Typhoon track simulation method coupled with machine learning and multi-source sea-air environmental factors

Abstract

The application discloses a typhoon path simulation method coupled with machine learning and multi-source sea-air environmental factors, and belongs to the field of typhoon disaster simulation. The application constructs a typhoon full-path machine learning prediction model fusing sea-air environmental factors. First, by integrating key sea-air parameters such as sea surface temperature, atmospheric vertical wind shear, relative humidity, absolute vorticity, vertical velocity and guiding airflow and typhoon path and intensity time series data, and combining a P value threshold method, significant factors affecting the typhoon path and intensity are selected. Then, based on a logistic regression model, a back propagation neural network, a recurrent neural network and a support vector machine machine learning model, and combining a hyperparameter tuning scheme, full-path four-stage optimal prediction models of the typhoon in the generation stage, the moving stage, the intensity evolution stage and the decay stage are constructed. Compared with a traditional statistical model, the machine learning model of the application can better capture the complex interaction mechanism between the typhoon and the environmental field.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of typhoon disaster simulation, and particularly relates to a typhoon path simulation method coupled with machine learning and multi-source sea-air environmental factors. BACKGROUND

[0002] Tropical cyclone is a strong disaster-causing weather system occurring on the tropical ocean, which is commonly known as typhoon in the Northwest Pacific. The Northwest Pacific is the sea area with the most frequent tropical cyclone activities in the world.

[0003] At present, the establishment of the generating model, moving model, intensity model and dissipation model in the typhoon empirical path model mostly adopts statistical methods or statistical dynamics methods, but the statistical methods are difficult to accurately depict the nonlinear variation law of typhoon generation, movement and intensity, and the statistical methods need a large number of typhoon samples to make more accurate estimates. Machine learning is a data-driven method that uses algorithms to learn rules from data and make predictions or decisions. Machine learning has strong generalization ability and adaptability, can effectively deal with complex nonlinear problems, and is suitable for typhoon path and intensity prediction. In recent years, many researchers have used different machine learning methods to improve the prediction accuracy of typhoon path and intensity from virtual and real-time aspects based on massive data. However, the movement path and intensity change of typhoon are not only controlled by atmospheric circulation field, but also have significant interaction relationship with the underlying sea-air environment. A large number of observation facts have shown that the sea-air environmental parameters have a key modulating effect on the evolution of typhoon, so it is necessary and feasible to include the sea-air environmental parameters into the typhoon full path prediction model. SUMMARY

[0004] To address the aforementioned issues, this invention proposes a typhoon path simulation method coupled with multi-source ocean-atmosphere environmental factors using machine learning. A machine learning prediction model for the entire typhoon path, integrating these factors, is constructed. First, by integrating key ocean-atmosphere environmental parameters such as sea surface temperature, vertical wind shear, relative humidity, absolute vorticity, vertical velocity, and steering flow with time-series data on typhoon path and intensity, significant factors influencing typhoon path and intensity are selected using the P-value thresholding method. Then, based on machine learning models including Logistic Regression, Back Propagation Neural Network (BPNN), Recurrent Neural Network (RNN), and Support Vector Machine (SVM), and combined with hyperparameter tuning schemes, optimal four-stage prediction models for the entire typhoon path are constructed, representing the formation, movement, intensity evolution, and attenuation stages. Finally, based on the optimal typhoon path prediction model, a 1000-year virtual typhoon event set in the Northwest Pacific is constructed. The results show good agreement with observed typhoon paths, further validating the effectiveness and accuracy of the proposed typhoon path prediction model.

[0005] The technical solution of the present invention is as follows:

[0006] A typhoon path simulation method coupled with multi-source air-sea environmental factors using machine learning is proposed. This method introduces air-sea environmental factors to construct a machine learning-based typhoon full-path prediction model, which includes four sub-models: a typhoon generation model, a typhoon movement speed prediction model, a typhoon intensity prediction model, and a land attenuation model. During model construction, a p-value thresholding method is used to select significant factors influencing typhoon path and intensity. The typhoon full-path prediction model is used to simulate and generate complete typhoon paths.

[0007] Furthermore, the construction process of the typhoon generation model is as follows:

[0008] Using the probability of tropical cyclone formation as the dependent variable and atmospheric absolute vorticity, relative humidity, relative ocean surface temperature, vertical velocity, and vertical wind shear as independent variables, a logistic regression model is established, as shown in the following formula:

[0009] ;

[0010] ;

[0011] in, The probability of a tropical cyclone forming point occurring; It is a function that is linearly correlated with large-scale environmental factors; This is the initial point generated; , , , , , All represent regression coefficients; The absolute vorticity of the 850 hPa pressure layer; This refers to the relative ocean surface temperature. The vertical velocity of the 500 hPa pressure layer; The relative humidity of a 600 hPa pressure layer; The vertical wind shear between the 200 hPa and 850 hPa pressure layers;

[0012] The typhoon generation model is used to construct any number of virtual typhoon formation points. The specific process is as follows: First, the entire Northwest Pacific Ocean is divided into a 1°×1° grid. Based on the historical typhoon data of each grid for each month, the coefficients of the logistic regression model are fitted to obtain 12 different sets of model coefficients. Then, a series of initial formation points are randomly generated and assigned a uniform random number between 0 and the maximum probability of occurrence. Topographic data and ocean-atmosphere environmental data are interpolated at the randomly generated initial formation points. The typhoon formation probability of the point is calculated according to the typhoon generation model. If the formation probability is greater than the assigned uniform random number, the randomly generated initial formation point is taken as the final formation point; otherwise, the randomly generated point is discarded.

[0013] Furthermore, the typhoon movement speed prediction model includes a typhoon meridional movement speed prediction model and a typhoon zonal movement speed prediction model, which are used to simulate the dynamic changes in the meridional and zonal movement speeds of the typhoon, respectively; the construction process is as follows:

[0014] The meridional and zonal velocity prediction models take the current meridional and zonal velocities as dependent variables, respectively, and select the meridional and zonal velocities of the typhoon, the longitude and latitude of the typhoon center, the steering flow, sea surface temperature, absolute vorticity, relative humidity, vertical wind shear, and vertical velocity air-sea environmental factors from the previous two time points as independent variables. The statistical model is expressed as follows:

[0015] ;

[0016] ;

[0017] in, to All represent regression coefficients; Longitude of the typhoon center; The latitude of the typhoon center; This represents the meridional speed of the typhoon. This represents the zonal movement speed of the typhoon. To guide the meridional velocity of the airflow; To guide the zonal velocity of the airflow; , , Indicates different times, and The interval between them is 6 hours; Sea surface temperature; The absolute vorticity of the 850 hPa pressure layer; The relative humidity of a 600 hPa pressure layer; The vertical velocity of the 500 hPa pressure layer; The vertical wind shear between the 200 hPa and 850 hPa pressure layers;

[0018] The meridional and zonal velocities of the guiding airflow are shown in the following formulas:

[0019] ;

[0020] ;

[0021] in, , , , , The average meridional wind speed in the atmospheric environment over the 5° annulus of each pressure layer; , , , , The average atmospheric zonal wind speed over the 5° annulus of each pressure layer;

[0022] The significance of the input parameters of the typhoon movement speed prediction model was analyzed using the P-value threshold method. Specifically, the study area of ​​the Northwest Pacific was divided into 5°×5° grids. Based on the fitted models of typhoon meridional and zonal movement speeds established for each grid, significant factors of typhoon movement speed within each grid were selected, with a P-value of 0.05 as the boundary. The frequency of each significant factor across all grids was then counted. Significant factors with a frequency exceeding 20 across all grids were selected as the final input factors for the meridional and zonal movement speed prediction models. The final typhoon meridional and zonal movement speed prediction formulas within each grid were simplified as follows:

[0023] ;

[0024] ;

[0025] Three different machine learning models—Backpropagation Neural Network (BPNN), Recurrent Neural Network (RNN), and Support Vector Machine (SVM)—were used to model the meridional and zonal movement speed of typhoons. Sensitivity experiments were then conducted to optimize the parameters, and the optimal machine learning model and its corresponding optimal parameters were selected for each grid. The parameters that needed to be optimized for BPNN and RNN included the number of nodes in the hidden layer, the transfer function, and the training function; the parameter that needed to be optimized for SVM was the kernel function.

[0026] The sensitivity experiment specifically uses the coefficient of determination and root mean square error as evaluation indicators; the coefficient of determination and root mean square error between the prediction result of each grid and the target result of each machine learning model are calculated; the machine learning model with the largest coefficient of determination and the smallest root mean square error is selected as the optimal machine learning prediction model for the current grid.

[0027] Furthermore, an ocean-atmosphere environmental factor is introduced to construct a typhoon intensity prediction model, which is used to simulate the relationship between the typhoon center pressure and significant ocean-atmosphere environmental factors, and to predict changes in the typhoon center pressure. The specific construction process of the typhoon intensity prediction model is as follows:

[0028] The typhoon intensity model takes the current central pressure of the typhoon as the dependent variable, and selects the central pressure of the typhoon at the previous two moments, the longitude and latitude of the typhoon center, sea surface temperature, absolute vorticity, relative humidity, vertical wind shear, and atmospheric vertical velocity as air-sea environmental factors as independent variables. The statistical model is expressed by the following formula:

[0029] ;

[0030] in, to All represent regression coefficients; Longitude of the typhoon center; The latitude of the typhoon center; The central pressure of the typhoon; , , Indicates different times, and The interval between them is 6 hours; Sea surface temperature; The absolute vorticity of the 850 hPa pressure layer; The relative humidity of a 600 hPa pressure layer; The vertical velocity of the 500 hPa pressure layer; The vertical wind shear between the 200 hPa and 850 hPa pressure layers;

[0031] The significance of the input parameters of the typhoon intensity model was analyzed using the p-value threshold method. Specifically, the study area of ​​the Northwest Pacific was divided into 5°×5° grids. Based on the typhoon intensity prediction model established for each grid, the frequency of significant factors for typhoon intensity in each 5°×5° grid was counted, with a p-value of 0.05 as the cutoff. Significant factors with a frequency exceeding 20 in all grids were selected as the final input factors for the typhoon intensity model. The final typhoon intensity prediction formula for each grid was simplified to the following formula:

[0032] ;

[0033] Three different machine learning models, BPNN, RNN, and SVM, were used to model the typhoon intensity prediction model. Each grid underwent sensitivity experiments and parameter tuning. Based on the criteria of maximizing the coefficient of determination and minimizing the root mean square error, the optimal machine learning prediction model for typhoon intensity in each 5°×5° grid in the Northwest Pacific was finally determined.

[0034] Furthermore, an ocean-atmosphere environmental factor was introduced to construct a land attenuation model to simulate the changes in central air pressure during the land attenuation phase of a typhoon; the specific construction process of the typhoon attenuation model is as follows:

[0035] The typhoon attenuation model takes the current typhoon center pressure as the dependent variable and selects the center pressure, longitude, latitude, absolute atmospheric vorticity, relative humidity, vertical wind shear, and vertical velocity air-sea environmental factors from the previous two time points as independent variables. The statistical model is expressed by the following formula:

[0036] ;

[0037] in, to All represent regression coefficients; Longitude of the typhoon center; The latitude of the typhoon center; The central pressure of the typhoon; , , Indicates different times, and The interval between them is 6 hours; The absolute vorticity of the 850 hPa pressure layer; The relative humidity of a 600 hPa pressure layer; The vertical velocity of the 500 hPa pressure layer; The vertical wind shear between the 200 hPa and 850 hPa pressure layers;

[0038] The significance of the input parameters of the land attenuation model was analyzed using the p-value threshold method. With a p-value of 0.05 as the cutoff, the significant factors in the typhoon land attenuation model were determined to be: , , , Therefore, the typhoon land attenuation model simplifies to the following formula:

[0039] ;

[0040] Three different machine learning models, BPNN, RNN, and SVM, were used to model the typhoon attenuation model. Through sensitivity experiments and parameter tuning, RNN was finally determined as the optimal prediction model for the land attenuation model.

[0041] The beneficial technical effects of this invention are as follows: This invention innovatively uses a machine learning model to replace the statistical regression model in traditional typhoon empirical path models. Furthermore, the model construction process considers the influence of marine and atmospheric environmental factors, including key marine and atmospheric environmental parameters such as sea surface temperature, vertical wind shear, relative humidity, absolute vorticity, vertical velocity, and steering flow. Simultaneously, the study focuses on meridional and zonal movement velocities in the typhoon movement model, taking into account the role of steering flow during typhoon movement. This aims to more accurately depict the entire typhoon path evolution process and improve the accuracy of typhoon path and intensity forecasts. Compared to traditional statistical models, the machine learning model of this invention can better capture the complex interaction mechanisms between typhoons and the environmental field, with the average coefficient of determination of the prediction results being superior to statistical results. The achievements of this invention can provide more reliable technical support for typhoon early warning and disaster prevention and mitigation, and also provide new research ideas for a deeper understanding of the typhoon-environment interaction mechanism. Attached Figure Description

[0042] Figure 1 This is a flowchart of the typhoon path simulation method of the present invention, which uses machine learning coupled with multiple source air-sea environmental factors.

[0043] Figure 2 This is a frequency distribution map of the significant influence factors on the meridional movement speed of typhoons in the Northwest Pacific 5°×5° grid in this invention.

[0044] Figure 3 This is a frequency distribution map of the significant influencing factors of the zonal movement speed of typhoons in the Northwest Pacific 5°×5° grid in this invention.

[0045] Figure 4 This is a frequency distribution diagram of the influence factors in the 5°×5° grid typhoon intensity model of the Northwest Pacific Ocean in this invention. Detailed Implementation

[0046] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0047] like Figure 1 As shown, this invention first constructs a typhoon full-path prediction model based on machine learning. This typhoon full-path prediction model consists of four core sub-models: a typhoon generation model, a typhoon movement speed prediction model, a typhoon intensity prediction model, and a land attenuation model. Then, it selects significant factors (hereinafter referred to as significant factors) that affect the typhoon path and intensity by combining the P-value threshold method. Through segmented modeling and integration of ocean-atmosphere environmental parameters and historical time series characteristics, it realizes the simulation of the entire life cycle of a typhoon from its formation to its dissipation.

[0048] The typhoon formation model uses a logistic regression model to establish the relationship between ocean-atmosphere environmental parameters and the probability of typhoon formation at the initial formation point on a monthly basis. The typhoon movement speed model uses three machine learning models—BPNN, RNN, and SVM—to simulate the dynamic changes in the meridional and zonal movement speed of typhoons, capturing the nonlinear relationship between the path and significant ocean-atmosphere environmental factors. The typhoon intensity model uses three machine learning models to simulate the relationship between typhoon central pressure and significant ocean-atmosphere environmental factors, predicting typhoon central pressure. The land attenuation model uses three machine learning models to handle the non-stationary temporal characteristics of typhoon central pressure during the land attenuation phase. Each sub-model achieved good predictive results through performance optimization. Integrating the four sub-models ultimately generated a 1000-year virtual typhoon event set for the Northwest Pacific. The following details the construction and result verification process of each sub-model.

[0049] The data used in this invention mainly comes from the China Meteorological Administration's optimal tropical cyclone track dataset (which includes typhoon track data) and the European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis 5 (ERA5) dataset (ERA5 is the full name of the ECMWF fifth generation global climate reanalysis dataset, which includes ocean-atmosphere environmental data). The optimal tropical cyclone track dataset comes from the China Meteorological Administration's Tropical Cyclone Data Center. This dataset systematically records the 6-hour location and intensity information of tropical cyclones in the Northwest Pacific. Specifically, this dataset contains typhoon track data from 1949 to 2023, and key parameters include typhoon timestamps, intensity indices, latitude, longitude, minimum central pressure, 2-minute average maximum sustained wind speed near the center, and 2-minute average wind speed. The ERA5 dataset system records relevant data on the global ocean and atmosphere environment, specifically including ocean and atmosphere environment data from 1940 to 2025, including data types, variable information, pressure layers, time information, spatial information, data format, etc. The key parameters required by this invention include atmospheric relative vorticity, atmospheric relative humidity, atmospheric vertical velocity, wind speed U component, wind speed V component, and ocean temperature.

[0050] To improve the prediction accuracy of the typhoon full-path model, this invention performs systematic preprocessing on the raw data. The data is standardized using the max-min normalization method, and missing values ​​are imputed using the average value of the same feature dimension.

[0051] (a) Typhoon formation model;

[0052] The typhoon formation model primarily establishes a monthly probabilistic model of typhoon formation points based on various air-sea environmental factors within a 1°×1° grid over the Northwest Pacific Ocean (100°E-180°E, 5°N-45°N). The environmental factors required for the monthly probabilistic model of tropical cyclone formation points in the Northwest Pacific are interpolated to the grid center of each tropical cyclone. Using the probability of tropical cyclone formation as the dependent variable and atmospheric absolute vorticity, relative humidity, relative sea surface temperature, vertical velocity, and vertical wind shear as independent variables, a logistic regression model is established, as shown in the following formula:

[0053] (1);

[0054] (2);

[0055] in, The probability of a tropical cyclone forming point occurring; It is a function that is linearly correlated with large-scale environmental factors; This is the initial point generated; , , , , , All represent regression coefficients; The absolute vorticity of the 850 hPa pressure layer (10 -5 s -1 ); The relative ocean surface temperature (K); The vertical velocity (m / s) of the 500 hPa pressure layer; The relative humidity (%) of the 600 hPa pressure layer; The vertical wind shear (m / s) between the 200 hPa and 850 hPa pressure layers;

[0056] It is worth noting that in this invention, the initial formation location of a tropical cyclone is defined as the location where the maximum sustained wind speed of 15 m / s is first reached. The relative ocean surface temperature is the ocean surface temperature at each 1°×1° grid point minus the average ocean surface temperature of the 20°S-20°N region. This invention uses relative ocean surface temperature because numerous studies have shown that rising ocean temperatures do not necessarily lead to an increase in the number and intensity of typhoons, as rising sea temperatures lead to an increase in the moisture entropy of the mid-atmosphere, thereby inhibiting typhoon formation.

[0057] The specific process of constructing typhoon formation points based on the typhoon formation model is as follows: First, the entire Northwest Pacific Ocean is divided into a 1°×1° grid. Based on the monthly historical typhoon data for each grid, the coefficients of a logistic regression model are fitted, resulting in 12 different sets of model coefficients. Then, a series of initial formation points are randomly generated, each assigned a uniformly random number between 0 and the maximum probability of occurrence. Topographic and atmospheric environmental data are interpolated at these randomly generated initial formation points. The typhoon formation probability at each point is calculated using the typhoon formation model. If the probability of formation is greater than the assigned uniformly random number, the randomly generated initial point is selected as the final formation point; otherwise, the randomly generated point is discarded. The typhoon formation model based on this invention can accurately simulate the distribution of formation points.

[0058] The typhoon formation model in this invention provides a relatively accurate simulation of the formation point in the Northwest Pacific Ocean, and can be used to predict typhoon movement and intensity models.

[0059] (ii) Typhoon movement speed prediction model;

[0060] This invention, based on empirical path models, incorporates steering airflow into typhoon movement speed prediction models, establishing a prediction model for typhoon meridional and zonal movement speeds based on air-sea environmental factors and steering airflow. Significance factors of the model are determined using the P-value threshold method, and the optimal machine learning prediction model for meridional and zonal movement speeds and its corresponding hyperparameters are determined through sensitivity experiments and hyperparameter tuning.

[0061] The meridional and zonal speed prediction models take the current meridional and zonal speeds as dependent variables, respectively, and select atmospheric and marine environmental factors such as the meridional and zonal speeds of the typhoon at the previous two moments, the longitude and latitude of the typhoon center, the steering flow, sea surface temperature, absolute vorticity, relative humidity, vertical wind shear, and vertical velocity as independent variables. The statistical model can be expressed as follows:

[0062] (3);

[0063] (4);

[0064] in, to All represent regression coefficients; Longitude (°) of the typhoon center; The latitude (°) of the typhoon center; The meridional speed of the typhoon (m / s); The zonal movement speed of the typhoon is (m / s); The meridional velocity (m / s) used to guide the airflow; The zonal velocity (m / s) used to guide the airflow; , , Indicates different times, and The interval between them is 6 hours; Sea surface temperature (K); The absolute vorticity of the 850 hPa pressure layer (10 -5 s -1 ); The relative humidity (%) of the 600 hPa pressure layer; The vertical velocity (m / s) of the 500 hPa pressure layer; The vertical wind shear (m / s) between the 200 hPa and 850 hPa pressure layers.

[0065] The steering flow is generally defined as the average wind speed of the pressure layer in the annulus at a certain distance from the center of a tropical cyclone. This invention selects the average wind speed of the pressure layer (300-850 hPa) in the annulus at 5° distance from the center of the tropical cyclone as the steering flow, because the overall average path simulated using the average wind in the 5° annulus best matches the historical average path. The specific meridional and zonal velocities of the steering flow are shown in the following formulas:

[0066] (5); (6);

[0067] in, , , , , The mean atmospheric meridional wind speed (m / s) is located in the 5° annulus of each pressure layer. , , , , The mean atmospheric zonal wind speed (m / s) is given in the 5° annular zone of each pressure layer.

[0068] To accurately depict the changes in typhoon movement speed, this invention re-divides the study area of ​​the Northwest Pacific into 5°×5° grids. Based on the typhoon movement speed data and air-sea environmental factor data for each grid, an optimal machine learning prediction model for the typhoon movement speed of that grid is established. First, based on historical typhoon data from the China Meteorological Administration, sample data of the meridional and zonal movement speeds of typhoons within each grid are statistically analyzed. Based on the statistical results, contour maps of the sample size for the typhoon movement path prediction model within each grid are drawn. Among them, the grid areas with relatively large sample sizes are mainly concentrated in 110°E-135°E and 10°N-25°N, with sample sizes reaching approximately 800-1000.

[0069] Since formulas (3) and (4) consider too many factors affecting typhoon movement speed, the complexity of the machine learning model will inevitably increase. Moreover, there may be multicollinearity among these factors, which will affect the generalization ability of the model. Therefore, the significance of the input factors of typhoon meridional and zonal movement speeds was analyzed by using the P-value threshold method. Based on the fitting models of typhoon meridional and zonal movement speeds established for each grid, the significant factors affecting typhoon movement speeds in each grid (hereinafter referred to as significant factors) can be selected with a P-value of 0.05 as the boundary. However, since the significant factors affecting each grid are not the same, and in order to reduce the complexity of the model, it is necessary to select common significant factors for all grids. Therefore, the frequency of each significant factor appearing in all grids was further counted. The factor with the higher frequency was used as the common significant factor for the typhoon movement speed prediction model in all grids. Thus, the frequency distribution map of the significant factors affecting typhoon meridional and zonal movement speeds in the 5°×5° grid of the Northwest Pacific was obtained, as shown in the figure. Figure 2 and Figure 3 As shown.

[0070] In this invention, significant influence factors with a frequency exceeding 20 in all grids are selected as the final input factors for the longitudinal and latitudinal movement speed prediction models. Figure 2 It can be seen that in the typhoon meridional speed prediction model, the input factor is... , , , , , In the zonal movement speed prediction model for typhoons, the input factor is... , , , , , , Finally, the formulas for predicting the meridional and zonal movement speed of typhoons within each grid can be simplified to the following formulas:

[0071] (7);

[0072] (8);

[0073] Subsequently, this invention models the typhoon meridional and zonal movement speed prediction models using three different machine learning models: BPNN, RNN, and SVM. Many parameters need to be determined in the machine learning models, such as the number of nodes in the hidden layer, the transfer function, and the training function, as shown in Table 1. For the BPNN / RNN model, the transfer functions of its hidden and output layers are composed of a combination of two activation functions, including the hyperbolic tangent function (MATLAB: tansig), the logistic sigmoid function (MATLAB: logsig), and the linear function (MATLAB: purelin). The training functions include the Levenberg-Marquardt algorithm (MATLAB: trainlm), the elastic backpropagation algorithm (MATLAB: trainrp), the scaled conjugate gradient method (MATLAB: trainscg), the gradient descent method with a driving term and adaptive learning rate (MATLAB: traindx), the one-step secant method (MATLAB: trainoss), and the Bayesian regularization algorithm (MATLAB: trainbr). For the SVM model, its kernel function is selected from the linear kernel (MATLAB: linear), the polynomial kernel (MATLAB: polynomial), and the Gaussian kernel (MATLAB: rbf). This invention optimizes parameters through sensitivity experiments, ultimately selecting the optimal machine learning model and corresponding optimal model parameters for each grid.

[0074] Table 1 Adjustable parameters of machine learning models

[0075]

[0076] The core of Backpropagation Neural Networks (BPNNs) lies in error backpropagation and gradient descent optimization, enabling multi-layer neural networks to effectively learn complex nonlinear mappings. It is a typical multi-layer feedforward neural network trained using the error backpropagation algorithm, consisting of an input layer, hidden layers, and an output layer. The complete process of BPNN operation is as follows: neurons in the input layer receive input signals from the external environment and pass them to one or more hidden layer neurons in the intermediate layers. These neurons process and transform the input signals, then pass the results to the output layer. The output layer neurons output the processed results, compare them with the actual expected output, and calculate the error. If the actual output value does not match the predicted output value, error backpropagation occurs. The output error is used to adjust the weights of each layer using error gradient descent, and the error is backpropagated to the hidden and input layers. This process continues until the operator is satisfied with the output error or a certain number of learning cycles have been completed. Therefore, BPNNs have a powerful nonlinear fitting capability. The gradient descent algorithm they employ calculates the gradient of the loss function with respect to the parameters, updates the parameters along the opposite direction of the gradient, and gradually approaches the minimum value of the loss function.

[0077] A Recurrent Neural Network (RNN) is a type of neural network specifically designed for processing sequential data. Its core feature is the use of recurrent connections to allow information to pass across time steps, thus endowing the network with memory capabilities. The hidden state at each time step is determined by both the current input and the previous state. An RNN consists of an input layer, recurrent hidden layers, and an output layer. The hidden layers simultaneously receive both the current input and historical states, forming a temporal feedback loop. This structure enables RNNs to model the dynamic features of sequences, making them suitable for time-series data prediction and generation tasks.

[0078] Support Vector Machines (SVMs) are supervised learning algorithms based on statistical learning theory, primarily used for classification and regression tasks. Their core idea is to find the optimal hyperplane that maximizes the margin between different classes of data, thereby improving the model's generalization ability.

[0079] To evaluate the prediction accuracy of the model under different parameter conditions, this invention employs the coefficient of determination (R²) between the predicted result and the target result. 2 The root mean square error (RMSE) and root mean square error are used as evaluation metrics. The calculation formulas are as follows:

[0080] (9);

[0081] (10);

[0082] in, The coefficient of determination; This is the root mean square error; It is the sum of squared residuals; The total sum of squares; Indicates the first The true value result of each observed object; Indicates the first Model prediction results for each observed object; It is the first The average of the actual observation results for each observation object; This represents the total number of observed objects.

[0083] The coefficient of determination is used to evaluate the goodness of fit of a model to the actual data, reflecting the proportion of variability in the actual results that can be explained by the predicted results. 2 The closer the RMSE is to 1, the better the model fits the true data, and the greater the proportion of variance the model can explain. RMSE represents the overall magnitude of the difference between the predicted and actual results. The smaller the RMSE, the higher the average prediction accuracy of the model.

[0084] After sensitivity experiments and parameter tuning, based on R 2 The criteria of maximizing and minimizing RMSE were used to determine the optimal machine learning prediction models for meridional and zonal typhoon movement speeds within each 5°×5° grid in the Northwest Pacific. For the optimal machine learning prediction models for meridional typhoon movement speed in each grid, RNN had the highest grid share at 46.6%, followed by BPNN at 28.1%, and finally SVM at 24.2%. For the optimal machine learning prediction models for zonal typhoon movement speed in each grid unit, RNN had the highest grid share at 44.6%, followed by SVM at 29.1%, and finally BPNN at 26.2%.

[0085] For the parameters of the optimal machine learning model for the meridional and zonal movement speed of the typhoon in each grid cell, the most common parameters are purelin-purelin for the hidden and output layer transfer functions, trainlm for the training function, and linear for the kernel function in SVM.

[0086] To verify the accuracy of the machine learning model's predictions, the coefficient of determination and root mean square error (RMSE) between the optimal machine learning model's predictions of typhoon meridional and zonal movement speeds within a 5°×5° grid in the Northwest Pacific and the actual typhoon movement speeds were compared with the corresponding results from the traditional statistical regression model. The R-squared values ​​of the machine learning model and the statistical model were calculated separately. 2The difference between the machine learning model and the RMSE is shown in the figure. For the vast majority of grid cells, the coefficients of determination between the machine learning model and the actual values ​​of the typhoon's meridional and zonal movement speeds are significantly higher than those of the statistical regression model. For the vast majority of grid cells, the RMSE between the machine learning model and the actual values ​​of the typhoon's meridional and zonal movement speeds is significantly lower than those of the statistical regression model. This indicates that the machine learning model's prediction results for the typhoon's meridional and zonal movement speeds are better than those of the statistical regression model.

[0087] (III) Typhoon Intensity Prediction Model;

[0088] Similar to typhoon movement models, this invention introduces ocean-atmosphere environmental factors to establish a typhoon intensity prediction model based on ocean-atmosphere environmental parameters. First, the significant factors of the typhoon intensity prediction model are determined using the P-value threshold method. Then, through sensitivity experiments and parameter tuning, the optimal machine learning model for typhoon intensity prediction and the corresponding optimal parameters are determined.

[0089] The typhoon intensity model takes the current central pressure of the typhoon as the dependent variable, and selects the central pressure of the typhoon at the previous two moments, the longitude and latitude of the typhoon center, sea surface temperature, absolute vorticity, relative humidity, vertical wind shear, and atmospheric vertical velocity, etc., as independent variables. The statistical model can be expressed as follows:

[0090] (11);

[0091] in, to All represent regression coefficients; Longitude (°) of the typhoon center; The latitude (°) of the typhoon center; The central pressure of the typhoon (hPa); , , Indicates different times, and The interval between them is 6 hours; Sea surface temperature (K); The absolute vorticity of the 850 hPa pressure layer (10 -5 s -1 ); The relative humidity (%) of the 600 hPa pressure layer; The vertical velocity (m / s) of the 500 hPa pressure layer; The vertical wind shear (m / s) between the 200 hPa and 850 hPa pressure layers.

[0092] This invention continues to use the 5°×5° grid in the Northwest Pacific region for typhoon movement speed prediction models. Based on the typhoon intensity data and air-sea environmental factor data for each grid, an optimal machine learning prediction model for the typhoon intensity of that grid is established. The significance of the input factors for typhoon intensity is analyzed using the P-value threshold method. Based on the typhoon intensity prediction model established for each grid, with a P-value of 0.05 as the cutoff, the frequency of significant influencing factors of typhoon intensity in the 5°×5° grid is statistically analyzed, resulting in a frequency distribution map of significant influencing factors for the 5°×5° grid typhoon intensity model in the Northwest Pacific region, as shown below. Figure 4 As shown.

[0093] In this invention, significant influence factors with a frequency exceeding 20 in all grids are selected as the final input factors for the typhoon intensity model. Figure 4 It can be seen that the significant input factor is , , , , Finally, the typhoon intensity prediction formula within each grid can be simplified to the following formula:

[0094] (12);

[0095] Subsequently, this invention uses three different machine learning models—BPNN, RNN, and SVM—to model the typhoon intensity prediction model. Each grid undergoes sensitivity experiments and parameter tuning (Table 1), based on R... 2 The criteria of maximum and minimum RMSE were used to determine the optimal machine learning prediction model for typhoon intensity within each 5°×5° grid in the Northwest Pacific.

[0096] For the machine learning prediction model that optimizes typhoon intensity for each grid, RNN has the highest grid ratio at 50.4%, followed by BPNN at 33.9%, and finally SVM at 15.5%.

[0097] For the parameters of the optimal machine learning model for typhoon intensity in each grid cell, the most common parameters are purelin-purelin transfer function for hidden and output layers, trainlm for training function, and linear kernel function for SVM.

[0098] To verify the accuracy of the machine learning model's predictions, the coefficient of determination and root mean square error (RMSE) between the predictions of the optimal machine learning model for typhoon intensity within a 5°×5° grid in the Northwest Pacific and the actual typhoon intensity were compared with the corresponding results of the traditional statistical regression model. The R-squared values ​​of the machine learning model and the statistical model were calculated separately. 2The difference between the predicted typhoon intensity and the actual RMSE, where, for most grid cells, the R-value between the predicted typhoon intensity and the actual RMSE is... 2 Both the RMSE and RMSE values ​​are significantly better than those of the regression model. This indicates that the machine learning prediction model outperforms the statistical regression model in predicting typhoon intensity.

[0099] (iv) Land attenuation model;

[0100] After making landfall, typhoons rapidly weaken due to the influence of the underlying surface. To fully characterize the evolution of a typhoon's entire life cycle, a model for its intensity decay after landfall is needed. Similar to the typhoon movement and intensity model, an ocean-atmosphere environmental factor-based typhoon landfall attenuation model is established. It's important to note that due to the limited sample data of typhoons making landfall, the attenuation model is not divided into grids; instead, all typhoons that have made landfall are used as samples to establish a unified typhoon attenuation model for the Northwest Pacific land region.

[0101] The typhoon attenuation model takes the current typhoon center pressure as the dependent variable and selects the center pressure at the previous two moments, the longitude and latitude of the typhoon center, atmospheric absolute vorticity, relative humidity, vertical wind shear, and vertical velocity, among other air-sea environmental factors, as independent variables. The statistical model can be expressed as follows:

[0102] (13);

[0103] in, to All represent regression coefficients; Longitude (°) of the typhoon center; The latitude (°) of the typhoon center; The central pressure of the typhoon (hPa); , , Indicates different times, and The interval between them is 6 hours; The absolute vorticity of the 850 hPa pressure layer (10 -5 s -1 ); The relative humidity (%) of the 600 hPa pressure layer; The vertical velocity (m / s) of the 500 hPa pressure layer; The vertical wind shear (m / s) between the 200 hPa and 850 hPa pressure layers.

[0104] First, the significance of the input parameters of the land attenuation model was analyzed using the p-value threshold method. With a p-value of 0.05 as the cutoff, the significant influencing factors in the typhoon land attenuation model were determined to be: , , , Therefore, the typhoon land attenuation model can be simplified to the following formula:

[0105] (14);

[0106] Then, this invention models the typhoon attenuation model using three different machine learning models: BPNN, RNN, and SVM. Through sensitivity experiments and parameter tuning (Table 1), the optimal prediction model for the land attenuation model was finally determined to be RNN. In this model, the transfer functions of both the hidden and output layers are tansig, the training function is trainbr, and the number of hidden layers is 20. Similarly, to verify the accuracy of the machine learning model's prediction results, the R² value between the typhoon intensity predicted by the RNN typhoon attenuation model and the actual typhoon intensity was calculated. 2 The R² values ​​for RMSE were 0.78 and 0.57, respectively, while the corresponding values ​​predicted by the statistical regression model were 0.53 and 2.37. This shows that the R² values ​​predicted by the machine learning-based typhoon attenuation model are significantly lower than those predicted by the RMSE. 2 The results were significantly higher than those of statistical regression, while the RMSE was significantly lower, demonstrating the effectiveness of machine learning in predicting the intensity decay of typhoons over land.

[0107] The above process completes the modeling of the entire life cycle of a typhoon, from its formation to its movement, intensity evolution, and decay. Based on the above four sub-models, any number of virtual typhoons can be generated, providing reliable data support for typhoon risk analysis.

[0108] The process of constructing virtual typhoons begins by determining the number of virtual tropical cyclones for each month over a 40-year period (1979-2018) using the average number of tropical cyclones generated each month. Based on a typhoon formation model, the starting points of virtual typhoons in the Northwest Pacific over the past 1000 years are generated. The initial intensity of the virtual typhoon is then randomly sampled from the pressure of all historical typhoon starting points within a 1°×1° grid, thus initializing the virtual typhoon's position and intensity. Next, the Northwest Pacific is divided into 5°×5° grids. Based on historical typhoon data for each generated point's grid, machine learning prediction models are built to predict the typhoon's meridional and zonal movement speeds and intensity, forecasting its position and intensity at the next moment. After landfall, a land attenuation model is used to predict its intensity. Finally, by repeating these steps, the complete typhoon path is obtained.

[0109] This invention constructs a 1000-year virtual typhoon dataset in the Northwest Pacific based on a machine learning model, containing 30,074 virtual typhoon events. The distribution trends of the virtual typhoon paths constructed in this invention highly match those of the real paths, and the pressure distribution is also quite similar to the measured typhoon paths. Therefore, the virtual typhoons constructed based on the machine learning model in this invention have high reliability.

[0110] In addition, the simulation of a virtual typhoon in a single month based on a machine learning model in this invention also yields good results.

[0111] In summary, the statistical characteristics of the virtual typhoons constructed by the machine learning model are highly consistent with those of the observed typhoons, indicating that the virtual typhoons constructed by the machine learning model can reproduce the statistical characteristics of typhoons along the Chinese coast.

[0112] Machine learning algorithms possess good adaptability, self-learning ability, and nonlinear mapping capability, making them more suitable for handling complex nonlinear problems. This invention innovatively replaces the statistical regression model in traditional typhoon empirical path models with a machine learning model, and considers the influence of atmospheric-sea environmental factors during model building, including key atmospheric-sea parameters such as sea surface temperature, vertical wind shear, humidity, vorticity, vertical velocity, and steering flow. First, a typhoon generation model is built based on a logistic regression model, which can construct any number of virtual typhoon starting points. Then, for the prediction models of typhoon meridional speed, zonal speed, and typhoon central pressure, the p-value threshold method is used to determine significant influencing factors. Sensitivity experiments and hyperparameter tuning are used to select the optimal machine learning model and its corresponding parameters from BPNN, RNN, and SVM models. Next, the machine learning prediction results are compared with those of traditional statistical regression models. The results show that the machine learning model of this invention can better capture the complex interaction mechanism between typhoons and the environmental field, and the average coefficient of determination of the prediction results is higher than that of statistical models. Finally, based on the optimal four-stage prediction model of the typhoon's full path in the formation, movement, intensity evolution and attenuation stages, a virtual typhoon event set for 1000 years in the Northwest Pacific was constructed. The results are in good agreement with the observed typhoon paths, further verifying the effectiveness and accuracy of the typhoon full path machine learning prediction model proposed in this invention.

[0113] This invention provides a new approach to typhoon stochastic simulation, representing a paradigm shift from empirical statistics to data-driven intelligence in typhoon stochastic simulation. This is significant for improving disaster prevention and control capabilities in the context of global climate change.

[0114] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A typhoon path simulation method using machine learning coupled with multi-source air-sea environmental factors, characterized in that, A typhoon full-path prediction model based on machine learning was constructed by introducing air-sea environmental factors. The model includes four sub-models: typhoon generation model, typhoon movement speed prediction model, typhoon intensity prediction model, and land attenuation model. During the model construction process, the p-value threshold method was used to select significant factors affecting typhoon path and intensity. The typhoon full-path prediction model is used to simulate and generate complete typhoon paths. The typhoon movement speed prediction model includes a typhoon meridional movement speed prediction model and a typhoon zonal movement speed prediction model, which are used to simulate the dynamic changes in the meridional and zonal movement speeds of typhoons, respectively; the construction process is as follows: The meridional and zonal velocity prediction models take the current meridional and zonal velocities as dependent variables, respectively, and select the meridional and zonal velocities of the typhoon, the longitude and latitude of the typhoon center, the steering flow, sea surface temperature, absolute vorticity, relative humidity, vertical wind shear, and vertical velocity air-sea environmental factors from the previous two time points as independent variables. The statistical model is expressed as follows: ； ； in, to All represent regression coefficients; Longitude of the typhoon center; The latitude of the typhoon center; This represents the meridional speed of the typhoon. This represents the zonal movement speed of the typhoon. To guide the meridional velocity of the airflow; To guide the zonal velocity of the airflow; , , Indicates different times, and The interval between them is 6 hours; Sea surface temperature; The absolute vorticity of the 850 hPa pressure layer; The relative humidity of a 600 hPa pressure layer; The vertical velocity of the 500 hPa pressure layer; The vertical wind shear between the 200 hPa and 850 hPa pressure layers; The meridional and zonal velocities of the guiding airflow are shown in the following formulas: ； ； in, , , , , The average meridional wind speed in the atmospheric environment over the 5° annulus of each pressure layer; , , , , The average atmospheric zonal wind speed over the 5° annulus of each pressure layer; The significance of the input parameters of the typhoon movement speed prediction model was analyzed using the P-value threshold method. Specifically, the study area of ​​the Northwest Pacific was divided into 5°×5° grids. Based on the fitted models of typhoon meridional and zonal movement speeds established for each grid, significant factors of typhoon movement speed within each grid were selected, with a P-value of 0.05 as the boundary. The frequency of each significant factor across all grids was then counted. Significant factors with a frequency exceeding 20 across all grids were selected as the final input factors for the meridional and zonal movement speed prediction models. The final typhoon meridional and zonal movement speed prediction formulas within each grid were simplified as follows: ； ； Three different machine learning models—Backpropagation Neural Network (BPNN), Recurrent Neural Network (RNN), and Support Vector Machine (SVM)—were used to model the meridional and zonal movement speed of typhoons. Sensitivity experiments were then conducted to optimize the parameters, and the optimal machine learning model and its corresponding optimal parameters were selected for each grid. The parameters that needed to be optimized for BPNN and RNN included the number of nodes in the hidden layer, the transfer function, and the training function; the parameter that needed to be optimized for SVM was the kernel function. The sensitivity experiment specifically uses the coefficient of determination and root mean square error as evaluation indicators; the coefficient of determination and root mean square error between the prediction result of each grid and the target result of each machine learning model are calculated; the machine learning model with the largest coefficient of determination and the smallest root mean square error is selected as the optimal machine learning prediction model for the current grid.

2. The typhoon path simulation method based on machine learning coupled with multi-source air-sea environmental factors according to claim 1, characterized in that, The process of constructing the typhoon generation model is as follows: Using the probability of tropical cyclone formation as the dependent variable and atmospheric absolute vorticity, relative humidity, relative ocean surface temperature, vertical velocity, and vertical wind shear as independent variables, a logistic regression model is established, as shown in the following formula: ； ； in, The probability of a tropical cyclone forming point; It is a function that is linearly correlated with large-scale environmental factors; This is the initial point generated; , , , , , All represent regression coefficients; The absolute vorticity of the 850 hPa pressure layer; This refers to the relative ocean surface temperature. The vertical velocity of the 500 hPa pressure layer; The relative humidity of a 600 hPa pressure layer; The vertical wind shear between the 200 hPa and 850 hPa pressure layers; The typhoon generation model is used to construct any number of virtual typhoon formation points. The specific process is as follows: First, the entire Northwest Pacific Ocean is divided into a 1°×1° grid. Based on the historical typhoon data of each grid for each month, the coefficients of the logistic regression model are fitted to obtain 12 different sets of model coefficients. Then, a series of initial formation points are randomly generated and assigned a uniform random number between 0 and the maximum probability of occurrence. Topographic data and ocean-atmosphere environmental data are interpolated at the randomly generated initial formation points. The typhoon formation probability of the point is calculated according to the typhoon generation model. If the formation probability is greater than the assigned uniform random number, the randomly generated initial formation point is taken as the final formation point; otherwise, the randomly generated point is discarded.

3. The typhoon path simulation method based on machine learning coupled with multi-source air-sea environmental factors according to claim 1, characterized in that, A typhoon intensity prediction model was constructed by introducing marine-atmospheric environmental factors to simulate the relationship between typhoon central pressure and significant marine-atmospheric environmental factors, and to predict changes in typhoon central pressure. The specific construction process of the typhoon intensity prediction model is as follows: The typhoon intensity model takes the current central pressure of the typhoon as the dependent variable, and selects the central pressure of the typhoon at the previous two moments, the longitude and latitude of the typhoon center, sea surface temperature, absolute vorticity, relative humidity, vertical wind shear, and atmospheric vertical velocity as air-sea environmental factors as independent variables. The statistical model is expressed by the following formula: ； in, to All represent regression coefficients; Longitude of the typhoon center; The latitude of the typhoon center; The central pressure of the typhoon; , , Indicates different times, and The interval between them is 6 hours; Sea surface temperature; The absolute vorticity of the 850 hPa pressure layer; The relative humidity of a 600 hPa pressure layer; The vertical velocity of the 500 hPa pressure layer; The vertical wind shear between the 200 hPa and 850 hPa pressure layers; The significance of the input parameters of the typhoon intensity model was analyzed using the p-value threshold method. Specifically, the study area of ​​the Northwest Pacific was divided into 5°×5° grids. Based on the typhoon intensity prediction model established for each grid, the frequency of significant factors for typhoon intensity in each 5°×5° grid was counted, with a p-value of 0.05 as the cutoff. Significant factors with a frequency exceeding 20 in all grids were selected as the final input factors for the typhoon intensity model. The final typhoon intensity prediction formula for each grid was simplified to the following formula: ； Three different machine learning models, BPNN, RNN, and SVM, were used to model the typhoon intensity prediction model. Each grid underwent sensitivity experiments and parameter tuning. Based on the criteria of maximizing the coefficient of determination and minimizing the root mean square error, the optimal machine learning prediction model for typhoon intensity in each 5°×5° grid in the Northwest Pacific was finally determined.

4. The typhoon path simulation method based on machine learning coupled with multi-source air-sea environmental factors according to claim 1, characterized in that, A land attenuation model was constructed by introducing air-sea environmental factors to simulate the changes in central air pressure of typhoons during their land attenuation phase. The specific construction process of the typhoon attenuation model is as follows: The typhoon attenuation model takes the current typhoon center pressure as the dependent variable and selects the center pressure, longitude, latitude, absolute atmospheric vorticity, relative humidity, vertical wind shear, and vertical velocity air-sea environmental factors from the previous two time points as independent variables. The statistical model is expressed by the following formula: ； in, to All represent regression coefficients; Longitude of the typhoon center; The latitude of the typhoon center; The central pressure of the typhoon; , , Indicates different times, and The interval between them is 6 hours; The absolute vorticity of the 850 hPa pressure layer; The relative humidity of a 600 hPa pressure layer; The vertical velocity of the 500 hPa pressure layer; The vertical wind shear between the 200 hPa and 850 hPa pressure layers; The significance of the input parameters of the land attenuation model was analyzed using the p-value threshold method. With a p-value of 0.05 as the cutoff, the significant factors in the typhoon land attenuation model were determined to be: , , , Therefore, the typhoon land attenuation model simplifies to the following formula: ； Three different machine learning models, BPNN, RNN, and SVM, were used to model the typhoon attenuation model. Through sensitivity experiments and parameter tuning, RNN was finally determined as the optimal prediction model for the land attenuation model.

Images