Intelligent forecasting method and device for tropical cyclone intensity based on feature adaptive optimization

Abstract

The present application relates to a method and device for intelligent forecasting of tropical cyclone intensity based on feature adaptive optimization. The method comprises: constructing an initial data set; dividing the initial data set into model training samples and independent test samples; selecting a pre-trained model; inputting the model training samples into the pre-trained model for feature adaptive optimization to obtain a set of high-impact factors that affect the accuracy of tropical cyclone intensity forecast; model optimization training; model robustness test and intelligent forecasting of tropical cyclone intensity. The present application can quickly extract the set of high-impact factors that are most critical to tropical cyclone intensity forecasting, significantly improving the accuracy of tropical cyclone intensity forecasting.

Description

Technical Field

[0001] The present application relates to the technical field of tropical cyclone intensity forecasting, and in particular to a method and device for intelligently forecasting tropical cyclone intensity based on feature adaptive optimization. Background Art

[0002] A tropical cyclone (TC) is a low-pressure system formed over tropical or subtropical oceans, usually accompanied by rotating air currents and strong winds. Accurate prediction of the path and intensity of a tropical cyclone is an important part of responding to extreme weather disasters and ensuring social security. Currently, mature forecasting technology has emerged for tropical cyclone path forecasting, but progress in tropical cyclone intensity forecasting has been slow.

[0003] Deep Learning (DL) models have shown great potential in the field of tropical cyclone intensity forecasting due to their ability to mine complex nonlinear relationships. Previous studies have used the Multilayer Perceptron (MLP) model for tropical cyclone intensity forecasting. This model is a DL model with multiple feedforward and fully connected hidden layers between the input layer and the output layer. The mean absolute error (MAE) in the 6-hour tropical cyclone intensity forecast is 1.68. , the model performs better than the simple linear regression model.

[0004] However, the MLP model is not accurate enough in predicting tropical cyclone intensity in practical applications. In addition, due to the lack of feature extraction engineering in the MLP model, it is difficult to fully explore the potential information in the data, and the generalization ability of the model is limited. In addition, the feature extraction method of the current deep learning model is not only time-consuming, but also has high requirements for computing resources. Therefore, it is urgent to propose a method that can quickly extract tropical cyclone intensity-related features and significantly improve the accuracy of tropical cyclone intensity forecasts. Summary of the invention

[0005] Based on this, it is necessary to provide a method and device for intelligent forecasting of tropical cyclone intensity based on feature adaptive optimization to address the above technical problems.

[0006] A method for intelligent forecasting of tropical cyclone intensity based on feature adaptive optimization, the method comprising:

[0007] Obtain SHIPS data and tropical cyclone intensity data within the study period, and preprocess the SHIPS data; the SHIPS data contains multiple forecasting factors that affect the accuracy of tropical cyclone intensity forecasts;

[0008] The tropical cyclone intensity data and pre-processed SHIPS data within the study period are organized into the initial data set, and the initial data set is divided into model training samples and independent test samples according to the time series;

[0009] The interpretable deep learning network KAN2.0 is selected as the pre-training model;

[0010] Input the model training samples into the pre-training model for feature adaptive optimization, obtain the weights of the prediction factors output by the model and the root mean square error of the intensity prediction, screen and prune the prediction factors based on the weight sorting, and repeatedly input the pruned prediction factor set into the pre-training model until the number of prediction factors is less than the set value, and screen the prediction factor sets of different numbers obtained in the feature adaptive optimization process according to the sorting of the root mean square error output by each round of the model to obtain a high-impact factor set;

[0011] The model training samples after screening the high impact factor set are input into the pre-training model, and the pre-training model is optimized and trained using the triple cross-validation algorithm and the TPE Bayesian optimization algorithm until the TCI-KAN intelligent forecasting model is obtained;

[0012] The model training samples after screening the high impact factor set are input into the TCI-KAN intelligent forecasting model for model robustness test, and the independent test samples after screening the high impact factor set are input into the TCI-KAN intelligent forecasting model for forecasting, and the tropical cyclone intensity forecast results are obtained as output.

[0013] In one embodiment, the SHIPS data is preprocessed, including:

[0014] The interquartile range algorithm was used to screen outliers and default values ​​in the SHIPS data. After replacing the outliers and default values ​​with the mean value, the standard deviation normalization algorithm was used to normalize the SHIPS data to obtain the preprocessed SHIPS data.

[0015] In one embodiment, the predictors are screened and pruned based on the weight ranking, and the pruned predictor set is repeatedly input into the pre-trained model until the number of predictors is less than a set value, including:

[0016] A weight analysis algorithm is used to screen the 10% of the prediction factors with the lowest weights output by the pre-trained model for pruning. The pruned prediction factor set and tropical cyclone intensity data are re-input into the pre-trained model for feature adaptive optimization, and the new root mean square error of the intensity forecast and the new prediction factor weight ranking are output. The above prediction factor pruning process is repeated until the number of prediction factors is less than 10.

[0017] In one embodiment, according to the ranking of the root mean square error output by each round of the model, different numbers of prediction factor sets obtained in the feature adaptive optimization process are screened to obtain a high-impact factor set, including:

[0018] The root mean square errors of the intensity predictions output by the pre-training model in each round are sorted, and when the root mean square error is the lowest, the prediction factor set of the pre-training model is input as the high-impact factor set.

[0019] In one embodiment, the model training samples after screening the high impact factor set are input into the pre-trained model, and the pre-trained model is optimized and trained using the triple cross-validation algorithm and the TPE Bayesian optimization algorithm until the TCI-KAN intelligent forecasting model is obtained, including:

[0020] Using the triple cross-validation algorithm, the model training samples after the high impact factor set screening were divided into three parts according to the year. One of them was selected as the test set in turn, and the other two were combined as the training set to input into the pre-training model for optimization training;

[0021] During the model optimization training process, the TPE Bayesian optimization algorithm is further combined to adjust the number of hidden layers, the number of hidden layer nodes, the regularization parameters and the number of training steps of the pre-trained model until a TCI-KAN intelligent forecasting model that meets the preset accuracy is obtained.

[0022] In one embodiment, the model training samples after screening the high impact factor set are input into the TCI-KAN intelligent forecasting model to perform a model robustness test, including:

[0023] The LOYO test scheme is adopted. The data of each year in the model training sample after the high impact factor set is screened is used as the test set, and the remaining data is used as the training set. The TCI-KAN intelligent forecasting model is tested year by year to test the robustness of the acquisition model.

[0024] In one embodiment, the independent test samples after screening the high impact factor set are input into the TCI-KAN intelligent forecast model for forecasting, and the tropical cyclone intensity forecast results are output, including:

[0025] The independent test samples screened by the high-impact factor set are input into the TCI-KAN intelligent forecast model for forecasting, and the tropical cyclone intensity forecast results are obtained as output. The forecast accuracy and generalization ability of the model are tested based on the mean absolute error and root mean square error between the tropical cyclone intensity forecast results and the true value.

[0026] A tropical cyclone intensity intelligent forecasting device based on feature adaptive optimization, the device comprising:

[0027] The data preprocessing module is used to obtain SHIPS data and tropical cyclone intensity data within the study period and preprocess the SHIPS data; the SHIPS data contains multiple forecasting factors that affect the accuracy of tropical cyclone intensity forecasts;

[0028] The data partitioning module is used to organize the tropical cyclone intensity data and the preprocessed SHIPS data within the study period into an initial data set, and divide the initial data set into model training samples and independent test samples according to the time series;

[0029] Model selection module, used to select the interpretable deep learning network KAN2.0 as the pre-training model;

[0030] A feature adaptive optimization module is used to input the model training samples into the pre-trained model for feature adaptive optimization, obtain the weights of the prediction factors output by the model and the root mean square error of the intensity prediction, screen and prune the prediction factors based on the weight sorting, and repeatedly input the pruned prediction factor set into the pre-trained model until the number of prediction factors is less than the set value, and screen the prediction factor sets of different numbers obtained in the feature adaptive optimization process according to the sorting of the root mean square error output by each round of the model to obtain a high-impact factor set;

[0031] The model optimization module is used to input the model training samples after screening the high impact factor set into the pre-trained model, and optimize the pre-trained model using the triple cross-validation algorithm and the TPE Bayesian optimization algorithm until the TCI-KAN intelligent forecasting model is obtained;

[0032] The model verification module is used to input the model training samples after the high impact factor set is screened into the TCI-KAN intelligent forecasting model for model robustness verification, and to input the independent test samples after the high impact factor set is screened into the TCI-KAN intelligent forecasting model for forecasting, and output the tropical cyclone intensity forecast results.

[0033] Compared with the existing tropical cyclone intensity forecasting methods, the above-mentioned tropical cyclone intensity intelligent forecasting method and device based on feature adaptive optimization have the following beneficial effects:

[0034] 1. This application uses KAN2.0 as the pre-trained model for tropical cyclone intensity forecasting. This model can achieve higher forecast accuracy than the MLP model with fewer parameters, and can effectively capture the periodicity and trend in the time series, making it show higher accuracy and efficiency in the time series forecasting task of tropical cyclone intensity forecasting.

[0035] 2. The feature adaptive optimization algorithm proposed in this application can effectively and quickly screen out the most critical set of high-impact factors for tropical cyclone intensity forecasting from the SHIPS data containing large-scale forecast factors, thereby reducing the consumption of computing resources and significantly improving the model training speed and the accuracy and generalization ability of the model intensity forecast. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 A schematic diagram of a flow chart of a method for intelligent forecasting of tropical cyclone intensity based on feature adaptive optimization in one embodiment;

[0037] Figure 2 A schematic diagram of a specific implementation process of a tropical cyclone intensity intelligent forecasting method based on feature adaptive optimization in one embodiment;

[0038] Figure 3 is a forecast result of the 6-hour intensity change of tropical cyclones from 1982 to 2022 in an embodiment; wherein, Figure 3 (a) is the root mean square error, Figure 3 (b) is the mean absolute error, unit: . DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0040] In one embodiment, Figure 1 and Figure 2 As shown, a method for intelligent forecasting of tropical cyclone intensity based on feature adaptive optimization is provided, comprising the following steps:

[0041] Step 1: Obtain SHIPS (Hurricane Intensity Statistical Forecast Scheme) data and tropical cyclone intensity data within the study period and preprocess the SHIPS data. The SHIPS data contains 121 forecast factors that affect the accuracy of tropical cyclone intensity forecasts.

[0042] Specifically, the research period is from 1982 to 2022, and the SHIPS data within the research period are preprocessed, including:

[0043] The interquartile range (IQR) algorithm was used to screen outliers and default values ​​in the SHIPS data. After replacing the outliers and default values ​​with the mean value, the standard deviation normalization algorithm was used to normalize the SHIPS data to obtain the preprocessed SHIPS data.

[0044] Step 2: The tropical cyclone intensity data within the study period and the preprocessed SHIPS data are organized into an initial data set, and the initial data set is divided into model training samples and independent test samples according to the time series.

[0045] Specifically, the model training samples are data from 1982 to 2020, and the independent test samples are data from 2021 to 2022.

[0046] Step 3: Select the interpretable deep learning network KAN2.0 as the pre-training model.

[0047] Among them, the KAN (Kolmogorov-Arnold Networks) architecture was proposed in April 2024. The KAN network places the activation function on the edge (connection) of the network, and each weight in the KAN network becomes a learnable parameter of a single variable function. KAN2.0 (Kolmogorov-Arnold Networks 2.0) is based on the original KAN, introduces multiplication nodes (i.e., MultKAN), and converts the KAN2.0 architecture into a tree diagram, which enhances the expressiveness, practicality, and interpretability of the KAN model. Compared with MLP, the KAN network can achieve higher forecasting accuracy than the MLP model with fewer parameters, and studies have shown that the KAN model can effectively capture the periodicity and trend in time series, making it show higher accuracy and efficiency in time series forecasting tasks. In addition, the KAN network provides model interpretability and interactivity that are difficult to achieve with the MLP model. It can output symbolic formulas and dynamic changes in module structure, which is helpful for the discovery and understanding of scientific laws, and its application scenarios are more extensive.

[0048] Step 4: Input the model training samples into the pre-trained model for feature adaptive optimization, obtain the weights of the prediction factors output by the model and the root mean square error of the intensity prediction, screen and prune the prediction factors based on the weight sorting, and repeatedly input the pruned prediction factor set into the pre-trained model until the number of prediction factors is less than the set value, and screen the prediction factor sets of different numbers obtained in the feature adaptive optimization process according to the sorting of the root mean square error output by the model in each round to obtain a set of high-impact factors.

[0049] Among them, the forecast factors are screened and pruned based on weight sorting, and the pruned forecast factor set is repeatedly input into the pre-trained model until the number of forecast factors is less than the set value, including: using a weight analysis algorithm to screen the 10% forecast factors with the lowest weights output by the pre-trained model for pruning, re-inputting the pruned forecast factor set and tropical cyclone intensity data into the pre-trained model for feature adaptive optimization, and outputting a new root mean square error of the intensity forecast and a new forecast factor weight ranking, and repeating the above forecast factor pruning process until the number of forecast factors is less than 10.

[0050] Among them, according to the ranking of the root mean square error of each round of model output, the prediction factor sets of different numbers obtained in the feature adaptive optimization process are screened to obtain a high impact factor set, including: sorting the root mean square error of the intensity prediction output of each round of the pre-training model, and when the root mean square error is the lowest, the prediction factor set of the pre-training model is input as the high impact factor set, with a total of 18 high impact factors, as shown in Table 1:

[0051]

[0052] The factors with the suffix "t6" in Table 1 represent the 6-hour forecast values ​​in the SHIPS dataset, and the forecast factors are ranked from high to low in importance. NCEP reanalysis data is a set of global climate and meteorological element data generated by assimilating a variety of observational data around the world and using advanced numerical models for calculations. It is widely used in meteorology and climate research.

[0053] It can be understood that this feature adaptive optimization algorithm is based on weight and root mean square error sorting, and can quickly screen out 18 key factors affecting tropical cyclone intensity forecast from 121 forecast factors in SHIPS data, thereby reducing the consumption of computing resources and significantly improving the model training speed and the accuracy and generalization ability of the model intensity forecast.

[0054] Step 5: Input the model training samples after screening the high impact factor set into the pre-trained model, and use the triple cross-validation algorithm and TPE Bayesian optimization algorithm to optimize the pre-trained model until the TCI-KAN (Typhoon Cyclone Intensity-Kolmogorov-Arnold Networks) intelligent forecasting model is obtained.

[0055] Among them, the model training samples after the high impact factor set screening are the SHIPS data and tropical cyclone intensity data of 18 high impact factors from 1982 to 2020. Specifically, the specific implementation process of step 5 is: using the triple cross-validation algorithm, the model training samples after the high impact factor set screening are divided into three parts according to the year, one of which is selected as the test set in turn, and the other two are combined as the training set to input the pre-trained model for optimization training; in the process of model optimization training, the TPE (Tree Structured Parzen Estimator) Bayesian optimization algorithm is further combined to adjust the number of hidden layers, the number of hidden layer nodes, the regularization parameters and the number of training steps of the pre-trained model until a TCI-KAN intelligent forecasting model that meets the preset accuracy is obtained.

[0056] Step 6: Input the model training samples after the high impact factor set is screened into the TCI-KAN intelligent forecasting model for model robustness test, and input the independent test samples after the high impact factor set is screened into the TCI-KAN intelligent forecasting model for forecasting, and output the tropical cyclone intensity forecast result.

[0057] Specifically, the specific implementation process of step 6 includes:

[0058] Firstly, the LOYO test scheme is adopted. The data of each year in the model training samples after the high impact factor set is screened is used as the test set, and the remaining data is used as the training set. The TCI-KAN intelligent forecasting model is tested year by year to test the robustness of the acquisition model.

[0059] Among them, the LOYO (leave one year out) test scheme uses data from all years except one year to train the model, and then uses the years that did not participate in the training to evaluate the model's forecasting performance.

[0060] Secondly, the independent test samples after screening the high-impact factor set were input into the TCI-KAN intelligent forecasting model for forecasting, and the tropical cyclone intensity forecast results were obtained as output. The forecast accuracy and generalization ability of the model were tested based on the mean absolute error and root mean square error between the tropical cyclone intensity forecast results and the true value.

[0061] Specifically, Figure 3 As shown, based on the feature adaptive optimization-based intelligent forecasting method for tropical cyclone intensity proposed in this application, the 6-hour intensity change forecast of tropical cyclones from 1982 to 2022 is carried out. The forecast object is the 6-hour intensity change of tropical cyclones: the maximum wind speed near the center of the tropical cyclone 6 hours later minus the maximum wind speed near the center of the tropical cyclone at the current moment. Figure 3 It can be seen that the average root mean square error of the tropical cyclone 6-hour intensity change forecast from 1982 to 2020 is 1.76 , the mean absolute error is 1.38 , the average root mean square error of independent tests in 2021 and 2022 is 1.83 , the mean absolute error is 1.43 , among which, the mean absolute error from 1982 to 2020 is 1.68 ) was reduced by 18%, significantly improving the accuracy of 6h tropical cyclone intensity forecast. In addition, the feature adaptive optimization method proposed in this application can effectively and quickly screen out the most critical forecast factor set for tropical cyclone intensity forecast, thereby reducing the consumption of computing resources, significantly improving the model training speed and model forecast accuracy and generalization ability.

[0062] In one embodiment, a tropical cyclone intensity intelligent forecasting device based on feature adaptive optimization is provided, comprising:

[0063] The data preprocessing module is used to obtain SHIPS data and tropical cyclone intensity data within the study period and preprocess the SHIPS data; the SHIPS data contains multiple forecasting factors that affect the accuracy of tropical cyclone intensity forecasts;

[0064] The data partitioning module is used to organize the tropical cyclone intensity data and the preprocessed SHIPS data within the study period into an initial data set, and divide the initial data set into model training samples and independent test samples according to the time series;

[0065] Model selection module, used to select the interpretable deep learning network KAN2.0 as the pre-training model;

[0066] A feature adaptive optimization module is used to input the model training samples into the pre-trained model for feature adaptive optimization, obtain the weights of the prediction factors output by the model and the root mean square error of the intensity prediction, screen and prune the prediction factors based on the weight sorting, and repeatedly input the pruned prediction factor set into the pre-trained model until the number of prediction factors is less than the set value, and screen the prediction factor sets of different numbers obtained in the feature adaptive optimization process according to the sorting of the root mean square error output by each round of the model to obtain a high-impact factor set;

[0067] The model optimization module is used to input the model training samples after screening the high impact factor set into the pre-trained model, and optimize the pre-trained model using the triple cross-validation algorithm and the TPE Bayesian optimization algorithm until the TCI-KAN intelligent forecasting model is obtained;

[0068] The model verification module is used to input the model training samples after the high impact factor set is screened into the TCI-KAN intelligent forecasting model for model robustness verification, and to input the independent test samples after the high impact factor set is screened into the TCI-KAN intelligent forecasting model for forecasting, and output the tropical cyclone intensity forecast results.

[0069] For the specific definition of the intelligent forecasting device for tropical cyclone intensity based on feature adaptive optimization, please refer to the definition of the intelligent forecasting method for tropical cyclone intensity based on feature adaptive optimization above, which will not be repeated here. Each module in the above-mentioned intelligent forecasting device for tropical cyclone intensity based on feature adaptive optimization can be implemented in whole or in part through software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0070] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0071] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A tropical cyclone intensity intelligent forecasting method based on feature adaptive optimization, characterized in that: The method comprises: Acquire SHIPS data and tropical cyclone intensity data within a study period, and preprocess the SHIPS data; wherein the SHIPS data contains a plurality of forecast factors that affect the accuracy of tropical cyclone intensity forecast; The tropical cyclone intensity data and the pre-processed SHIPS data within the study period are collated into an initial data set, and the initial data set is divided into model training samples and independent test samples according to the time series; The interpretable deep learning network KAN2.0 is selected as the pre-training model; Inputting the model training samples into the pre-trained model for feature adaptive optimization, obtaining the weights of the prediction factors output by the model and the root mean square error of the intensity prediction, screening and pruning the prediction factors based on the weight sorting, and repeatedly inputting the pruned prediction factor set into the pre-trained model until the number of prediction factors is less than the set value, and screening the prediction factor sets of different numbers obtained in the feature adaptive optimization process according to the sorting of the root mean square error output by each round of the model to obtain a high-impact factor set; The model training samples after screening by the high impact factor set are input into the pre-training model, and the pre-training model is optimized and trained by using the triple cross validation algorithm and the TPE Bayesian optimization algorithm until the TCI-KAN intelligent forecasting model is obtained; The model training samples after the high impact factor set is screened are input into the TCI-KAN intelligent forecast model for model robustness test, and the independent test samples after the high impact factor set is screened are input into the TCI-KAN intelligent forecast model for forecasting, and the tropical cyclone intensity forecast result is output; Among them, the predictor factors are screened and pruned based on the weight sorting, and the pruned predictor factor set is repeatedly input into the pre-training model until the number of predictor factors is less than the set value, including: The weight analysis algorithm is used to screen the 10% of the lowest-weighted forecast factors output by the pre-trained model for pruning. The pruned forecast factor set and tropical cyclone intensity data are re-input into the pre-trained model for feature adaptive optimization, and the new root mean square error of the intensity forecast and the new forecast factor weight ranking are output. The above forecast factor pruning process is repeated until the number of forecast factors is less than 10. Among them, according to the sorting of the root mean square error of each round of model output, the different numbers of prediction factor sets obtained in the feature adaptive optimization process are screened to obtain a high-impact factor set, including: sorting the root mean square error of the intensity forecast output by the pre-trained model in each round, and when the root mean square error is the lowest, the prediction factor set of the pre-trained model is input as the high-impact factor set, with a total of 18 high-impact factors, sorted as follows: near-surface wind speed in the center of the tropical cyclone 6 hours ago; near-surface wind speed in the center of the tropical cyclone at the current moment; minimum brightness temperature 20-120km from the center of the tropical cyclone in the satellite cloud map; distance from the center of the tropical cyclone to the land; tropical cyclone size estimation related parameters 2; tropical cyclone size estimation related parameters 1; within 50-200km from the center of the tropical cyclone, the brightness temperature is less than 40 ℃ proportion; 26℃ isotherm depth; average brightness temperature 20-120km from tropical cyclone center in satellite cloud images; average vertical velocity of air mass rise within 0-500km from tropical cyclone center; change of generalized 850-200hPa shear size over time; maximum brightness temperature 0-30km from tropical cyclone center in satellite cloud images; longitude of 850hPa low vortex center in NCEP reanalysis data; latitude of 850hPa low vortex center in NCEP reanalysis data; maximum symmetrical tangential wind at 850hPa in NCEP reanalysis data; average precipitable water within 0-1000km from tropical cyclone center; average precipitable water within 400-600km from tropical cyclone center; azimuth average of tangential wind at 500hPa altitude within 500km from the center.

2. The method according to claim 1, characterized in that: The SHIPS data is preprocessed, including: The interquartile range algorithm is used to screen outliers and default values ​​in the SHIPS data. After replacing the outliers and default values ​​with the mean value, the standard deviation normalization algorithm is used to normalize the SHIPS data to obtain the preprocessed SHIPS data.

3. The method according to claim 1, characterized in that The model training samples after screening the high impact factor set are input into the pre-training model, and the pre-training model is optimized and trained using the triple cross-validation algorithm and the TPE Bayesian optimization algorithm until the TCI-KAN intelligent forecasting model is obtained, including: Using a triple cross-validation algorithm, the model training samples after the high impact factor set screening are divided into three parts according to the year, one of which is selected as the test set in turn, and the other two are combined as the training set to input into the pre-trained model for optimization training; During the model optimization training process, the TPE Bayesian optimization algorithm is further combined to tune the number of hidden layers, the number of hidden layer nodes, the regularization parameters and the number of training steps of the pre-trained model until a TCI-KAN intelligent forecasting model that meets the preset accuracy is obtained.

4. The method according to claim 1, characterized in that The model training samples after screening the high impact factor set are input into the TCI-KAN intelligent forecasting model for model robustness testing, including: The LOYO test scheme is adopted, and the data of each year in the model training sample after the high impact factor set is screened is used as the test set, and the remaining data is used as the training set. The TCI-KAN intelligent forecasting model is tested year by year to test the robustness of the acquisition model.

5. The method according to claim 1, characterized in that The independent test samples after screening the high impact factor set are input into the TCI-KAN intelligent forecast model for forecasting, and the output is the tropical cyclone intensity forecast results, including: The independent test samples screened by the high-impact factor set are input into the TCI-KAN intelligent forecast model for forecasting, and the tropical cyclone intensity forecast results are obtained as output. The forecast accuracy and generalization ability of the model are tested based on the mean absolute error and root mean square error between the tropical cyclone intensity forecast results and the true values.

6. A tropical cyclone intensity intelligent forecasting device based on feature adaptive optimization, characterized in that: The device comprises: A data preprocessing module, used to obtain SHIPS data and tropical cyclone intensity data within a study period, and preprocess the SHIPS data; wherein the SHIPS data contains a plurality of forecast factors that affect the accuracy of tropical cyclone intensity forecast; A data partitioning module is used to organize the tropical cyclone intensity data and the preprocessed SHIPS data within the study period into an initial data set, and divide the initial data set into model training samples and independent test samples according to the time series; Model selection module, used to select the interpretable deep learning network KAN2.0 as the pre-training model; A feature adaptive optimization module, used for inputting the model training samples into the pre-trained model for feature adaptive optimization, obtaining the weights of the prediction factors output by the model and the root mean square error of the intensity prediction, screening and pruning the prediction factors based on the weight sorting, and repeatedly inputting the pruned prediction factor set into the pre-trained model until the number of prediction factors is less than a set value, and screening the prediction factor sets of different numbers obtained in the feature adaptive optimization process according to the sorting of the root mean square error output by each round of the model to obtain a high-impact factor set; The model optimization module is used to input the model training samples after the high impact factor set is screened into the pre-trained model, and optimize the pre-trained model using the triple cross-validation algorithm and the TPE Bayesian optimization algorithm until the TCI-KAN intelligent forecasting model is obtained; The model verification module is used to input the model training samples after the high impact factor set is screened into the TCI-KAN intelligent forecast model for model robustness verification, and input the independent test samples after the high impact factor set is screened into the TCI-KAN intelligent forecast model for forecasting, and output the tropical cyclone intensity forecast result; Among them, the predictor factors are screened and pruned based on the weight sorting, and the pruned predictor factor set is repeatedly input into the pre-training model until the number of predictor factors is less than the set value, including: The weight analysis algorithm is used to screen the 10% of the lowest-weighted forecast factors output by the pre-trained model for pruning. The pruned forecast factor set and tropical cyclone intensity data are re-input into the pre-trained model for feature adaptive optimization, and the new root mean square error of the intensity forecast and the new forecast factor weight ranking are output. The above forecast factor pruning process is repeated until the number of forecast factors is less than 10. Among them, according to the sorting of the root mean square error of each round of model output, the different numbers of prediction factor sets obtained in the feature adaptive optimization process are screened to obtain a high-impact factor set, including: sorting the root mean square error of the intensity forecast output by the pre-trained model in each round, and when the root mean square error is the lowest, the prediction factor set of the pre-trained model is input as the high-impact factor set, with a total of 18 high-impact factors, sorted as follows: near-surface wind speed in the center of the tropical cyclone 6 hours ago; near-surface wind speed in the center of the tropical cyclone at the current moment; minimum brightness temperature 20-120km from the center of the tropical cyclone in the satellite cloud map; distance from the center of the tropical cyclone to the land; tropical cyclone size estimation related parameters 2; tropical cyclone size estimation related parameters 1; within 50-200km from the center of the tropical cyclone, the brightness temperature is less than 40 ℃ proportion; 26℃ isotherm depth; average brightness temperature 20-120km from tropical cyclone center in satellite cloud images; average vertical velocity of air mass rise within 0-500km from tropical cyclone center; change of generalized 850-200hPa shear size over time; maximum brightness temperature 0-30km from tropical cyclone center in satellite cloud images; longitude of 850hPa low vortex center in NCEP reanalysis data; latitude of 850hPa low vortex center in NCEP reanalysis data; maximum symmetrical tangential wind at 850hPa in NCEP reanalysis data; average precipitable water within 0-1000km from tropical cyclone center; average precipitable water within 400-600km from tropical cyclone center; azimuth average of tangential wind at 500hPa altitude within 500km from the center.

Images