A tropical cyclone intensity prediction method and device based on a meteorological large model

Abstract

The present application relates to a kind of tropical cyclone intensity prediction method and device based on meteorological big model, method includes following stage: first stage: with satellite image and reanalysis atmospheric field data as condition, tropical cyclone intensity information is mapped to discrete vector hidden space by encoder, compare the mapping result with all vectors of learnable discrete codebook, take the closest vector and restore back original tropical cyclone intensity information by decoder, and iteration training is carried out to encoder, discrete codebook and decoder;Second stage: historical meteorological data is input into trained encoder and discrete codebook, is converted into discrete hidden variable, and is carried out autoregressive iteration, obtains forecast hidden variable, and with wind-uw forecast field data as physical constraint, tropical cyclone intensity information is restored back by decoder, obtains tropical cyclone intensity prediction result.Compared with prior art, the present application has the advantages of long-term prediction, good performance, reduced computing cost and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of tropical cyclone intensity prediction, and particularly relates to a tropical cyclone intensity prediction method and device based on a meteorological large model. BACKGROUND

[0002] Tropical cyclone (TC) is a kind of natural disaster with great destructive power. In the past few decades, great progress has been made in the track prediction of tropical cyclones, but the intensity prediction of tropical cyclones is still very difficult. Traditional tropical cyclone intensity prediction is usually divided into dynamic model and statistical model. The dynamic model is mainly based on fluid dynamics equations to solve the intensity of tropical cyclone from the atmospheric field, such as China Meteorological Administration Global Assimilation Forecast System (CMA-GFS), National Environmental Prediction Center Global Forecast System (NCEP-GFS), and European Center for Medium-Range Weather Forecasts Integrated Forecast System (ECMWF-IFS). The dynamic model often needs to consume a large amount of computing resources, and because the physical process of the intensity of the tropical cyclone is not clear and the influencing factors are complex, the prediction accuracy of the intensity of the tropical cyclone 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 realize prediction, such as the famous hurricane intensity statistical prediction scheme (SHIPS). The statistical model is simple and easy to implement and deploy, but it cannot well realize the modeling of the nonlinear process of the tropical cyclone, the prediction accuracy is poor, and it is difficult to predict the extreme case of sudden enhancement of the tropical cyclone.

[0003] In recent years, deep learning methods have shown great advantages in various prediction tasks, are good at processing nonlinear processes, and can utilize multi-source data to capture potential physical relationships. Currently, many researchers have made various attempts and have made some progress in short-term prediction in the Northwest Pacific. The multi-modal spatial-temporal network (MMSTN) in the literature "AMulti-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 to use a generative adversarial network for ensemble prediction. Then, the multi-generator tropical cyclone forecasting model (MGTCF) proposed in the literature "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.) further improves the performance by introducing multi-source data fusion. However, the performance of these two methods in actual deterministic prediction is poor and cannot meet the actual demand.

[0004] There are also some methods that focus on global ocean areas, and the literature "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.) discloses a multi-modal multi-scale causal autoregressive model (MSCAR). The MSCAR model is based on multi-source data, including satellite image data, ECMWF Reanalysis v5 (ERA5) and tropical cyclone intensity information, and its process can be roughly divided into the following three steps:

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

[0006] 2) Cross-attention calculation of spatial feature information and cyclone intensity information. The calculation method is changed according to the natural time causal relationship, i.e. the cyclone intensity information at the current time will only be calculated with the spatial feature information at the current time and historical time, and the initial hidden space vector representing the intensity is obtained.

[0007] 3) Fusion of the input tropical cyclone intensity information at the last time and the initial hidden space vector, cross-attention calculation of the next forecast time hidden space vector, and then addition of the last time tropical cyclone intensity information through a feedforward neural network to obtain the final forecast tropical cyclone intensity.

[0008] 4) The hidden space vector obtained in step 3) is subjected to the same operation as step 3), and T steps are repeated to obtain the T-step prediction result.

[0009] This scheme mainly calculates the hidden space vector calculation result by directly calculating the multi-source data through cross-attention, and iterates to obtain the T-step prediction result. The real-time calculation of this scheme requires a large amount of computing resources, and the long-term prediction performance is not good. The final official prediction often needs to be manually corrected, testing the quality and ability of the forecasters.

[0010] In summary, the existing scheme has the following defects:

[0011] 1) Current numerical prediction models used by various meteorological agencies require a large amount of computing resources, and the final official forecast often needs to be manually corrected, testing the quality and ability of the forecasters.

[0012] 2) Most artificial intelligence methods focus on 24-hour short-term forecasts, and long-term forecast performance is poor

[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 and rely on reanalysis data, which cannot be used for real-time forecasting. SUMMARY

[0015] The purpose of the present application is to overcome the defects of the prior art, such as the need for a large amount of computing resources and poor long-term forecast performance, and to provide a tropical cyclone intensity prediction method and device based on a meteorological large model.

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

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

[0018] First stage: Pre-acquire satellite images, reanalysis atmospheric field data and tropical cyclone intensity information of the area to be measured, use satellite images and reanalysis atmospheric field data as conditions, map tropical cyclone intensity information to discrete vector latent space through an encoder, compare the mapping result with all vectors of a learnable discrete codebook, take the closest vector, restore the original tropical cyclone intensity information through a decoder, and iteratively train the encoder, discrete codebook and decoder;

[0019] Second stage: Obtain historical meteorological data and wind-uv forecast field data of the area to be measured, input the historical meteorological data into the trained encoder and discrete codebook, convert it into discrete latent variables, and perform autoregressive iteration to obtain prediction latent variables, and restore the tropical cyclone intensity information through the trained decoder with the wind-uv forecast field data as a physical constraint, to obtain the tropical cyclone intensity prediction result.

[0020] Further, in the first stage, the encoder uses a convolutional neural network to downsample the satellite images and reanalysis atmospheric field data, outputs feature maps of different scales, uses a top-down feature fusion mechanism, uses linear interpolation to gradually upsample from deep feature maps, and fuses with shallow feature maps of the same scale output in the previous downsampling process, finally obtains multi-scale features, decomposes each scale feature map into a tile, and then flattens it according to the feature dimension to obtain a spatial feature basic unit; use a linear layer to map the tropical cyclone intensity information to an intensity information basic unit;

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

[0022] Further, the decoder takes the quantized feature vector as a query matrix, performs down-sampling and feature fusion mechanism operation on the reanalysis atmospheric field data through the encoder to obtain a spatial feature basic unit, takes the spatial feature basic unit as a key matrix and a value matrix, performs cross attention calculation, and finally maps back to the original intensity information value through a feedforward neural network.

[0023] Further, for satellite images and reanalysis atmospheric field data with inconsistent spatial scales, two different convolutional neural networks are used to down-sample the satellite images and the reanalysis atmospheric field data to the same scale for splicing and fusion, and then up-sampling is performed according to the scale of the reanalysis atmospheric field data.

[0024] Further, 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] Further, in the second stage, a hidden space vector iterative model is used to iterate the discrete hidden variable for T steps to obtain a hidden variable predicted at the Tth step, which is used to predict the tropical cyclone intensity at the Tth step in the future.

[0026] Further, the hidden space vector iterative model is a causal cross attention calculation module in the MSCAR, and the information at the current time is only cross-attention calculated with the information at the current time and the historical time.

[0027] Further, in the second stage, the wind and cloud forecast field data and the predicted hidden variable are input into the decoder to obtain a preliminary result, the wind and cloud forecast field data are processed through a physical constraint model to obtain a prediction result, and the prediction result is superimposed with the preliminary result to obtain a final tropical cyclone intensity prediction result.

[0028] Further, the physical constraint model sets a basis function conforming to the physical law of the tropical cyclone, and uses a learnable sparse matrix to select the basis function to obtain the nonlinear characteristics of the tropical cyclone intensity, so as to improve the accuracy of the tropical cyclone intensity prediction.

[0029] The application also provides a tropical cyclone intensity prediction device based on a meteorological large model, which comprises a memory and a processor, the memory stores a computer program, and the processor calls the computer program to execute the steps of the method as 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 the tropical cyclone intensity information to a latent space through an encoder. The output is compared with all vectors of the learnable codebook, and the closest vector is output to the decoder. The decoder restores the original tropical cyclone intensity. 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 the discrete latent representation of the tropical cyclone intensity.

[0032] In the real-time prediction process, latent variables can be quickly calculated using a trained discrete codebook for predicting tropical cyclone intensity. Furthermore, wind and fog forecast field data can be introduced and the tropical cyclone intensity information can be restored through a decoder, enabling high-precision long-term forecasts of tropical cyclone intensity with lower inference costs and faster inference speed.

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

[0034] (3) This invention overcomes the problem that most deep learning methods are limited to short-term forecasts, and can achieve long-term forecasts with arbitrary step lengths while maintaining good performance. Attached Figure Description

[0035] Figure 1 This is 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 This is a schematic diagram of the second-stage processing of a tropical cyclone intensity prediction method based on a large meteorological model provided in an embodiment of the present invention.

[0037] Figure 3 This is a schematic diagram of the 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 Implementation

[0038] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0039] Therefore, the following detailed description of the embodiments of the application provided in the accompanying drawings is not intended to limit the scope of the application claimed, but merely represents selected embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the application.

[0040] It should be noted that similar reference numbers and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0041] Example 1

[0042] The Fengwu model proposed in the paper "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 10.75-day effective prediction. The prediction results of the Fengwu model are used in the prediction of tropical cyclone intensity in the present application to achieve high-precision long-term prediction of tropical cyclone intensity. Specifically, a tropical cyclone intensity prediction method based on a meteorological large model is proposed, including 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 region to be measured in advance, taking the satellite images and reanalysis atmospheric field data as conditions, mapping the tropical cyclone intensity information to a discrete vector hidden space through an encoder, comparing the mapping result with all vectors of a learnable discrete codebook, taking the closest vector to restore the original tropical cyclone intensity information through a decoder, and iteratively training the encoder, discrete codebook and decoder;

[0044] In the second stage, historical meteorological data and Fengwu forecast field data of the region to be measured are obtained, the historical meteorological data are input into the trained encoder and discrete codebook, converted into discrete hidden variables, and iteratively autoregressively obtained prediction hidden variables, and the Fengwu forecast field data are taken as physical constraints to restore the tropical cyclone intensity information through the trained decoder to obtain the prediction result of the tropical cyclone intensity.

[0045] Specifically, in the first stage, the encoder uses a convolutional neural network to down-sample the satellite image and the reanalysis atmospheric field data, output different scale feature maps, and then uses a top-down feature fusion mechanism to gradually up-sample from the deep feature maps using linear interpolation, and fuse with the same scale shallow feature maps output in the previous down-sampling process, and finally obtain multi-scale features. The feature maps of each scale are decomposed into tiles, and then flattened according to the feature dimension to obtain spatial feature basic units. The tropical cyclone intensity information is represented as intensity information basic units by using a linear layer mapping.

[0046] The intensity information basic units are used as a query matrix, the spatial feature basic units are used as a key matrix and a value matrix, cross-attention calculation is performed, and hidden space features are obtained. Then, the vector in the learnable discrete codebook is matched, and 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 uses a feature pyramid network (FPN) similar to MSCAR, which uses a convolutional neural network to down-sample the reanalysis field data and satellite image data, output different scale feature maps, and then uses a top-down feature fusion mechanism to gradually up-sample from the deep feature maps using linear interpolation, and fuse with the same scale shallow feature maps output in the previous down-sampling process, and finally obtain rich multi-scale features. Through this multi-scale extraction method, different scale structural features of the tropical cyclone throughout its life cycle can be effectively captured, such as cloud band change features and typhoon eye at different times.

[0048] It is worth noting that the original FPN is different in that, due to the inconsistent spatial scales of the input sources, two different convolutional neural networks are used to sample the two types of data samples to the same scale and then fused, and then up-sampled according to the scale of the reanalysis field data. The different scale feature maps obtained are similar to the segmentation method of 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.), which decomposes the feature maps into 5x5 patches, and then flattens them according to the feature dimension to obtain spatial feature basic units (Tokens). The encoding of TC intensity information is relatively simple, using a linear layer mapping to represent the intensity information basic units (Tokens).

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

[0050] Specifically, the cross attention calculation can be specifically represented 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, W Q , W k and W v represent the learnable mapping matrix parameters of the three, Q is generally an input sequence, K and V take the same input, which is another input sequence, and only Q is calculated with K and V of the historical time, not with the future time. In actual calculation, the intensity information basic unit sequence is taken as Q, the spatial feature basic unit is taken as K and V, the hidden space feature is obtained through cross attention calculation, and then the vector in the learnable discrete codebook is matched to obtain the closest vector. The quantized feature vector is input into the decoder. The quantized feature vector is taken as Q, and the reanalysis data is processed through the same process as the encoder to obtain the spatial feature basic unit as the condition K and V, and the cross attention calculation is performed with Q, and finally the feedforward neural network is used to map back to the original intensity information value.

[0053] Through the above steps of training, a discrete codebook representing the tropical cyclone intensity information in the hidden space can be obtained, which can be used for the training task in the next stage.

[0054] In the second stage, the wind and cloud forecast field data are input into the decoder together with the prediction hidden variable to obtain a preliminary result; the wind and cloud forecast field data are processed through a physical constraint model to obtain a prediction result, and the preliminary result is superimposed to obtain the final tropical cyclone intensity prediction result.

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

[0056] The input historical meteorological data is encoded by an encoder and a discrete codebook to obtain a 4-step sequence length hidden variable, and then iterated for T steps by a hidden space vector iteration model to obtain the hidden variable of the T-step forecast. The atmospheric field data predicted by the wind and cloud is used to replace the ERA5 reanalysis atmospheric field data originally input into the decoder, and the hidden variable is input into the decoder to obtain a preliminary result. At the same time, the atmospheric field data predicted by the wind and cloud is also output by a physical constraint model, and is superimposed on the preliminary prediction result to obtain the final prediction of the tropical cyclone intensity at the T-step in the future.

[0057] The present scheme transfers the prediction task to the hidden space to solve the problem of asymmetric input and output that cannot be self-recurrently predicted. The hidden space vector iteration model uses the causal cross-attention calculation module in MSCAR, that is, the information at the current time only performs cross-attention calculation with the information at the current time and the historical time. 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 conforming to the physical law of tropical cyclones, using a learnable sparse matrix to select the basis function, and obtaining the physically constrained tropical cyclone intensity prediction result.

[0058] The present embodiment performs a simple experiment, trains a model with 24-hour historical input and 24-hour prediction output, and then tests the extension to 120-hour prediction. The test is performed on the data in 2022, and the results are shown in Figure 3 It can be found that the error growth of long-term prediction is significantly reduced, which meets our expectations.

[0059] The present embodiment also provides a tropical cyclone intensity prediction device based on a meteorological large model, which includes a memory and a processor. 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 the meteorological large model as described above.

[0060] The preferred embodiments of the present application have been described above in detail. It should be understood that modifications and variations to the preferred embodiments could be made by those skilled in the art in light of the teachings above. It is therefore contemplated that the application can encompass other variations and modifications that fall within the scope of the claims.

Claims

1. A method for predicting the intensity of tropical cyclones based on a large meteorological model, characterized in that, Includes the following stages: The first stage involves acquiring satellite images, reanalysis atmospheric field data, and tropical cyclone intensity information for the area to be tested in advance. Using the satellite images and reanalysis atmospheric field data as conditions, the tropical cyclone intensity information is mapped to a discrete vector latent space through an encoder. The mapping result is compared with all vectors in the learnable discrete codebook. The closest vector is taken and used to restore the original tropical cyclone intensity information through a decoder. The encoder, discrete codebook, and decoder are then iteratively trained. The second stage involves acquiring historical meteorological data and wind and cyclone forecast field data for the area to be measured. The historical meteorological data is then input into the trained encoder and discrete codebook, transformed into discrete latent variables, and subjected to autoregressive iteration to obtain forecast latent variables. Using the wind and cyclone forecast field data as physical constraints, the tropical cyclone intensity information is restored through the trained decoder to obtain the tropical cyclone intensity prediction results. In the first stage, the encoder uses a convolutional neural network to downsample satellite images and reanalysis atmospheric field data, outputting feature maps at different scales. Then, a top-down feature fusion mechanism is adopted, using linear interpolation to gradually upsample from deep feature maps and fuse them with shallow feature maps of the same scale output in the previous downsampling process to finally obtain multi-scale features. The feature maps at each scale are decomposed into patches and then flattened according to the feature dimensions to obtain basic spatial feature units. The intensity information of tropical cyclones is characterized as basic units 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 the latent space features. Then, they are matched with the vectors in the learnable discrete codebook. The closest vector is taken to obtain the quantized feature vector, which is then input into the decoder to restore the original tropical cyclone intensity. The decoder uses the quantized feature vector as the query matrix, and performs downsampling and feature fusion operations on the reanalyzed atmospheric field data through the encoder to obtain the basic spatial feature units, which are used as the key matrix and value matrix. Cross-attention calculation is then performed, and finally, a feedforward neural network is used to map back to the original intensity information values.

2. The method for predicting the intensity of tropical cyclones based on a large meteorological model according to claim 1, characterized in that, For satellite images and reanalysis atmospheric field data with inconsistent spatial scales, the encoder uses two different convolutional neural networks to downsample the satellite images and reanalysis atmospheric field data to the same scale before stitching and fusing them, and then upsamples them according to the scale of the reanalysis atmospheric field data.

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

4. The method for predicting the intensity of tropical cyclones based on a large meteorological model according to claim 1, characterized in that, In the second stage, the latent space vector iterative model is used to iterate the discrete latent variables for T steps to obtain the latent variables for the Tth step forecast, which are then used to predict the intensity of the tropical cyclone in the future Tth step.

5. The method for predicting the intensity of tropical cyclones based on a large meteorological model according to claim 4, characterized in that, The latent space vector iterative model is the causal cross-attention calculation module in MSCAR, which performs cross-attention calculation only on the information at the current moment and the information at the previous moment and the previous moment.

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

7. The method for predicting the intensity of tropical cyclones based on a large meteorological model according to claim 6, characterized in that, The physical constraint model sets basis functions that conform to the physical laws of tropical cyclones and selects basis functions using learnable sparse matrices to obtain the nonlinear characteristics of tropical cyclone intensity, thereby improving the accuracy of tropical cyclone intensity prediction.

8. A tropical cyclone intensity prediction device based on a large meteorological model, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, and the processor invokes the computer program to perform the steps of the method as described in any one of claims 1 to 7.

Images