is worth 16x16 words: Transformers for image recognition at scale [J]. arXiv preprint arXiv: 2010.11929, 2020.) to decompose the feature map into 5×5 patches, and then flatten them according to the feature dimension to obtain the basic unit of spatial features (Token). The encoding of TC intensity information is relatively simple, using a linear layer mapping to represent the basic unit of intensity information (Token).

[0049] The decoder is composed of a cross attention calculation module and a feedforward neural network (FNN). The quantized feature vector is used as the query matrix. The reanalyzed atmospheric field data is operated through the downsampling and feature fusion mechanism of the encoder to obtain the basic unit of spatial features. The cross attention calculation is performed as the key matrix and the value matrix. Finally, the feedforward neural network is used to map back to the original intensity information value.

[0050] Specifically, the cross attention calculation can be expressed as the following formula:

[0051]

[0052] Where Q (Query) is the query matrix, K (Key) is the key matrix, V (Value) is the value matrix, d is the dimension of the key, and W Q , W k and W v They represent the learnable mapping matrix parameters of the three. Q is generally an input sequence. K and V take the same input and are another input sequence. When calculating, only Q is calculated with K and V at the historical moment, not with the future time. In the actual calculation, the sequence of basic units of intensity information is used as Q, and the basic units of spatial features are used as K and V. The latent spatial features are obtained through cross-attention calculation, and then matched with the vectors in the learnable discrete codebook to obtain the closest vector. The quantized feature vector is then input into the decoder. The quantized feature vector is used as Q. After the reanalyzed data is obtained through the same process as the encoder to obtain the basic units of spatial features, it is used as conditions K and V, and cross-attention calculation is performed with Q. Finally, the feedforward neural network is used to map back to the original intensity information value.

[0053] Through the above training steps, a discrete codebook that represents the tropical cyclone intensity information in the latent space can be obtained and can be used for the next stage of training tasks.

[0054] In the second stage, the wind and rain forecast field data are input into the decoder together with the forecast latent variables to obtain preliminary results; the wind and rain forecast field data are processed through the physical constraint model to obtain the prediction results, which are superimposed with the preliminary results to obtain the final tropical cyclone intensity prediction results.

[0055] In this embodiment, the historical meteorological data of the second stage is 6 hours in time interval, with a total of 4 steps, including 24 hours of historical satellite images, ERA5 reanalysis atmospheric field data and tropical cyclone intensity information.

[0056] The input historical meteorological data is passed through the encoder and discrete codebook to obtain a 4-step sequence length hidden variable, and then iterated T steps through a hidden space vector iteration model to obtain the hidden variable for the T-th step forecast. The atmospheric field data originally input into the decoder is replaced with the ERA5 reanalysis atmospheric field data of the Fengwu forecast, and input into the decoder together with the forecast hidden variables to obtain a preliminary result. At the same time, the atmospheric field data of the Fengwu forecast will also pass through a physical constraint model, and the output will be superimposed with the preliminary forecast results to obtain the final forecast of the tropical cyclone intensity in the T-th step in the future.

[0057] This scheme transfers the forecasting task to the latent space to deal with the problem that the original input-output asymmetry leads to the inability to autoregressive forecast. The latent space vector iteration model adopts the causal cross-attention calculation module in MSCAR, that is, the current moment information will only be cross-attention calculated with the current moment and the historical moment information. The physical constraint model can refer to the literature "TC-SINDy: Improving physics-based deterministic tropical cyclone track and intensity model via data-driven sparse identification of Nonlinear Dynamics" (Zhong X, Jiang W, Zhang J. Journal of Wind Engineering and Industrial Aerodynamics, 2024, 250: 105758.) By setting the basis function that conforms to the physical laws of tropical cyclones, the basis function is selected using a learnable sparse matrix to obtain the tropical cyclone intensity forecast result with physical constraints.

[0058] This example conducts a simple experiment, training a model with 24-hour historical input and outputting a 24-hour forecast, and then tests the model with a 120-hour forecast and the data from 2022. The results are as follows: Figure 3 As shown, it can be found that the error growth of long-term forecasts has dropped significantly, meeting our expectations.

[0059] This embodiment also provides a tropical cyclone intensity prediction device based on a large meteorological model, including a memory and a processor, wherein the memory stores a computer program, and the processor calls the computer program to execute the steps of the tropical cyclone intensity prediction method based on a large meteorological model as described above.

[0060] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A tropical cyclone intensity prediction method based on a large meteorological model, characterized in that: The following phases are included: Phase 1: Obtain satellite images, reanalysis atmospheric field data, and tropical cyclone intensity information of the area to be tested in advance. Using satellite images and reanalysis atmospheric field data as conditions, use the encoder to map the tropical cyclone intensity information to a discrete vector latent space. Compare the mapping result with all vectors of the learnable discrete codebook, take the closest vector and restore it back to the original tropical cyclone intensity information through the decoder, and iteratively train the encoder, discrete codebook, and decoder. The second stage: obtain the historical meteorological data and wind and rain forecast field data of the test area, input the historical meteorological data into the trained encoder and discrete codebook, convert it into discrete latent variables, and perform autoregressive iteration to obtain the forecast latent variables. Use the wind and rain forecast field data as physical constraints, and restore the tropical cyclone intensity information through the trained decoder to obtain the tropical cyclone intensity prediction result.

2. The method for predicting tropical cyclone intensity based on a large meteorological model according to claim 1, characterized in that: In the first stage, the encoder uses a convolutional neural network to downsample satellite images and reanalysis atmospheric field data, outputs feature maps of different scales, and then adopts a top-down feature fusion mechanism to gradually upsample from deep feature maps using linear interpolation, and fuses them with shallow feature maps of the same scale output in the previous downsampling process, finally obtaining multi-scale features, decomposing feature maps of each scale into blocks, and then flattening them according to feature dimensions to obtain basic units of spatial features; The tropical cyclone intensity information is represented as the basic unit of intensity information using linear layer mapping; The basic unit of intensity information is used as the query matrix, and the basic unit of spatial features is used as the key matrix and value matrix. Cross-attention calculation is performed to obtain latent spatial features, which are then matched with the vectors in the learnable discrete codebook. The closest vector is taken to obtain the quantized feature vector, which is input into the decoder to restore the original tropical cyclone intensity.

3. The method for predicting tropical cyclone intensity based on a large meteorological model according to claim 2, characterized in that: The decoder uses the quantized feature vector as the query matrix, and operates the reanalyzed atmospheric field data through the downsampling and feature fusion mechanism of the encoder to obtain the basic unit of spatial features as the key matrix and value matrix for cross-attention calculation, and finally uses a feedforward neural network to map back to the original intensity information value.

4. The method for predicting tropical cyclone intensity based on a large meteorological model according to claim 2, characterized in that: For satellite images and reanalysis atmospheric field data with inconsistent spatial scales, when the encoder performs downsampling, two different convolutional neural networks are used to downsample the satellite images and the reanalysis atmospheric field data to the same scale, then splice and fuse them, and then upsample them according to the scale of the reanalysis atmospheric field data.

5. The method for predicting tropical cyclone intensity based on a large meteorological model according to claim 1, characterized in that: In the second stage, the historical meteorological data includes historical satellite image data, historical reanalysis atmospheric field data and historical tropical cyclone intensity information.

6. A method for predicting tropical cyclone intensity based on a large meteorological model according to claim 1, characterized in that: In the second stage, the latent space vector iteration model is used to iterate the discrete latent variables for T steps to obtain the latent variables for the T-th step forecast, which are used to predict the tropical cyclone intensity in the T-th step in the future.

7. A tropical cyclone intensity prediction method based on a large meteorological model according to claim 6, characterized in that: The latent space vector iteration model is a causal cross-attention calculation module in MSCAR, which performs cross-attention calculation on the information at the current moment only with the information at the current moment and the information at the historical moment.

8. A method for predicting tropical cyclone intensity based on a large meteorological model according to claim 1, characterized in that: In the second stage, the wind and rain forecast field data are input into the decoder together with the forecast latent variables to obtain preliminary results; the wind and rain forecast field data are processed through the physical constraint model to obtain the prediction results, and superimposed with the preliminary results to obtain the final tropical cyclone intensity prediction results.

9. A method for predicting tropical cyclone intensity based on a large meteorological model according to claim 8, characterized in that: The physical constraint model sets basis functions that conform to the physical laws of tropical cyclones and selects basis functions using a learnable sparse matrix to obtain the nonlinear characteristics of tropical cyclone intensity, thereby improving the accuracy of tropical cyclone intensity prediction.

10. A tropical cyclone intensity prediction device based on a large meteorological model, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and the processor calls the computer program to execute the steps of any one of the methods according to claims 1 to 9.

Images