A tropical cyclone atmospheric duct prediction method and system based on an explainable XGBoost model

Abstract

The application belongs to the technical field of atmospheric waveguide judgment, and specifically provides a tropical cyclone atmospheric waveguide prediction method and system based on an explainable XGBoost model, wherein the method comprises the following steps: corresponding different features to meteorological parameters, tropical cyclone parameters and position parameters respectively, combining respective labels to form a data set; using a grid search algorithm to adjust the hyperparameters of the XGBoost algorithm on the training set, performing cross-validation training on the training set to obtain an optimal parameter combination; using SHAP to explain the prediction result of the XGBoost algorithm, analyzing the relationship between the tropical cyclone atmospheric waveguide and different features, and performing feature importance sorting; using the XGBoost model with the optimal parameter combination, inputting the to-be-tested features according to the feature importance corresponding weight, and predicting the atmospheric waveguide formation probability. Through the scheme, the relationship between the atmospheric waveguide formation and the features is obtained, which has a guiding effect on the atmospheric waveguide formation prediction in the tropical storm area, and is beneficial to combat command and early warning.

Description

Technical Field

[0001] This invention relates to the field of atmospheric waveguide identification technology, and more specifically, to a method and system for predicting atmospheric waveguides in tropical cyclones based on an interpretable XGBoost model. Background Technology

[0002] Evaporation waveguides, as a special atmospheric refraction structure, have a significant impact on the performance of shipborne radar and radio communication systems. Due to their importance to communication systems, the characteristics of evaporation waveguides in many regions of the world have been studied. Zhang Qi et al. (2016) investigated the spatiotemporal variations of evaporation waveguide height in the Gulf of Aden and found that the main influencing factor was the seasonal reverse monsoon. Zhao Xiaofeng et al. (2013) analyzed the statistical characteristics of evaporation waveguides in the tropical regions of the South China Sea and the eastern Indian Ocean; however, due to the limitations of data points and the discontinuity of observation time, waveguide climatology research has been difficult to conduct. Therefore, due to the low temporal resolution of datasets, past research has mainly focused on the monthly and seasonal variations of evaporation waveguides.

[0003] The evaporation waveguide model is a theoretical model that allows for convenient calculation of the evaporation waveguide height (EDH) using a relatively small number of easily detectable parameters. EDH is a key characteristic parameter used to determine the presence of an evaporation waveguide and to measure its intensity.

[0004] To better understand the patterns of atmospheric waveguides, it is necessary not only to study their own spatiotemporal variations but also to investigate the relationship between atmospheric waveguides and various weather systems. Therefore, determining the relationship between weather and atmospheric waveguides is a key factor in identifying atmospheric waveguides. Currently, there are no technologies that can predict atmospheric waveguide changes based on weather variations, and the determination of the various influencing factors on the formation of atmospheric waveguides in tropical cyclones is particularly urgent. Summary of the Invention

[0005] This invention addresses the technical problem of how to determine the relationship between weather and atmospheric waveguides in the prior art.

[0006] This invention provides a method for predicting atmospheric waveguides in tropical cyclones based on an interpretable XGBoost model, comprising:

[0007] S1, assigning meteorological parameters, tropical cyclone parameters, and location parameters to different characteristics x. i Combine their respective tags y i Combined into a dataset (x) i y i ), where i is a natural number;

[0008] S2, take 70% of the dataset as the training set, use the grid search algorithm to adjust the hyperparameters of the XGBoost algorithm on the training set, use the remaining 30% of the dataset as the test set, and perform cross-validation training on the training set to obtain the optimal parameter combination.

[0009] S3 uses SHAP to interpret the prediction results of the XGBoost algorithm and analyzes the relationship between atmospheric waveguides of tropical cyclones and different features, and ranks the features by importance.

[0010] S4 uses the XGBoost model with optimal parameter combination, inputs the features to be measured according to the weights corresponding to the importance of the features, and predicts the probability of atmospheric waveguide formation.

[0011] Preferably, S2 specifically includes:

[0012] The loss function is binary cross-entropy. When adjusting the parameter values, each parameter is first set within a fixed range based on prior knowledge. Then, a grid search algorithm is used to obtain all parameter combinations within the given range. Finally, cross-validation is used on the training set to calculate the performance of each parameter combination.

[0013] Preferably, the hyperparameters in the XGBoost algorithm in S3 include:

[0014] learning_rate is used to control the learning speed;

[0015] max_depth, the maximum depth of the decision tree;

[0016] n_estimators, the maximum number of decision trees;

[0017] min_child_weight, the minimum sum of weights required for a child object;

[0018] reg_lambda, L2 regularization term;

[0019] reg_alpha, L1 regularization term;

[0020] subsample controls the proportion of random sampling;

[0021] colsample_bytree controls the proportion of randomly extracted features;

[0022] gamma is the minimum decrease in the loss function required for node segmentation.

[0023] Preferably, step S2 further includes selecting the Kappa coefficient and classification accuracy as evaluation metrics, and the expression for the Kappa coefficient is as follows:

[0024]

[0025]

[0026] Where p0 is the percentage of correctly classified samples in each category out of the total number of samples, a1 and a2 represent the true number of samples in each category, and b1 and b2 represent the number of samples predicted as the two label categories. The relationship between the Kappa coefficient and the degree of consistency is as follows:

[0027]

[0028] Preferably, S1 specifically includes: acquiring feature parameters using a GPS-based drop-down detector;

[0029] The meteorological parameters include specific humidity, temperature, meridional wind and zonal wind; the tropical cyclone parameters include tropical cyclone class, maximum wind speed radius of tropical cyclone, quadrant of the detector position, and distance between the detector and the center of the tropical cyclone; the location parameters include longitude and latitude.

[0030] Preferably, the label is 0 or 1, where 1 represents no atmospheric waveguide under the feature, and 0 represents the opposite. In addition, an oversampling method is used to randomly and manually copy points in the sample category with a label value of 0, so that the number of samples corresponding to the labels of the two categories is the same.

[0031] Preferably, S3 specifically includes: finding the relationship between the atmospheric waveguide of a tropical cyclone and different features. The Shapely value is the predicted value assigned to the features in the dataset. The formula for the Shapely value is as follows:

[0032] z i =z base +h(x i1 )+h(x i2 )+...+h(x ij )

[0033] Where x i j z represents the j-th feature of the i-th sample. i z represents the predicted value of the i-th sample. base It is the model baseline value, h(x) ij ) is the contribution value of the j-th feature to the final prediction result of the i-th sample.

[0034] This invention also provides a tropical cyclone atmospheric waveguide prediction system based on an interpretable XGBoost model. The system is used to implement a tropical cyclone atmospheric waveguide prediction method based on an interpretable XGBoost model, comprising:

[0035] The feature acquisition module is used to map meteorological parameters, tropical cyclone parameters, and location parameters to different features x. i Combine their respective tags y i Combined into a dataset (x) i y i ), where i is a natural number;

[0036] The feature training module is used to take 70% of the dataset as the training set, use the grid search algorithm to adjust the hyperparameters of the XGBoost algorithm on the training set, and use the remaining 30% of the dataset as the test set to perform cross-validation training on the training set to obtain the optimal parameter combination.

[0037] The feature importance ranking module is used to interpret the prediction results of the XGBoost algorithm using SHAP, and to analyze the relationship between the atmospheric waveguide of tropical cyclones and different features, and to rank the features by importance.

[0038] The prediction module is used to predict the probability of atmospheric waveguide formation by using the XGBoost model with the optimal parameter combination and inputting the features to be measured according to the weights corresponding to the importance of the features.

[0039] The present invention also provides an electronic device, including a memory and a processor, wherein the processor is used to implement the steps of a tropical cyclone atmospheric waveguide prediction method based on an interpretable XGBoost model when executing a computer management program stored in the memory.

[0040] The present invention also provides a computer-readable storage medium storing a computer management program thereon, which, when executed by a processor, implements the steps of a tropical cyclone atmospheric waveguide prediction method based on an interpretable XGBoost model.

[0041] Beneficial Effects: This invention provides a method and system for predicting atmospheric ducts in tropical cyclones based on an interpretable XGBoost model. The method includes: mapping meteorological parameters, tropical cyclone parameters, and location parameters to different features, combining them with their respective labels to form a dataset; using 70% of the dataset as a training set, adjusting the hyperparameters of the XGBoost algorithm on the training set using a grid search algorithm, and using the remaining 30% of the dataset as a test set, performing cross-validation training on the training set to obtain the optimal parameter combination; using SHAP to interpret the prediction results of the XGBoost algorithm and analyzing the relationship between atmospheric ducts in tropical cyclones and different features, ranking the features by importance; and using the XGBoost model with the optimal parameter combination, inputting the features to be tested according to the weights corresponding to the feature importance to predict the probability of atmospheric duct formation. This scheme obtains the relationship between atmospheric duct formation and features, providing guidance for predicting atmospheric duct formation in tropical storm areas, which is beneficial for combat command and early warning. Attached Figure Description

[0042] Figure 1 A flowchart of a tropical cyclone atmospheric waveguide prediction method based on an interpretable XGBoost model provided by the present invention;

[0043] Figure 2 A schematic diagram of the hardware structure of a possible electronic device provided by the present invention;

[0044] Figure 3 A schematic diagram of the hardware structure of a possible computer-readable storage medium provided by the present invention;

[0045] Figure 4 The trajectory map of 164 tropical cyclones that occurred in the eastern Pacific and North Atlantic between 1996 and 2020 is provided for this invention. Detailed Implementation

[0046] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0047] Figure 1 The present invention provides a method for predicting atmospheric waveguides in tropical cyclones based on an interpretable XGBoost model, comprising:

[0048] S1, assigning meteorological parameters, tropical cyclone parameters, and location parameters to different characteristics x. i Combine their respective tags y i The data are combined into a dataset (xi, yi), where i is a natural number. Temperature, pressure, and humidity profiles obtained from GPS drop-down sensors are used to calculate the corrected refractive index profile. Since some sensors are located on land, elevation variations are also considered. Due to turbulence and random noise, the raw data is preprocessed using the Atmospheric Sensing Processing Environment (ASPEN) software released by the National Center for Atmospheric Research (NCAR) (http: / / www.eol.ucar.edu / software / aspen; ASPEN can analyze data, perform smoothing, sensor time-based correction, and eliminate suspicious data points). After calculating the corrected refractive index profile, each potential waveguide layer corresponds to a cutoff wavelength, which is a parameter measuring the waveguide's ability to trap electromagnetic waves. The longer the wavelength, the more stable the waveguide layer. The cutoff wavelength can be calculated using the following formula:

[0049]

[0050] Where λ maxδ is the cutoff wavelength (meters), d is the thickness of the waveguide layer (meters), and δM is the waveguide intensity (M). The constant C has a value of 5.66 × 10⁻⁶ for a suspended waveguide. -3 For surface waveguides, the value is 3.773 × 10⁻⁶. -3 In this invention, λ is considered to be... max A potential waveguide layer with a strength ≥0.5 is considered an effective waveguide layer. This constraint is used because it considers two characteristic parameters of the waveguide: too low a strength or too low a thickness may result in an inaccurate waveguide caused by instrument errors and random interference. Setting this condition filters out these inaccurate waveguides. Through these steps, negative effects are eliminated, and the true waveguide layer is obtained. Finally, 15216 profiles of 164 tropical cyclones from 1996 to 2020 are calculated as follows: Figure 4 As shown.

[0051] S2, take 70% of the dataset as the training set, use the grid search algorithm to adjust the hyperparameters of the XGBoost algorithm on the training set, use the remaining 30% of the dataset as the test set, and perform cross-validation training on the training set to obtain the optimal parameter combination.

[0052] S3 uses SHAP to interpret the prediction results of the XGBoost algorithm and analyzes the relationship between atmospheric waveguides of tropical cyclones and different features, and ranks the features by importance.

[0053] S4 uses the XGBoost model with optimal parameter combination, inputs the features to be measured according to the weights corresponding to the importance of the features, and predicts the probability of atmospheric waveguide formation.

[0054] In this invention, the temperature, pressure, and humidity data for studying the connection between tropical cyclones and atmospheric waveguides are derived from GPS drop-sonde instruments. These drop-sonde instruments were deployed by aircraft specifically designed for tropical cyclone research, launched by the National Oceanic and Atmospheric Administration (NOAA) between 1996 and 2020, with over 20,000 measurements conducted (https: / / www.aoml.noaa.gov / hrd / data_sub / dropsonde.html). The data includes variables such as temperature, humidity, wind speed, and wind direction, with a time resolution of 0.5 seconds and a vertical resolution of approximately 5-15 meters. In addition, the elevation data used in this invention comes from the SRTM15 Digital Elevation Model (DEM) data from NASA and NIMA. This dataset has a horizontal resolution of 450 meters and global coverage. By combining the temperature, pressure, humidity, and elevation data, a corrected refractive index vertical profile can be calculated.

[0055] The data on tropical cyclones are from the International Best Tracking Archive for Climate Management (IBTrACS) (ncdc.noaa.gov / ibtracs / index.php) published by the National Oceanic and Atmospheric Administration (NOAA)

[81] . The dataset has a time resolution of 3 hours, and based on prior knowledge and simple data analysis, the following four variables are selected as waveguide-related features: tropical cyclone level (divided into 11 levels, from -5 to 0 are unknown type, post-tropical cyclone, micro-disturbance, subtropical cyclone, tropical depression and tropical storm, from 1 to 5 are divided using Saffir-Simpson scale, based on the 10-minute average maximum sustained wind speed), the maximum wind speed radius of the tropical cyclone, the distance between the detector and the center of the cyclone and the positional relationship between the detector and the center of the cyclone (the position of the detector can be four types: left front, left rear, right front and right rear of the tropical cyclone, which are represented by 1, 2, 3 and 4 respectively in this invention, and are called the detector quadrants). When selecting data points for the tropical cyclone path corresponding to the detector, the data points corresponding to the time points closest to the time of the detector deployment are selected as the features of a certain atmospheric refractive index profile.

[0056] Secondly, the meteorological parameters of the standard pressure layers used in this study were obtained from the ECMWF ERA-5 dataset. According to a 2001 study by Kursinski et al., the highest height of waveguides generally does not exceed 5000 meters; therefore, only standard pressure layers below 5000 meters were considered. Thus, based on the correspondence between pressure and altitude, this invention selected a total of 16 pressure layers: 1000, 975, 950, 925, 900, 875, 850, 825, 800, 775, 750, 700, 650, 600, 550, and 500 hPa. For ease of description later, these pressure layers are divided into three parts: upper (500-600 hPa), middle (600-750 hPa), and lower (775-1000 hPa). The characteristic variables are the specific humidity, temperature, and zonal wind components of these pressure layers. Regarding the selection of grid points for the corresponding meteorological data, referring to the 2010 study by Shu Shoujuan and Peng Liran, even during Super Typhoon Longwang, the horizontal movement range of the drop-sonde was far less than 10 km. Since the study area was limited to an altitude of 5000 meters and the tropical cyclone was of a low intensity, the horizontal movement distance of the meteorological sensor was even smaller. Therefore, within a grid with a resolution of 0.25° × 0.25°, the meteorological sensor can be considered to have remained at a relatively fixed position, i.e., its latitude and longitude at the time of release. Thus, it is believed that the sensor can detect local meteorological parameters relatively accurately. Furthermore, in the selection of the data time, the hourly data closest to the time of the sensor's release was chosen.

[0057] In summary, a feature set is formed by combining local meteorological parameters, tropical cyclone information parameters, and the location of the detector, which is used to predict the generation of atmospheric ducts in tropical cyclones, as shown in Table 1.

[0058] Table 1 shows the characteristics used to predict the existence of tropical cyclone waveguides.

[0059]

[0060]

[0061] XGBoost is a variant of the gradient descent tree algorithm that avoids overfitting by adding a regularization term to the loss function. Its basic theory is as follows:

[0062]

[0063] Where t is the number of basic tree models, f t (x i ) represents the prediction result of the t-th tree for the i-th sample. This is the combined prediction result of t trees for the i-th sample. The goal of this algorithm is to achieve optimal performance for the ensemble model, which means minimizing the loss function.

[0064]

[0065] Where l is a 2-order differentiable convex function used to measure the true value y. i and predicted value The error between the nodes is given by T, where T is the number of nodes in the decision tree, and w is the weight of the j-th node among all leaf nodes. γ and λ are the node splitting difficulty and regularization coefficient, respectively. Ω(f) is the sum of the complexities of t trees, used as a penalty function.

[0066] Since XGBoost uses forward iteration, when focusing on the t-th tree, the prediction results of the first t-1 trees can be considered constant:

[0067]

[0068] Then, the loss function is expanded using a Taylor series, and the original loss function is rewritten as follows:

[0069]

[0070] Where g i h represents the first derivative i Let represent the second derivative. Combining all the equations, we get:

[0071]

[0072] To obtain the weight w j The loss function with respect to w j Find the first derivative to obtain w j The expressions for ζ are as follows:

[0073]

[0074] The hyperparameters in the XGBoost algorithm used in this invention include: learning_rate (to control the learning speed), max_depth (maximum depth of the decision tree), n_estimators (maximum number of decision trees), min_child_weight (minimum sum of weights required for a child object), reg_lambda (L2 regularization term), reg_alpha (L1 regularization term), subsample (controls the proportion of random sampling), colsample_bytree (controls the proportion of randomly extracted features), and gamma (minimum descent of the loss function required for node segmentation).

[0075] As for the features and labels, these features are used to predict whether atmospheric ducts are generated. Labels are 0 and 1, where 1 represents the absence of atmospheric ducts under that feature, and 0 represents the opposite. The dataset contains 5727 samples with a label of 0 and 9489 samples with a label of 1. There is a significant difference in the number of samples between the two categories. Therefore, the sampling method used in this invention is oversampling, that is, randomly and manually replicating points in the category of samples with a label value of 0 to make the number of samples in both categories equal.

[0076] Subsequently, the entire dataset was divided into a training set (70% of the total samples) and a test set (30% of the total samples). The loss function was binary cross-entropy. When adjusting the parameter values, each parameter was first set within a fixed range based on prior knowledge. Using a grid search algorithm, all parameter combinations within the given range were obtained, and the performance of each parameter combination was calculated on the training set using cross-validation. After extensive computer experiments, the optimal parameter combinations were obtained, as shown in Table 2.

[0077] Table 2. Best-performing hyperparameter combinations for the XGBoost model.

[0078]

[0079] Kappa coefficient and classification accuracy were chosen as evaluation metrics. The Kappa coefficient measures the consistency between observed and simulated data. It is a commonly used metric in machine learning, especially in spatial data matching. The expression for the Kappa coefficient is as follows:

[0080]

[0081] Where p0 is the percentage of correctly classified samples in each category out of the total sample size. a1 and a2 represent the true number of samples in each category, and b1 and b2 represent the number of samples predicted as belonging to these two categories, respectively. The relationship between the Kappa coefficient and the degree of consistency is as follows:

[0082]

[0083] Shapely Additional Explanation (SHAP) is a Python package for interpreting the output of any machine learning model. SHAP can quantitatively analyze the relationship between machine learning algorithm predictions and input variables. The SHAP value of each variable represents its impact and importance on the prediction. SHAP can rank variables based on their importance to the prediction, thus it can be used for feature selection. Furthermore, compared to the original XGBoost ranking, SHAP has two major advantages: consistency and accuracy. Consistency means that when the number of features changes, the original ranking of feature importance changes as little as possible. Accuracy means that when some features are removed from the feature set, the contribution of each feature to the total importance remains unchanged. The basic theory of SHAP is the Shapely value method proposed by Shapely in 1953, which resolves the contradictions caused by the distribution of benefits in cooperative processes and belongs to the field of cooperative game theory. One advantage of applying the Shapely value method is that benefits are distributed according to the marginal contribution rate of members to the alliance; that is, the benefit enjoyed by member i is equal to the average marginal benefit created by the member for participating in the alliance. The basic theory is as follows:

[0084]

[0085] Where n is the number of members in the cooperative system, N = {1, 2, ..., n}. S is a subset of N containing distinct sub-members. v(S) represents the total benefit of the alliance S. Let $v(S)$ represent the benefit gained by member $i$ in the alliance $S$. $S$ represents the number of members in the alliance. $n!$ represents the factorial of $n$, and $S\{i}$ represents the set after removing element $i$ from $S$. The marginal benefit created by member $i$ in different alliances $S$ is denoted as $[v(S) - v(S\{i})]$. The proportion of the benefit created by member $i$ in the entire alliance is denoted as $v(S) - v(S\{i})$.

[0086] The basic principle of Shapely values ​​is applied in SHAP. In machine learning, each feature in a feature set is a member of a consortium. The machine learning algorithm gives a prediction value for each sample; the Shapely value is the predicted value assigned to a feature in the dataset. The formula for the Shapely value is as follows:

[0087] z i =z base +h(x i1 )+h(x i2 )+...+h(x ij (10)

[0088] Where x ij z represents the j-th feature of the i-th sample. i z represents the predicted value of the i-th sample. base It is the model baseline value, h(x) ij This means the contribution of the j-th feature to the final prediction result of the i-th sample.

[0089] In this invention, SHAP is used to find the relationship between tropical cyclone atmospheric ducts and different features. Furthermore, the importance of various features is ranked to identify the most important features. In addition, the main reasons for the appearance of tropical cyclone atmospheric ducts in each profile can also be analyzed from this.

[0090] The trained model was tested on a test set, achieving a Kappa coefficient of 0.6258 and an accuracy of 81.30%. This indicates a high degree of consistency between the selected features and the atmospheric waveguide of tropical cyclones. To highlight the effectiveness of the algorithm, Logistic Regression was used as the benchmark model. Logistic Regression is a generalized linear regression analysis model belonging to supervised learning. Its derivation and calculation methods are similar to those of regression methods. However, it is primarily used to solve bisection problems. Here, it is used as the benchmark model for performance comparison with the XGBoost algorithm. Logistic Regression has the following parameters: regularization parameter (C), penalty term, and optimization method. Cross-validation was performed on the training set using a grid search algorithm, yielding the optimal parameter combination: C = 0.1, penalty term = L2, and optimization method = "lbfgs". The model achieved a test accuracy of approximately 69.5% and a test Kappa coefficient of approximately 0.391. The test results demonstrate that the XGBoost model significantly improves performance compared to Logistic Regression.

[0091] This invention constructs a dataset using different features and custom labels, and then builds an XGBoost model to predict the presence of waveguides in tropical cyclones. A grid search algorithm is used to perform cross-validation on the training set to obtain the optimal XGBoost model parameter combination, and the model's performance is tested on the test set, achieving an optimal accuracy of 81.30%.

[0092] Then, SHAP was used to interpret the results of the XGBoost model's prediction of atmospheric waveguides for tropical cyclones, obtaining the feature importance ranking of tropical cyclone characteristics, local meteorological parameters, and geographical location parameters, and specifically analyzing the impact of each relatively important feature on waveguide formation.

[0093] Using the method of this invention, an example of a tropical cyclone nestor was investigated, analyzing the main influencing factors of waveguide formation and the specific role of each factor. The main conclusions are as follows:

[0094] (1) The most important factor in the formation of a tropical cyclone is the distance between the weather vane and the center of the tropical cyclone. Local meteorological parameters are secondary, with humidity and temperature in the upper atmosphere being the most critical.

[0095] (2) Tropical cyclone waveguides are prone to form at the center of a tropical cyclone, while the probability of waveguide formation is very low away from the center. Subsequently, as the distance between the instrument and the center of the tropical cyclone increases, the probability of waveguide formation gradually increases, and after the distance increases to about 250 kilometers, the probability of waveguide formation remains basically unchanged. Secondly, this probability is positively correlated with the radius of the maximum wind speed. Furthermore, tropical cyclone waveguides are more likely to form to the right and rear of the tropical cyclone's orbit.

[0096] (3) Increased humidity in the mesosphere inhibits the formation of atmospheric waveguides. In addition, the mesosphere temperature is most favorable for the existence of atmospheric waveguides when it is between 270-275K.

[0097] This invention also provides a tropical cyclone atmospheric waveguide prediction system based on an interpretable XGBoost model. The system is used to implement a tropical cyclone atmospheric waveguide prediction method based on an interpretable XGBoost model, comprising:

[0098] The feature acquisition module is used to map meteorological parameters, tropical cyclone parameters, and location parameters to different features x. i Combine their respective tags y i Combined into a dataset (x) i y i ), where i is a natural number;

[0099] The feature training module is used to take 70% of the dataset as the training set, use the grid search algorithm to adjust the hyperparameters of the XGBoost algorithm on the training set, and use the remaining 30% of the dataset as the test set to perform cross-validation training on the training set to obtain the optimal parameter combination.

[0100] The feature importance ranking module is used to interpret the prediction results of the XGBoost algorithm using SHAP, and to analyze the relationship between the atmospheric waveguide of tropical cyclones and different features, and to rank the features by importance.

[0101] The prediction module is used to predict the probability of atmospheric waveguide formation by using the XGBoost model with the optimal parameter combination and inputting the features to be measured according to the weights corresponding to the importance of the features.

[0102] Please see Figure 2 This is a schematic diagram illustrating an embodiment of the electronic device provided in this invention. For example... Figure 2 As shown, this embodiment of the invention provides an electronic device, including a memory 1310, a processor 1320, and a computer program 1311 stored in the memory 1310 and executable on the processor 1320. When the processor 1320 executes the computer program 1311, it performs the following steps: S1, mapping meteorological parameters, tropical cyclone parameters, and location parameters to different features x. i Combine their respective tags y i Combined into a dataset (x) i y i ), where i is a natural number;

[0103] S2, take 70% of the dataset as the training set, use the grid search algorithm to adjust the hyperparameters of the XGBoost algorithm on the training set, use the remaining 30% of the dataset as the test set, and perform cross-validation training on the training set to obtain the optimal parameter combination.

[0104] S3 uses SHAP to interpret the prediction results of the XGBoost algorithm and analyzes the relationship between atmospheric waveguides of tropical cyclones and different features, and ranks the features by importance.

[0105] S4 uses the XGBoost model with optimal parameter combination, inputs the features to be measured according to the weights corresponding to the importance of the features, and predicts the probability of atmospheric waveguide formation.

[0106] Please see Figure 3 This is a schematic diagram illustrating an embodiment of a computer-readable storage medium provided by the present invention. (See diagram below.) Figure 3 As shown, this embodiment provides a computer-readable storage medium 1400, on which a computer program 1411 is stored. When the computer program 1411 is executed by a processor, it performs the following steps: S1, assigning meteorological parameters, tropical cyclone parameters, and location parameters to different features x. i Combine their respective tags y i Combined into a dataset (x) i y i ), where i is a natural number;

[0107] S2, take 70% of the dataset as the training set, use the grid search algorithm to adjust the hyperparameters of the XGBoost algorithm on the training set, use the remaining 30% of the dataset as the test set, and perform cross-validation training on the training set to obtain the optimal parameter combination.

[0108] S3 uses SHAP to interpret the prediction results of the XGBoost algorithm and analyzes the relationship between atmospheric waveguides of tropical cyclones and different features, and ranks the features by importance.

[0109] S4 uses the XGBoost model with optimal parameter combination, inputs the features to be measured according to the weights corresponding to the importance of the features, and predicts the probability of atmospheric waveguide formation.

[0110] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for predicting atmospheric waveguides in tropical cyclones based on an interpretable XGBoost model, characterized in that, include: S1, which assigns meteorological parameters, tropical cyclone parameters, and location parameters to different characteristics. Combine their respective tags Combined into a dataset , It is a natural number; S2, take 70% of the dataset as the training set, use the grid search algorithm to adjust the hyperparameters of the XGBoost algorithm on the training set, use the remaining 30% of the dataset as the test set, and perform cross-validation training on the training set to obtain the optimal parameter combination. S3 uses SHAP to interpret the prediction results of the XGBoost algorithm and analyzes the relationship between atmospheric waveguides of tropical cyclones and different features, and ranks the features by importance. Specifically, by finding the relationship between atmospheric waveguides of tropical cyclones and different features, the Shapely value is the predicted value assigned to the features in the dataset. The formula for the Shapley value is as follows: ； in Representing the The first sample One characteristic, Representing the Predicted values ​​for each sample, It is the model baseline value. It is the first The feature is related to the first The contribution value of each sample to the final prediction result; SHAP is a model interpretation package developed in Python, used to interpret the output of machine learning models and to quantitatively analyze the relationship between machine learning algorithm predictions and input variables. The SHAP value of each variable represents its impact and importance on the prediction. SHAP is also used to rank variables according to their importance to the prediction results. S4 uses the XGBoost model with optimal parameter combination, inputs the features to be measured according to the weights corresponding to the importance of the features, and predicts the probability of atmospheric waveguide formation.

2. The method for predicting atmospheric waveguides of tropical cyclones based on an interpretable XGBoost model according to claim 1, characterized in that, S2 specifically includes: The loss function is binary cross-entropy. When adjusting the parameter values, each parameter is first set within a fixed range based on prior knowledge. Then, a grid search algorithm is used to obtain all parameter combinations within the given range. Finally, cross-validation is used on the training set to calculate the performance of each parameter combination.

3. The method for predicting atmospheric waveguides of tropical cyclones based on an interpretable XGBoost model according to claim 2, characterized in that, The hyperparameters in the XGBoost algorithm in S3 include: learning_rate is used to control the learning speed; max_depth, the maximum depth of the decision tree; n_estimators, the maximum number of decision trees; min_child_weight, the minimum sum of weights required for a child object; reg_lambda, L2 regularization term; reg_alpha, L1 regularization term; subsample controls the proportion of random sampling; colsample_bytree controls the proportion of randomly extracted features; gamma is the minimum decrease in the loss function required for node segmentation.

4. The method for predicting atmospheric waveguides of tropical cyclones based on an interpretable XGBoost model according to claim 1, characterized in that, S1 specifically includes: acquiring feature parameters through a GPS-based drop-down detector; The meteorological parameters include specific humidity, temperature, meridional wind and zonal wind; the tropical cyclone parameters include tropical cyclone class, maximum wind speed radius of tropical cyclone, quadrant of the detector position, and distance between the detector and the center of the tropical cyclone; the location parameters include longitude and latitude.

5. The method for predicting atmospheric waveguides of tropical cyclones based on an interpretable XGBoost model according to claim 1, characterized in that, The label is 0 or 1, where 1 represents no atmospheric waveguide under the feature, and 0 represents the opposite. In addition, an oversampling method is used to randomly and manually copy points in the sample category with a label value of 0, so that the number of samples corresponding to the labels of the two categories is the same.

6. A tropical cyclone atmospheric waveguide prediction system based on an interpretable XGBoost model, characterized in that, The system is used to implement the tropical cyclone atmospheric waveguide prediction method based on the interpretable XGBoost model as described in any one of claims 1-5, including: The feature acquisition module is used to map meteorological parameters, tropical cyclone parameters, and location parameters to different features. Combine their respective tags Combined into a dataset , It is a natural number; The feature training module is used to take 70% of the dataset as the training set, use the grid search algorithm to adjust the hyperparameters of the XGBoost algorithm on the training set, and use the remaining 30% of the dataset as the test set to perform cross-validation training on the training set to obtain the optimal parameter combination. The feature importance ranking module is used to interpret the prediction results of the XGBoost algorithm using SHAP, and to analyze the relationship between the atmospheric waveguide of tropical cyclones and different features, and to rank the features by importance. The prediction module is used to predict the probability of atmospheric waveguide formation by using the XGBoost model with the optimal parameter combination and inputting the features to be measured according to the weights corresponding to the importance of the features.

7. An electronic device, characterized in that, The system includes a memory and a processor, wherein the processor is used to execute computer management programs stored in the memory to implement the steps of the tropical cyclone atmospheric waveguide prediction method based on the interpretable XGBoost model as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, It stores a computer management program, which, when executed by a processor, implements the steps of the tropical cyclone atmospheric waveguide prediction method based on an interpretable XGBoost model as described in any one of claims 1-5.

Images