Tropical cyclone strength change mechanism analysis method and device based on KAN

Abstract

The invention relates to a KAN-based tropical cyclone strength change mechanism analysis method and device, and belongs to the technical field of TC strength change prediction. The method comprises the steps that a factor recursive pruning algorithm based on a KAN2.0 attribution score is adopted, and a high-impact forecasting factor set for TC intensity change forecasting is screened and obtained; performing factor contribution analysis based on symbolization regression by using KAN2.0 trained by a high-impact forecasting factor set, obtaining a complete linear equation from high-impact forecasting factors to TC intensity change, and calculating the TC intensity change according to the symbols and absolute values of various linear coefficients in the equation; and quantitatively analyzing the contribution direction and contribution degree of each high-influence forecasting factor to the TC intensity change. According to the method, the stability, the accuracy, the generalization ability and the interpretability of TC intensity forecasting can be effectively improved.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of TC (tropical cyclone) intensity change prediction, in particular to a tropical cyclone intensity change mechanism analysis method and device based on KAN. BACKGROUND

[0002] A TC is a strong cyclonic vortex generated on the tropical or subtropical ocean surface, and its intensity change is controlled by multiple-scale physical processes. The interaction mechanism between its internal dynamics and external environment has not been fully understood, so accurate TC intensity prediction remains a formidable challenge.

[0003] Deep learning (DL) models have shown great potential in TC intensity prediction due to their automatic feature extraction and complex nonlinear relationship mining capabilities. For example, some research has proposed a TC-Rolling model based on a convolutional neural network, established a correlation between TC convective structure and its intensity, and used a rolling strategy to effectively reduce the TC intensity multi-time step prediction bias and improve the accuracy of TC short-term prediction. Some research has also proposed a deep learning-driven TC intensity prediction framework, Deep-TCP, which integrates data constraint modules, spatiotemporal attention modules, and feature enhancement modules, effectively fusing multi-source data and significantly improving the accuracy and robustness of the prediction.

[0004] However, due to the "black box" nature of intelligent models, it is difficult to analyze their internal decision-making process, which largely limits the model's explainability. To analyze the contribution of prediction factors to TC intensity change, some research has used Garson variable importance scores to rank 121 SHIPS (Statistical Hurricane Intensity Prediction Scheme) prediction factors. This method calculates the contribution of each input variable to the output result, thereby quantifying the relative importance of different features in model prediction. However, due to the large size of the prediction factors, the calculation cost of this method is high, and this method can only evaluate the overall contribution of each prediction factor, making it difficult to explain the contribution direction of each prediction factor to the prediction result, and the explainability of TC intensity prediction is poor. SUMMARY

[0005] Therefore, it is necessary to provide a tropical cyclone intensity change mechanism analysis method and device based on KAN to address the weak explainability of current TC intensity intelligent prediction, effectively screen high-impact prediction factors, and quantitatively analyze the contribution of each factor to TC intensity change, thereby achieving high-accuracy and strongly-explained TC intensity prediction.

[0006] A tropical cyclone intensity change mechanism analysis method based on KAN, the method comprising:

[0007] The factor recursive pruning algorithm based on the KAN2.0 attribution score is used to screen and obtain a high-impact forecast factor set for TC intensity change prediction. The KAN2.0 is trained using the high-impact forecast factor set, and the factor contribution analysis is performed based on the symbolic regression of KAN2.0. The analysis process includes: First, the symbol function library of the network is limited to a linear function, and the optimal linear function is obtained by performing linear function fitting on the activation function of each edge in the network to replace the original activation function of each edge, thereby realizing the symbolic regression of the network structure. Second, the layer-by-layer expression calculation is performed in the network structure symbolic regression network, including the sub-node input function cumulative summation, sub-node linear transformation, node linear transformation, and node output function cumulative summation of each layer of the network. Through the layer-by-layer expression calculation and propagation to the last layer of the network, the linear symbolic output of the entire network is obtained. Finally, based on the linear symbolic output of the entire network, the linear equation representing the relationship between the high-impact forecast factor and the TC intensity change is obtained, and according to the signs and absolute values of the linear coefficients in the linear equation, the contribution direction and contribution degree of each high-impact forecast factor to the TC intensity change are quantitatively analyzed.

[0008] In one embodiment, a factor recursive pruning algorithm based on the KAN2.0 attribution score is used to screen and obtain a high-impact forecast factor set for TC intensity change prediction, including: The full forecast factor set for TC intensity change prediction is initialized, and is divided into a training set, a validation set, and an independent test set; The training set and the validation set are input into KAN2.0 for iterative training, and the attribution score and the root mean square error of the validation set output by KAN2.0 are obtained. The attribution score is used to quantify the global contribution of each forecast factor to the TC intensity change, including linear contribution and nonlinear contribution. The last 10% of the forecast factors sorted by the attribution score are pruned and removed to form a new forecast factor set, which is re-input into KAN2.0 to perform the forecast factor pruning and removal operation until the number of forecast factors is less than a set value. According to the ranking of the root mean square error of the validation set output by KAN2.0 in each round, the forecast factor set with the lowest root mean square error of the validation set is selected as the high-impact forecast factor set.

[0009] In one embodiment, the full forecast factor set for TC intensity change prediction is initialized, including: An initial prediction factor set in SHIPS data and a time series prediction factor set reflecting time sequence characteristics of TC intensity change are acquired, combined to form a full prediction factor set for TC intensity change prediction, and preprocessed, including default value filling and standardization processing; wherein the time series prediction factor set supplements 8 time series prediction factors, which are sea level pressure changes in the past 6h, 12h, 18h and 24h, TC intensity changes in the past 6h, 12h and 18h, and a TC life history label, the TC life history label is used for sequentially numbering each moment of the TC life history.

[0010] In one of the embodiments, KAN2.0 is trained by using the high-impact prediction factor set, including: KAN2.0 is trained by using the high-impact prediction factor set, and a TPE Bayesian optimization algorithm is used to optimize network hyperparameters in the training process, to obtain an optimized KAN2.0 for year-by-year testing of TC intensity change prediction, and root mean square error and mean absolute error are used to evaluate the prediction stability.

[0011] In one of the embodiments, the optimal linear function is obtained by performing linear function fitting on the activation function of each edge in the network to replace the original activation function of each edge, to realize network structure symbolic regression, including: For each edge in the KAN2.0 network, the pre-activation value of the activation function of each edge is obtained based on the data sample and the post-activation value , and the optimal linear parameters are searched within a preset parameter range to fit the optimal mapping relationship from the pre-activation value to the post-activation value , to finally obtain the optimal linear function to replace the original activation function of each edge, to realize network structure symbolic regression; wherein the optimal linear parameters include optimal linear coefficients and optimal intercepts.

[0012] In one of the embodiments, layer-by-layer expression calculation is performed in the network structure symbolic regression network, including sub-node input function cumulative summation, sub-node linear transformation, node linear transformation and node output function cumulative summation of each layer of the network, through layer-by-layer expression calculation and propagation to the last layer of the network, to obtain the linear symbolic output of the whole network, including: In each layer of the KAN2.0 network with symbolic regression, all input functions passed to the same child node of the current layer are summed to obtain the preliminary output function of each child node in the current layer. Scaling and bias adjustments at the child node level are applied to the preliminary output functions of each child node in the current layer to complete the linear transformation of the child nodes, resulting in the preliminary output function of each node in the current layer. Node-level scaling and bias adjustments are applied to the preliminary output functions of each node in the current layer to complete the linear transformation of the nodes, resulting in the final output function of each node in the current layer. The final output functions of each node in the current layer are summed to obtain the input function of the next layer. This layer-by-layer expression is calculated and propagated to the last layer of the KAN2.0 network to obtain the overall linear symbolic output of the KAN2.0 network. Here, the input function, preliminary output function, and final output function are all linear functions.

[0013] In one embodiment, based on the overall linear symbolic output of the network, a linear equation representing the relationship between high-impact forecasting factors and TC intensity changes is obtained. Then, based on the sign and absolute value of the linear coefficients in the linear equation, the contribution direction and degree of each high-impact forecasting factor to TC intensity changes are quantitatively analyzed, including: Based on the overall linear symbolic output of the KAN2.0 network, a linear equation representing the relationship between high-impact forecasting factors and TC intensity changes is obtained, expressed as follows: ;in, This represents the change in TC intensity. Indicates the first One high-impact forecasting factor, Indicates the first The linear coefficients corresponding to the high-impact forecasting factors The intercept is... , The number of high-impact forecasting factors; Based on the sign and absolute value of each linear coefficient, the contribution direction and degree of each high-impact forecasting factor to the change in TC intensity are quantitatively analyzed. The sign of the linear coefficient indicates the direction of contribution of the high-impact forecasting factor to the change in TC intensity. When the linear coefficient is positive and the corresponding high-impact forecasting factor is positive / negative, it indicates that the corresponding high-impact forecasting factor has a positive / negative contribution to the change in TC intensity, which is conducive to the enhancement / weakening of TC intensity. When the linear coefficient is negative and the corresponding high-impact forecasting factor is positive / negative, it indicates that the corresponding high-impact forecasting factor has a negative / positive contribution to the change in TC intensity, which is conducive to the weakening / enhancing of TC intensity. The absolute value of the linear coefficient indicates the degree of contribution of the high-impact forecasting factor to the change in TC intensity; the larger the absolute value of the linear coefficient, the greater the contribution of the corresponding high-impact forecasting factor to the change in TC intensity.

[0014] In one embodiment, high-impact forecasting factors are classified into climate and persistence factors, kinetic factors, and thermodynamic factors; wherein, climate and persistence factors are used to reflect the historical evolution characteristics, climate background, and persistence status of TC; kinetic factors are used to reflect the kinetic characteristics of the large-scale environmental field in which TC is located; and thermodynamic factors are used to reflect the energy supply conditions for TC development.

[0015] In one embodiment, after quantitatively analyzing the direction and degree of contribution of each high-impact forecasting factor to the change in TC intensity, the method further includes: A set of high-impact forecasting factors that can reflect the direction and degree of contribution to TC intensity changes will be transferred to the training of the existing TC intensity change forecasting model until the trained TC intensity change forecasting model is converged to perform the TC intensity change forecasting task; wherein, the existing TC intensity change forecasting model is an MLP model.

[0016] A device for analyzing the mechanism of tropical cyclone intensity change based on KAN, the device comprising: The factor recursive pruning module is used to screen and obtain a set of high-impact forecast factors for TC intensity change forecasts using a factor recursive pruning algorithm based on KAN2.0 attribution scores. The factor contribution analysis module is used to train KAN2.0 using a set of high-impact forecast factors and perform factor contribution analysis based on symbolic regression of KAN2.0. The analysis process includes: First, limiting the symbolic function library of the network to linear functions, and obtaining the optimal linear function by performing linear function fitting on the activation function of each edge in the network to replace the original activation function of each edge, thus realizing symbolic regression of the network structure; Second, performing layer-by-layer expression calculation in the network of symbolic regression of the network structure, including summing the input functions of each child node in each layer of the network, linear transformation of child nodes, linear transformation of nodes, and summing the output functions of nodes. Through layer-by-layer expression calculation and propagation to the last layer of the network, the overall linear symbolic output of the network is obtained; Finally, based on the overall linear symbolic output of the network, the linear equation representing the relationship between high-impact forecast factors and TC intensity changes is obtained, and the contribution direction and degree of each high-impact forecast factor to TC intensity changes are quantitatively analyzed according to the sign and absolute value of the linear coefficients in the linear equation.

[0017] The above-mentioned method and apparatus for analyzing the intensity change mechanism of tropical cyclones based on KAN have the following advantages compared with existing technologies: 1. The factor recursive pruning algorithm based on KAN2.0 attribution scores is used to screen forecast factors. It can effectively screen out important forecast factors related to TC intensity changes, improve the stability and accuracy of TC intensity change forecasts, and can be generalized to traditional TC intensity change forecast models.

[0018] 2. By using KAN2.0 trained with a set of high-impact forecasting factors to perform factor contribution analysis based on symbolic regression, a complete linear equation from high-impact forecasting factors to TC intensity changes is obtained. Based on the sign and absolute value of each linear coefficient in the equation, the direction and degree of contribution of each high-impact forecasting factor to TC intensity changes are quantitatively analyzed. Thus, the data-driven KAN2.0 model can be used to accurately understand the physical mechanism of TC intensity changes, making the high-impact forecasting factors highly interpretable and reusable, and ultimately achieving high-accuracy and highly interpretable TC intensity forecasts. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating a method for analyzing the intensity variation mechanism of tropical cyclones based on KAN in one embodiment. Figure 2 This is a line graph showing the RMSE (Root Mean Square Error) and MAE (Mean Absolute Error) of the KAN2.0 validation set based on factor recursive pruning in one embodiment; where, Figure 2 (a) is a line graph of RMSE. Figure 2 (b) is a line graph of MAE; Figure 3 This is a flowchart illustrating the factor contribution analysis steps in one embodiment; Figure 4 This is a schematic diagram illustrating the annual test results of KAN2.0 for 24-hour TC intensity forecasts from 1982 to 2020 under both the full forecast factor set and the high-impact forecast factor set in one embodiment; wherein, Figure 4 (a) RMSE of the test results year by year. Figure 4 (b) MAE for annual test results; Figure 5 This is a schematic diagram of the attribution scores and linear coefficients of symbolic regression for high-impact forecasting factors with attribution scores greater than 0.1 in one embodiment. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0021] Before providing a further detailed description of the embodiments of this application, the nouns and terms involved in the embodiments of this application will be explained, and the nouns and terms involved in the embodiments of this application shall be interpreted as follows.

[0022] 1. Attribution Score: The attribution score is a quantitative metric used to measure the importance of nodes and edges in KAN 2.0. Its core principle is to iteratively propagate downstream contributions from the output layer, quantifying edge contributions layer by layer and summing them into node contributions, thereby measuring the global importance of input features. For an L-layer KAN, the number of nodes in each layer is... , define the first Layer Nodes The attribution score is ,node The corresponding number Strip edge The attribution score is but: ; ; in, , For the edge The standard deviation of the activation values, For nodes The standard deviation of activation values. Attribution scores, through the logic of iteratively passing from the output layer to the input layer, use the contribution of downstream nodes / edges as the core weight for the importance of upstream nodes / edges. Compared to the traditional L1 norm, attribution scores can overcome the limitations of local information, exclude locally active but globally ineffective nodes / edges, and effectively measure the global importance of input features.

[0023] 2. MAE and RMSE: MAE represents the average absolute error between the predicted and actual values, and RMSE is the square root of MAE. They are expressed as follows: ; ; in, For the observed values, Here, is the predicted value, and N is the total number of samples. The smaller the RMSE and MAE, the closer the predicted value is to the observed value, and the better the prediction effect.

[0024] In one embodiment, such as Figure 1 As shown, a method for analyzing the intensity variation mechanism of tropical cyclones based on KAN is provided, including the following steps: Step 1, Factor recursive pruning: The factor recursive pruning algorithm based on KAN2.0 attribution scores is used to screen and obtain the set of high-impact forecast factors for TC intensity change forecast.

[0025] Step 2, Factor Contribution Analysis: KAN2.0 is trained using a set of high-impact forecasting factors, and factor contribution analysis is performed based on symbolic regression of KAN2.0. The analysis process includes: First, the symbolic function library of the network is limited to linear functions, and the optimal linear function is obtained by fitting the activation function of each edge in the network to replace the original activation function of each edge, thus realizing symbolic regression of the network structure; Second, layer-by-layer expression calculation is performed in the network of symbolic regression of the network structure, including the summation of the input functions of each child node in each layer of the network, linear transformation of child nodes, linear transformation of nodes, and summation of the output functions of nodes. Through layer-by-layer expression calculation and propagation to the last layer of the network, the linear symbolic output of the network as a whole is obtained; Finally, based on the linear symbolic output of the network as a whole, a linear equation representing the relationship between high-impact forecasting factors and TC intensity changes is obtained, and the contribution direction and degree of each high-impact forecasting factor to TC intensity changes are quantitatively analyzed according to the sign and absolute value of the linear coefficients in the linear equation.

[0026] The KAN-based method for analyzing the mechanisms of tropical cyclone intensity changes has two advantages: first, the set of high-impact forecasting factors is selected based on model training performance, ensuring good forecasting results; second, by utilizing attribution scores and symbolic regression in KAN2.0, the impact of different high-impact forecasting factors on forecast results can be quantitatively analyzed, enabling a data-driven model to understand the physical mechanisms. This method can effectively improve the stability, accuracy, generalization ability, and interpretability of TC intensity forecasts.

[0027] A factor recursive pruning algorithm based on KAN2.0 attribution scores was used to screen and obtain a set of high-impact forecasting factors for TC intensity change forecasts, including: Initialize the full forecast factor set for TC intensity change forecast and divide it into training set, validation set and independent test set; The training and validation sets are input into KAN2.0 for iterative training to obtain the attribution score and root mean square error of the validation set output by KAN2.0. The attribution score is used to quantify the global contribution of each forecast factor to the change in TC intensity. The global contribution includes linear and nonlinear contributions. The bottom 10% of forecast factors ranked by attribution score are pruned and removed to form a new set of forecast factors. This new set of forecast factors is then re-entered into KAN2.0 to perform the forecast factor pruning and removal operation until the number of forecast factors is less than a set value. Based on the ranking of the root mean square error of the validation set output by KAN2.0 in each round, the set of forecast factors with the lowest root mean square error of the validation set is selected as the high-impact forecast factor set.

[0028] In one embodiment, a factor recursive pruning algorithm based on KAN2.0 attribution scores is used to screen and obtain a set of high-impact forecast factors for TC intensity change forecasts, including: Initialize the full forecast factor set for TC intensity change forecast and divide it into training set, validation set and independent test set; The training and validation sets are input into KAN2.0 for iterative training to obtain the attribution score and root mean square error of the validation set output by KAN2.0. The attribution score is used to quantify the global contribution of each forecast factor to the change in TC intensity. The global contribution includes linear and nonlinear contributions. The bottom 10% of forecast factors ranked by attribution score are pruned and removed to form a new set of forecast factors. This new set of forecast factors is then re-entered into KAN2.0 to perform the forecast factor pruning and removal operation until the number of forecast factors is less than the set value (10). Based on the ranking of the root mean square error of the validation set output by KAN2.0 in each round, the set of forecast factors with the lowest root mean square error of the validation set is selected as the high-impact forecast factor set.

[0029] In one embodiment, initializing the full forecast factor set for TC intensity change forecasts includes: By acquiring the initial forecast factor set (including 121 forecast factors) and the time-series forecast factor set reflecting the temporal characteristics of TC intensity changes from SHIPS data, a full forecast factor set for TC intensity change forecasts is formed. The full forecast factor set is then preprocessed, including filling in default values ​​and standardization. Compared with the initial forecast factor set, the time-series forecast factor set is supplemented with 8 additional time-series forecast factors, namely the sea level pressure changes over the past 6h, 12h, 18h, and 24h (DPRE-6, DPRE-12, DPRE-18, DPRE-24), the TC intensity changes over the past 6h, 12h, and 18h ​​(DELV-6, DELV-12, DELV-18), and the TC life history label (Life). The TC life history label is used to sequentially number each moment in the TC life history. For example, the TC inception moment is recorded as 1, the next moment (6h later) is recorded as 2, and so on, with the extinction moment recorded as n.

[0030] Specifically, for the default value, the characteristic average value of the intensity level of the TC in the sample is used to fill it; the maximum wind speed of TC for 1 minute and 10m is used to characterize the intensity of TC, and the forecast quantity is the change in TC intensity in the next 24 hours (dvs24), that is, the difference between the intensity of TC in the next 24 hours and the intensity at the current moment; the full forecast factor set from 1982 to 2018 is used as the training set, the full forecast factor set from 2019 to 2020 is used as the validation set, and the full forecast factor set from 2021 to 2022 is used as the independent test set.

[0031] Figure 2 Line plots of RMSE and MAE for the KAN2.0 validation set after recursive pruning of 129 full forecast factor sets are presented. Figure 2It can be seen that KAN2.0 has the best forecast performance for dvs24 when the number of forecast factors is 87, with a validation set RMSE of 10.79kt and MAE of 7.86kt. Therefore, these 87 factors are the high-impact forecast factor set for dvs24, including: vs0, TWXC_t24, DTL_t24, VMPI_t24, SHGC_t24, TLAT_t24, DELV-6, V500_t24, TWAC_t24, MTPW_v2, IR00_v16, MTPW_v10, CD26_t24, DELV-12, T250_t24, Z850_t24, IR00_v6, CD20_t24, V000_t24, and DELV-24. life, V20C_t24, MTPW_v4, SHDC_t24, PSLV_v6, RHHI_t24, PSLV_v5, IR00_v20, IR00_v13, TLON_t24, HE07_t24, G250_t 24. MTPW_v3, PENV_t24, U200_t24, VVAC_t24, V300_t24, E000_t24, IR00_v11, PSLV_v10, RHMD_t24, ENEG_t24, RHLO_t 24. SHTD_t24, HE05_t24, IR00_v2, EPOS_t24, T150_t24, V850_t24, PSLV_v4, D200_t24, T200_t24, G200_t24, COHC_t2 4. PSLV_v17, TADV_t24, IR00_v7, MTPW_v5, RSST_t24, SHTS_t24, PSLV_v18, PSLV_v3, SHRS_t24, DELV-18, VVAV_t24, I R00_v14, SHRD_t24, U20C_t24, CFLX_t24, G150_t24, IR00_v18, PSLV_v12, PSLV_v19, IR00_v5, O500_t24, PSLV_v14, R EFC_t24, MTPW_v19, PSLV_v16, O700_t24, CSST_t24, TGRD_t24, PSLV_v11, PSLV_v15, EPSS_t24, MTPW_v11, IR00_v17.

[0032] In one embodiment, KAN2.0 is trained using a set of high-impact forecast factors, including: KAN2.0 was trained using a set of high-impact forecasting factors. The TPE Bayesian optimization algorithm was employed during training to optimize the network hyperparameters. The optimized KAN2.0 was then used for year-by-year testing of TC intensity change forecasts, and the root mean square error (RMSE) and mean absolute error (MAE) were used to evaluate forecast stability. TPE is a tree-based Bayesian optimization algorithm used to solve the global optimization problem of black-box functions.

[0033] Specifically, the optimized KAN2.0 in this application has 1 hidden layer, 10 nodes, an optimizer named "Adam", a learning rate of 0.05, a regularization coefficient of 0.2, an L1 regularization coefficient of 1, an entropy regularization coefficient of 10, 1500 training steps, and no early stopping.

[0034] In one embodiment, such as Figure 3 As shown, to analyze the impact of various high-impact forecasting factors on the change in TC intensity, this application designed a factor contribution analysis step based on KAN2.0 symbolic regression. The core of this step is to transform the KAN2.0 network structure, which has high fitting performance, into an explicit mathematical symbolic expression, thereby quantitatively analyzing the degree and direction of contribution of each high-impact forecasting factor to dvs24. Specifically, this includes the following steps: After training KAN2.0 using a set of high-impact forecasting factors (referred to as the high-impact factor set), firstly, to facilitate the analysis of the contributions of the forecasting factors, the symbolic function library (lib) of KAN2.0 is limited to linear functions. For each edge in the KAN2.0 network, the pre-activation value of the activation function for each edge is obtained based on data samples. and post-activation value And search for the optimal linear parameters within the preset parameter range to adjust the pre-activation value. Activation value The mapping relationship is best fitted, and finally the optimal linear function is obtained. To replace the original activation function of each edge, symbolic regression of the network structure is achieved; where the optimal linear parameters include the optimal linear coefficients ( ) and optimal intercept ( ).

[0035] Secondly, in each layer of the KAN2.0 network with symbolic regression, all input functions passed to the same child node of the current layer are summed to obtain the preliminary output function of each child node in the current layer. Scaling and bias adjustments at the child node level are applied to the preliminary output functions of each child node in the current layer to complete the linear transformation of the child nodes, resulting in the preliminary output function of each node in the current layer. Node-level scaling and bias adjustments are applied to the preliminary output functions of each node in the current layer to complete the linear transformation of the nodes, resulting in the final output function of each node in the current layer. The final output functions of each node in the current layer are summed to obtain the input function of the next layer. This layer-by-layer expression is calculated and propagated to the last layer of the KAN2.0 network to obtain the overall linear symbolic output of the KAN2.0 network. Here, the input function, preliminary output function, and final output function are all linear functions. .

[0036] Then, based on the overall linear symbolic output of the KAN2.0 network, a linear equation representing the relationship between high-impact forecasting factors and TC intensity changes is obtained, expressed as follows: ;in, This represents the change in TC intensity. Indicates the first One high-impact forecasting factor, Indicates the first The linear coefficients corresponding to the high-impact forecasting factors The intercept is... , This represents the number of high-impact forecasting factors.

[0037] Finally, based on the sign and absolute value of each linear coefficient, the contribution direction and degree of each high-impact forecasting factor to the change in TC intensity are quantitatively analyzed. The sign of the linear coefficient indicates the direction of contribution of the high-impact forecasting factor to the change in TC intensity. When the linear coefficient is positive and the corresponding high-impact forecasting factor is positive / negative, it indicates that the corresponding high-impact forecasting factor has a positive / negative contribution to the change in TC intensity, which is conducive to the enhancement / weakening of TC intensity. When the linear coefficient is negative and the corresponding high-impact forecasting factor is positive / negative, it indicates that the corresponding high-impact forecasting factor has a negative / positive contribution to the change in TC intensity, which is conducive to the weakening / enhancing of TC intensity. The absolute value of the linear coefficient indicates the degree of contribution of the high-impact forecasting factor to the change in TC intensity; the larger the absolute value of the linear coefficient, the greater the contribution of the corresponding high-impact forecasting factor to the change in TC intensity. Furthermore, if it is assumed that only a single high-impact forecasting factor affects the change in TC intensity, the threshold at which this high-impact forecasting factor will cause the TC to strengthen or weaken in the future can be estimated by combining the linear intercept.

[0038] In one embodiment, the types of high-impact forecasting factors include climate and persistence factors, kinetic factors, and thermodynamic factors. Climate and persistence factors reflect the historical evolution characteristics, climate background, and persistence status of the TC, primarily including the current TC intensity (vs0), the distance of the TC from land (DTL_t24), the latitude of the 850 hPa vortex center (TLAT_t24), TC intensity variations (DELV-6, DELV-12, DELV-24), the TC lifespan marker (Life), and the zonal movement speed of the TC (PSLV_v6). Kinetic factors reflect the dynamic characteristics of the large-scale environmental field in which the TC is located, primarily including the environmental tangential wind (V50). The thermodynamic factors used to reflect the energy supply conditions for TC development mainly include the maximum potential intensity of TC (VMPI_t24), total precipitable water (MTPW_v2, MTPW_v10, MTPW_v4), brightness temperature (IR00_v16, IR00_v6), ocean heat content (CD26_t24, CD20_t24), and 250 hPa air temperature (T250_t24).

[0039] In one embodiment, after quantitatively analyzing the direction and extent of contribution of each high-impact forecasting factor to the change in TC intensity, the above method further includes: A set of high-impact forecasting factors that reflect the direction and degree of contribution to TC intensity changes will be transferred to the training of the existing TC intensity change forecasting model until convergence is achieved, allowing the model to perform TC intensity change forecasting tasks. The existing TC intensity change forecasting model is an MLP (Multilayer Perceptron) model. During training, the hyperparameters of the MLP model are set as follows: 5 hidden layers with 128, 256, 128, 256, and 256 nodes per layer, respectively; the optimizer is "Adam"; adaptive learning rate; activation function is "ReLU"; L2 regularization coefficient is 0.0005; maximum number of iterations is 200; and there is no early stopping.

[0040] To verify the beneficial effects of the method proposed in this application, the following experiments were also conducted, as detailed below: The forecast stability of KAN2.0 from 1982 to 2020 was analyzed using the Leave-One-Year-Out (LOYO) method. Figure 4The annual test results of KAN2.0 for 24-hour TC intensity forecasts from 1982 to 2020 are presented under the full forecast factor set (129) and the high-impact forecast factor set (87). Figure 4 The red line represents KAN-129, the blue line represents KAN-87, and the unmarked years indicate that the model's forecasts did not converge. Analysis Figure 4 It can be seen that KAN-129 exhibits poor forecast stability, with convergence occurring in six years (1988, 1990, 1993, 1994, 1995, and 2007). The average RMSE for years with convergence is 11.56kt, and the average MAE is 8.56kt. KAN-87, on the other hand, shows significantly improved forecast stability, with no years exhibiting convergence. It demonstrates good forecasting ability and accuracy in year-by-year testing, with an average RMSE of 11.45kt and an MAE of 8.42kt. Therefore, using a recursive pruning method to select the high-impact forecasting factor set can effectively improve the model's forecast stability and accuracy.

[0041] To evaluate the generalization ability of the models, the independent test results of KAN-129 and KAN-87 for dvs24 were compared. Table 1 shows the forecast errors of KAN-129, KAN-87, MLP-129, MLP-87, and the official operational forecasts and various deep learning models for the 24-hour TC intensity change forecast. As shown in Table 1, KAN-129 has an RMSE of 11.20kt and a MAE of 8.11kt, while KAN-87 has an RMSE of 10.90kt and a MAE of 7.87kt. Compared with the best official operational forecast (JTWC) and the best single deep learning model (MLP-87), KAN-87's MAE is reduced by approximately 4.6% and 5.7%, respectively. This demonstrates that KAN2.0 has strong generalization ability, and the high-impact factor set after recursive pruning effectively improves the forecast accuracy of KAN2.0 for dvs24. However, its forecasting performance is slightly lower than that of the current best hybrid models TC-Pred (ConvGRU) and Deep-TCP (ConvGPU).

[0042] Table 1. Comparison of errors between official forecasts and deep learning models for 24-hour TC intensity forecasts (Unit: kt; Model suffix "-87" represents the high-impact forecast factor set, and "-129" represents the full forecast factor set.)

[0043] It is worth noting that comparing the independent test results of MLP-129 and MLP-87, the RMSE of MLP-87 was 11.73kt and the MAE was 8.35kt. Compared with MLP-129, the RMSE and MAE decreased by about 8% and 7%, respectively. This shows that the high-impact forecasting factor set screened by KAN2.0 can not only improve the forecasting accuracy of KAN2.0, but also improve the forecasting accuracy of MLP. This indicates that the factor recursive pruning algorithm based on attribution scores can effectively screen out important forecasting factors related to TC intensity changes and can be applied to other intelligent models for TC intensity forecasting.

[0044] Regarding model interpretability, the focus is on analyzing the contribution of 25 high-impact forecasting factors with attribution scores greater than 0.1 to dvs24. Figure 5 This is a schematic diagram showing the attribution scores and linear coefficients of the symbolic regression for 25 high-impact forecasting factors with attribution scores greater than 0.1. The linear intercept is 4.67, and the factors are sorted from highest to lowest attribution score. Figure 5 In the horizontal axis, green represents climate and persistence factors, orange represents thermodynamic factors, and blue represents kinetic factors; Figure 5 Only values ​​with an attribution score greater than 0.1 or an absolute value of a linear coefficient greater than 1 are indicated.

[0045] Depend on Figure 5 It can be seen that for climate and persistence factors, vs0, Life, TLAT_t24, and PSLV_v6 all show negative contributions. Among them, vs0 has the largest negative contribution to dvs24, that is, the larger the vs0, the more favorable it is for TC weakening, which is consistent with the characteristics of TC intensity change. Life has the second largest impact on dvs24, meaning that the later the TC life cycle stage, the less favorable it is for TC enhancement. TLAT_t24 has a negative impact on dvs24, that is, the higher the TC latitude, the less favorable it is for TC intensity enhancement. PSLV_v6 also shows a negative contribution, that is, the greater the zonal movement speed of TC, the less favorable it is for TC intensity enhancement. Relatively speaking, DELV-6, DTL_t24, DELV-12, and DELV-24 show positive contributions. Among them, the impact of the past 6-hour TC intensity change (DELV-6) is the most important. That is, when the past TC intensity is enhanced, it is more favorable for the TC enhancement in the next 24 hours; and when the past TC intensity is weakened, it is more favorable for the TC weakening in the next 24 hours. DTL_t24 has the second largest positive contribution to dvs24, meaning that the farther the TC is from the land, the more beneficial it is to the enhancement of TC strength.

[0046] Regarding the dynamic factors, the maximum symmetrical tangential wind at 850 hPa in the next 24 hours (TWXC_t24) and the tangential wind at an average of 1000 hPa 500 km from the TC center (V000_t24) both contribute positively to dvs24. Among them, TWXC_t24 has the greatest impact on dvs24, showing that the larger TWXC_t24 is, the more favorable it is for TC intensification. The positive contribution of V000_t24 to dvs24 is relatively small. On the other hand, the vertical wind shear at 850–200 hPa in the next 24 hours (SHGC_t24 and SHDC_t24), the tangential wind at an average of 500 hPa 500 km from the TC center (V500_t24), and the symmetrical tangential wind at an average of 850 hPa 0–600 km from the TC center (TWAC_t24) all contribute negatively to dvs24. Among them, SHGC_t24 and SHDC_t24 have the largest negative contributions to dvs24, meaning that the stronger the vertical wind shear, the more favorable it is for TC to weaken. This is consistent with the influence mechanism of environmental wind vertical shear on TC intensity, that is, when the vertical shear of the horizontal wind is large, the resulting "ventilation effect" will lead to heat loss from the upper layers of the TC, thereby inhibiting the development of the TC. The negative contributions of V500_t24 and TWAC_t24 are relatively small. In addition, the influence of the 200hPa meridional wind (V20C_t24) on dvs24 in the next 24 hours should be distinguished between northerly and southerly winds. When the upper-level wind is northerly (V20C_t24<0), it will carry a large amount of dry and cold air into the TC. The stronger the northerly wind, the less favorable it is for the TC to strengthen. When the upper-level wind is southerly (V20C_t24>0), the southerly wind will carry a large amount of warm and humid air into the TC. The stronger the southerly wind, the more favorable it is for the TC to strengthen.

[0047] For thermodynamic factors, VMPI_t24, total precipitable water (MTPW_v10) at a distance of 800–1000 km from the TC center, the depth of the ocean 26°C isotherm (CD26_t24), and the depth of the ocean 20°C isotherm (CD20_t24) all contribute positively to dvs24. VMPI_t24 has the greatest impact on TC intensity variation; a larger VMPI_t24 favors 24-hour TC enhancement. MTPW_v10 has the second largest positive contribution to dvs24. The total precipitable water in this region reflects the humidity conditions of the large-scale environmental field; a wetter large-scale environmental field is more conducive to TC enhancement. Ocean heat content has a significant impact on dvs24. CD26_t24 and CD20_t24 both contribute positively to dvs24; that is, the deeper CD26_t24 and CD20_t24 are, the higher the ocean heat content, which is more conducive to TC enhancement. The factors that negatively contribute to dvs24, in descending order of contribution, are: the standard deviation of brightness temperature at a distance of 100-300 km from the TC center (IR00_v6), the total precipitable water at a distance of 0-200 km from the TC center (MTPW_v2), the minimum brightness temperature at a distance of 20-120 km from the TC center (IR00_v16), the atmospheric temperature at 250 hPa at a distance of 200-800 km from the TC center (T250_t24), and the total precipitable water at a distance of 200-400 km from the TC center (MTPW_v4). Specifically: the larger IR00_v6 is, the greater the asymmetry in the spatial distribution of brightness temperature near the TC eyewall, leading to the outward loss of upper-level potential vorticity and equivalent potential temperature, ultimately resulting in the loss of the TC warm core and a weakening of TC intensity. The higher IR00_v16 and T250_t24 are, the higher the TC outflow layer temperature, which is more unfavorable for TC enhancement. This aligns with the Carnot engine principle, which states that the greater the temperature difference between the ocean surface and the outflow layer, the higher the efficiency of the Carnot engine, leading to a greater increase in total temperature (TC).

[0048] In summary, the KAN-based tropical cyclone intensity change mechanism analysis method proposed in this application can quantitatively analyze the impact of different types of high-impact forecast factors on TC intensity changes. Furthermore, the impact mechanism obtained based on symbolic formula analysis is basically consistent with existing physical understanding. This demonstrates that this method can effectively understand complex physical laws using data-driven models and has strong interpretability.

[0049] It is worth noting that the ranking of forecast factors by attribution scores is not entirely consistent with the ranking of the absolute values ​​of their linear coefficients in symbolic regression. Most factors with high attribution scores have large absolute values ​​of their linear coefficients, while some factors with low attribution scores still have large absolute values ​​of their linear coefficients (e.g., Life, SHDC_t24), or factors with high attribution scores have small absolute values ​​of their linear coefficients (e.g., TLAT_t24, V500_t24, TWAC_t24). This is mainly because attribution scores measure the global importance of a factor to the forecast result, including both linear and nonlinear contributions. However, the linear coefficients obtained from the linear symbolic output only reflect the linear contribution of a factor to the forecast result, and therefore are not entirely consistent with the attribution scores.

[0050] In one embodiment, a KAN-based tropical cyclone intensity change mechanism analysis device is provided, comprising: The factor recursive pruning module is used to screen and obtain a set of high-impact forecast factors for TC intensity change forecasts using a factor recursive pruning algorithm based on KAN2.0 attribution scores. The factor contribution analysis module is used to train KAN2.0 using a set of high-impact forecast factors and perform factor contribution analysis based on symbolic regression of KAN2.0. The analysis process includes: First, limiting the symbolic function library of the network to linear functions, and obtaining the optimal linear function by performing linear function fitting on the activation function of each edge in the network to replace the original activation function of each edge, thus realizing symbolic regression of the network structure; Second, performing layer-by-layer expression calculation in the network of symbolic regression of the network structure, including summing the input functions of each child node in each layer of the network, linear transformation of child nodes, linear transformation of nodes, and summing the output functions of nodes. Through layer-by-layer expression calculation and propagation to the last layer of the network, the overall linear symbolic output of the network is obtained; Finally, based on the overall linear symbolic output of the network, the linear equation representing the relationship between high-impact forecast factors and TC intensity changes is obtained, and the contribution direction and degree of each high-impact forecast factor to TC intensity changes are quantitatively analyzed according to the sign and absolute value of the linear coefficients in the linear equation.

[0051] Specific limitations regarding the KAN-based tropical cyclone intensity change mechanism analysis device can be found in the limitations of the KAN-based tropical cyclone intensity change mechanism analysis method above, and will not be repeated here. Each module in the aforementioned KAN-based tropical cyclone intensity change mechanism analysis device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0052] The technical features of the above embodiments can be combined in any way. For the sake of brevity, 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.

[0053] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application.

Claims

1. A method for analyzing the mechanism of tropical cyclone intensity change based on KAN, characterized in that, The method comprises: An algorithm of recursive pruning of factors based on KAN2.0 attribution score is used to screen a high-impact forecast factor set for TC intensity change prediction; KAN2.0 is trained using the high-impact forecast factor set, and factor contribution analysis is performed based on the symbolic regression of KAN2.0, and the analysis process comprises: First, the symbol function library of the network is limited to a linear function, and the optimal linear function is obtained by performing linear function fitting on the activation function of each edge in the network to replace the original activation function of each edge, thereby realizing symbolic regression of the network structure; second, layer-by-layer expression calculation is performed in the network structure symbolic regression network, including the cumulative summation of the input function of each layer of the subnode, linear transformation of the subnode, linear transformation of the node, and cumulative summation of the output function of the node, and the linear equation of the relationship between the high-impact forecast factor and the TC intensity change is obtained through layer-by-layer expression calculation and propagation to the last layer of the network, and the contribution direction and contribution degree of each high-impact forecast factor to the TC intensity change are quantitatively analyzed according to the signs and absolute values of the linear coefficients in the linear equation.

2. The KAN-based tropical cyclone intensity change mechanism analysis method according to claim 1, wherein, An algorithm of recursive pruning of factors based on KAN2.0 attribution score is used to screen a high-impact forecast factor set for TC intensity change prediction, comprising: Initializing a full forecast factor set for TC intensity change prediction, and dividing it into a training set, a validation set and an independent test set; The training set and the validation set are input into KAN2.0 for iterative training, and the attribution score output by KAN2.0 and the root mean square error of the validation set are obtained; wherein the attribution score is used to quantify the global contribution of each forecast factor to the TC intensity change, which includes linear contribution and nonlinear contribution; The last 10% of the forecast factors sorted by the attribution score are pruned and removed to form a new forecast factor set, which is re-input into KAN2.0 to perform forecast factor pruning and removal operations until the number of forecast factors is less than a set value, and the forecast factor set with the lowest root mean square error of the validation set is selected as the high-impact forecast factor set according to the sorting of the root mean square error of the validation set output by KAN2.0 in each round.

3. The KAN-based tropical cyclone intensity change mechanism analysis method according to claim 2, wherein, Initializing a full forecast factor set for TC intensity change prediction, comprising: Obtaining an initial forecast factor set in SHIPS data and a time series forecast factor set reflecting the time series characteristics of TC intensity change, combining to form a full forecast factor set for TC intensity change prediction, and preprocessing the full forecast factor set, including default value filling and standardization processing; wherein the time series forecast factor set supplements eight time series forecast factors, which are sea level pressure changes in the past 6h, 12h, 18h and 24h, TC intensity changes in the past 6h, 12h and 18h, and a TC life history label, which is used to sequentially number each time of the TC life history.

4. The KAN-based tropical cyclone intensity change mechanism analysis method according to claim 2, characterized in that, Training KAN2.0 using the high-impact forecast factor set, comprising: The high-influence predictor set is used to train KAN2.0, and the training process uses the TPE Bayesian optimization algorithm to optimize the network hyperparameters to obtain the optimized KAN2.0 for year-by-year testing of TC intensity change prediction, and the root mean square error and the mean absolute error are used to evaluate the prediction stability.

5. The KAN-based tropical cyclone intensity change mechanism analysis method according to claim 1, wherein, By performing linear function fitting on the activation function of each edge in the network, the optimal linear function is obtained to replace the original activation function of each edge, realizing symbolic regression of the network structure, including: For each edge in the KAN2.0 network, a pre-activation value of an activation function of each edge is obtained based on a data sample and a post-activation value , and an optimal linear parameter is searched within a preset parameter range to fit an optimal mapping relationship of the pre-activation value to the post-activation value , and finally an optimal linear function is obtained to replace an original activation function of each edge, so as to realize network structure symbolic regression; wherein the optimal linear parameter comprises an optimal linear coefficient and an optimal intercept.

6. The KAN-based tropical cyclone intensity change mechanism analysis method according to claim 5, wherein, The network structure symbolic regression network performs layer-by-layer expression calculation, including the accumulation sum of the input function of each layer of the network, the linear transformation of the subnode, the linear transformation of the node, and the accumulation sum of the output function of the node, which is propagated to the last layer of the network through layer-by-layer expression calculation to obtain the linear symbolic output of the entire network, including: In each layer of the KAN2.0 network of the network structure symbolic regression, the input functions of the same subnode passing through the current layer are accumulated and summed to obtain the preliminary output function of each subnode in the current layer; the preliminary output function of each subnode in the current layer is subjected to scaling and bias adjustment at the subnode level to complete the linear transformation of the subnode and obtain the preliminary output function of each node in the current layer; the preliminary output function of each node in the current layer is subjected to scaling and bias adjustment at the node level to complete the linear transformation of the node and obtain the final output function of each node in the current layer; the final output function of each node in the current layer is accumulated and summed to obtain the input function of the next layer; the linear symbolic output of the entire KAN2.0 network is obtained by layer-by-layer expression calculation and propagation to the last layer of the KAN2.0 network; wherein the input function, the preliminary output function and the final output function are all in the form of linear functions.

7. The KAN-based tropical cyclone intensity change mechanism analysis method according to claim 6, wherein, Based on the linear symbolic output of the entire network, the linear equation representing the relationship between the high-influence predictor and the TC intensity change is obtained, and according to the sign and absolute value of the linear coefficient in the linear equation, the contribution direction and contribution degree of each high-influence predictor to the TC intensity change are quantitatively analyzed, including: Based on the overall linear symbolic output of the KAN2.0 network, a linear equation representing the relationship between high-impact forecasting factors and TC intensity changes is obtained, expressed as follows: ;in, This represents the change in TC intensity. Indicates the first One high-impact forecasting factor, Indicates the first The linear coefficients corresponding to the high-impact forecasting factors The intercept is... , The number of high-impact forecasting factors; According to the sign and absolute value of each linear coefficient, the contribution direction and contribution degree of each high-influence predictor to the TC intensity change are quantitatively analyzed; wherein the sign of the linear coefficient represents the contribution direction of the high-influence predictor to the TC intensity change, when the linear coefficient is positive and the corresponding high-influence predictor is positive / negative, it represents that the corresponding high-influence predictor has positive / negative contribution to the TC intensity change, which is conducive to the enhancement / weakening of the TC intensity; when the linear coefficient is negative and the corresponding high-influence predictor is positive / negative, it represents that the corresponding high-influence predictor has negative / positive contribution to the TC intensity change, which is conducive to the weakening / enhancement of the TC intensity; the absolute value of the linear coefficient represents the contribution degree of the high-influence predictor to the TC intensity change, the larger the absolute value of the linear coefficient, the greater the contribution of the corresponding high-influence predictor to the TC intensity change.

8. The KAN-based tropical cyclone intensity change mechanism analysis method according to claim 1 or 7, characterized in that, The high-influence prediction factors are classified into climate and persistence factors, dynamic factors, and thermodynamic factors; the climate and persistence factors are used to reflect the historical evolution characteristics, climate background, and self-persistence state of the TC; the dynamic factors are used to reflect the dynamic characteristics of the large-scale environmental field where the TC is located; and the thermodynamic factors are used to reflect the energy supply conditions for the development of the TC.

9. The KAN-based tropical cyclone intensity change mechanism analysis method according to claim 1 or 7, characterized in that, After quantitatively analyzing the contribution direction and contribution degree of each high-influence prediction factor to the TC intensity change, the method further comprises: The high-influence prediction factor set capable of reflecting the contribution direction and contribution degree to the TC intensity change is migrated to training of an existing TC intensity change prediction model until a trained TC intensity change prediction model is obtained to perform a TC intensity change prediction task; and the existing TC intensity change prediction model is an MLP model.

10. A device for analyzing the mechanism of tropical cyclone intensity change based on KAN, characterized in that, The device comprises: a factor recursive pruning module configured to filter and obtain a high-influence prediction factor set for TC intensity change prediction by using a factor recursive pruning algorithm based on a KAN2.0 attribution score; a factor contribution analysis module configured to train KAN2.0 by using the high-influence prediction factor set and perform factor contribution analysis based on symbolic regression of KAN2.0, and the analysis process comprises: first, limiting the symbolic function library of the network to a linear function, and replacing the original activation function of each edge in the network with an optimal linear function obtained by performing linear function fitting on the activation function of each edge in the network, to realize symbolic regression of the network structure; second, performing layer-by-layer expression calculation in the network structure symbolic regression network, including cumulative summation of the input function of each layer of the network, linear transformation of the sub-nodes, linear transformation of the nodes, and cumulative summation of the output function of the nodes, and through the layer-by-layer expression calculation and propagation to the last layer of the network, the linear symbolic output of the entire network is obtained; and finally, based on the linear symbolic output of the entire network, a linear equation representing the relationship between the high-influence prediction factor and the TC intensity change is obtained, and the contribution direction and contribution degree of each high-influence prediction factor to the TC intensity change are quantitatively analyzed according to the sign and absolute value of the linear coefficient in the linear equation.

Images