Tropical gas cyclone radius prediction system based on statistics-dynamics and machine learning fusion model

Abstract

The invention relates to a tropical cyclone radius prediction system based on a statistics-dynamics and machine learning fusion model, which is provided with a statistics-dynamic model and a machine learning model, and is characterized in that historical cyclone data and climate statistical data are input through the statistics-dynamic model, and prediction of a target cyclone radius is obtained in combination with a historical cyclone development trend. The meteorological data and the data output by the statistical-dynamic model are input through the machine learning model, the predicted cyclone radius is output by the trained machine learning model, the complex non-linear relation in the meteorological data is identified and captured, and higher precision is achieved when the radius of the wind ring is predicted; and finally, based on an integrated learning method, combining the output results of the machine learning model and the statistical-dynamic model to form a comprehensive prediction result, so that the limitation of model hypothesis on the prediction result can be reduced, and the estimation precision of the wind ring radius is further optimized.

Description

Technical Field

[0001] The present invention belongs to the technical field of machine learning, and in particular relates to a tropical cyclone wind radius prediction system based on a statistical-dynamic and machine learning fusion model. Background Art

[0002] Tropical cyclones are important weather phenomena worldwide, and their intensity and wind circle structure pose a significant threat to human life and property safety in coastal areas. Currently, weather forecasts mainly rely on numerical weather prediction models and statistical-dynamic models, which predict the path and intensity of cyclones by analyzing the atmospheric environment and historical data.

[0003] Existing statistical-dynamic models, such as the CLIPER model, are often limited by model assumptions and parameterization when dealing with nonlinear changes in the wind circle radius, resulting in insufficient prediction accuracy. The model is too dependent on climatological averages and persistence, and fails to fully capture the dynamic characteristics of tropical cyclones and the impact of complex environmental factors. At the same time, statistical-dynamic models are mainly constructed based on static data and cannot reflect the latest environmental changes and cyclone development trends in a timely manner. There is a lag in real-time data updates and it is impossible to fully utilize real-time data to improve the timeliness of predictions. When faced with rapidly changing cyclone conditions, the predicted wind circle radius has defects in uncertainty and asymmetry.

[0004] Therefore, designing a tropical cyclone wind radius prediction system based on a statistical-dynamic and machine learning fusion model that can capture complex cyclone change relationships, enhance real-time adaptability, and reduce uncertainty and bias has become a technical problem that needs to be solved urgently. Summary of the invention

[0005] In order to solve the above problems, the present invention provides a tropical cyclone wind radius prediction system based on a statistical-dynamic and machine learning fusion model.

[0006] The technical solution of the present invention is a tropical cyclone wind radius prediction system based on a statistical-dynamic and machine learning fusion model, which is provided with a statistical-dynamic model and a machine learning model. The statistical-dynamic model is based on an improved Rankine vortex model, and the wind speed V is described as a function of the radius r and the azimuth angle θ; by collecting historical cyclone data and climate statistical data, a basic wind speed estimate is obtained, and then combined with the historical cyclone development trend, a persistence method is used to predict the short-term change of the cyclone radius; then real-time data is collected, and the model parameters are automatically adjusted using preset rules; finally, the iterative optimization technology genetic algorithm is used to obtain the optimal parameter settings during the model operation process; The machine learning model collects meteorological data and data output by the statistical-dynamic model, processes the data using feature engineering methods, obtains key features of historical data trends, environmental variables and statistical-dynamic model output data, and generates an input feature vector; after the features are input, historical cyclone data is used for training, and a cross-validation method is used in the process to optimize model parameters, and the index mean square error is calculated to evaluate model performance; after the training is completed, an integrated learning method is used to combine the output results of the statistical-dynamic model to form a comprehensive prediction result; finally, the model is dynamically updated by inputting real-time data to correct the comprehensive prediction result.

[0007] As a further improvement of the present invention, the function describing the wind speed V is: , , in, Indicates the maximum wind speed, represents the radius of maximum wind speed, x represents the shape parameter that controls the radial attenuation of wind speed, a is the asymmetry parameter of the wind field, is the position angle of asymmetry; the persistence method is described using a linear regression model, expressed as , in, and Represent the current and future wind circle radii, Indicates the current wind speed. , , represents the regression coefficient, represents the error term.

[0008] As a further improvement of the present invention, the preset rules scale the input and output of the variables in the statistical-dynamic model; then apply a penalty term to the error function, multiply the amount out of range in the vortex model parameters by a large coefficient, which is 10^6, to increase the root mean square error RMSE; finally, use the RMSE with the penalty term as the loss function of the model, find the minimum mean square error, and obtain the adjusted vortex model parameters.

[0009] As a further improvement of the present invention, the steps of the genetic algorithm include: s1, introduce the parameters of the statistical-dynamic model as the target, and randomly generate a set of parameter combinations from the target as the initial population; s2, the difference between the prediction result of the statistical-dynamic model and the actual data is used as the fitness function, which uses the mean square error MSE formula, expressed as , in, is the predicted wind circle radius, is the actual observed value; the fitness function is used to calculate the fitness of each individual in the population, and the size of the fitness is inversely proportional to the mean square error; s3, using a selection strategy to select individuals with higher fitness from the population as parents; the selection strategy includes roulette wheel selection and tournament selection; s4, randomly select two parent individuals and generate new offspring through crossover operation, which includes single-point crossover and multi-point crossover; s5, randomly select some offspring individuals to make small parameter changes to increase the diversity of the population and prevent falling into the local optimum; s6, replace individuals with lower fitness with newly generated offspring to form a new population; s7, when the fitness of individuals in the population reaches the preset threshold, or the number of iterations reaches the upper limit, the algorithm is terminated.

[0010] As a further improvement of the present invention, the machine learning model selects a model suitable for processing time series and nonlinear relationships for training, and the selected objects include random forest, XGBoost and LSTM models.

[0011] As a further improvement of the present invention, the meteorological data include the temperature, humidity, wind speed, moving speed and central air pressure of the cyclone, and the data output by the statistical-dynamic model include the wind circle radius, storm path and intensity changes of the cyclone; the historical data trends include the temperature at historical moments, the humidity at historical moments, the wind speed at historical moments and the central air pressure at historical moments, and the environmental variables include the current temperature, the current humidity, the current wind speed, the current moving speed and the current central air pressure.

[0012] As a further improvement of the present invention, the ensemble learning method obtains a comprehensive prediction result by weighted averaging the output results of machine learning and statistical-dynamic models, which is expressed as: ,in, and is the weight, is the prediction result of the machine learning model, is the prediction result of the statistical-dynamic model; the adjustment formula of the dynamic update model is , where w is the learning rate, L is the loss function, is the parameter value updated at time step t+1 after adjustment based on the model parameter at time t. Indicates that at time t, the loss function L has an effect on the parameter gradient.

[0013] As a further improvement of the present invention, the machine learning model monitors and collects the performance index data of the system, and uses feedback to optimize the machine learning model. The deviation formula is: ,in, represents the predicted value, Represents the true value; after optimizing the model, A / B testing is used to evaluate the model effect; the machine learning model also regularly uses reinforcement learning methods to collect new meteorological data to form an updated data set to improve the performance of the machine learning model.

[0014] After adopting the above method, the prediction of the target cyclone radius is first obtained through the statistical-dynamic model, and then the meteorological data and the data output by the statistical-dynamic model are input into the machine learning model. The trained machine learning model outputs the predicted cyclone radius, identifies and captures the complex nonlinear relationship in the meteorological data, and has higher accuracy in predicting the wind circle radius; based on the integrated learning method, it is combined with the output results of the statistical-dynamic model to form a comprehensive prediction result, which can reduce the restrictions of the model assumptions on the prediction results and further optimize the estimation accuracy of the wind circle radius. When calculating the comprehensive prediction results, weights w1 and w2 are assigned respectively to indicate the importance of the model prediction results in the overall prediction.

[0015] By collecting real-time data, using preset rules to automatically adjust the parameters of the statistical-dynamic model, and utilizing iterative optimization technology genetic algorithms, the optimal parameter settings are obtained during the model operation process, and the statistical-dynamic model is dynamically updated; the system's performance indicator data is monitored and collected through machine learning models, and feedback measurement is used to optimize the machine learning model. Real-time observation data is effectively used for forecast updates to improve the timeliness and adaptability of the forecast; and reinforcement learning methods are also used regularly to collect new meteorological data to form an updated data set to improve the performance of the machine learning model, so that the model can reflect the latest environmental and cyclone changes and adapt to the dynamic development process of cyclones.

[0016] By using an integrated learning method, the output results of the machine learning model and the statistical-dynamic model are combined to form a comprehensive prediction result, reducing the uncertainty and bias of a single model, ensuring the stability and accuracy of the model's prediction of cyclone radius changes, and providing more reliable support for meteorological forecasting and emergency response.

[0017] By continuously optimizing parameter combinations through genetic algorithms and automating parameter adjustments, it is possible to quickly respond to real-time data changes, improve the adaptability and real-time performance of statistical-dynamic models, and effectively explore a vast parameter space to avoid falling into local optimal solutions, thereby ensuring the robustness of the model under different meteorological conditions, significantly improving prediction accuracy, and reducing prediction errors. DETAILED DESCRIPTION

[0018] A tropical cyclone wind radius prediction system based on a statistical-dynamic and machine learning fusion model, which is equipped with a statistical-dynamic model and a machine learning model. The statistical-dynamic model is based on an improved Rankine vortex model, and the wind speed V is described as a function of the radius r and the azimuth angle θ; by collecting historical cyclone data and climate statistical data, a basic wind speed estimate is obtained, and then combined with the historical cyclone development trend, a persistence method is used to predict the short-term change of the cyclone radius; then real-time data is collected, and the model parameters are automatically adjusted using preset rules; finally, the iterative optimization technology genetic algorithm is used to obtain the optimal parameter settings during the model operation process; The machine learning model collects meteorological data and data output by the statistical-dynamic model, processes the data using feature engineering methods, obtains key features of historical data trends, environmental variables and statistical-dynamic model output data, and generates an input feature vector; after the features are input, historical cyclone data is used for training, and a cross-validation method is used in the process to optimize model parameters, and the index mean square error is calculated to evaluate model performance; after the training is completed, an integrated learning method is used to combine the output results of the statistical-dynamic model to form a comprehensive prediction result; finally, the model is dynamically updated by inputting real-time data to correct the comprehensive prediction result.

[0019] The function describing the wind speed V is , , in, Indicates the maximum wind speed, represents the radius of maximum wind speed, x represents the shape parameter that controls the radial attenuation of wind speed, a is the asymmetry parameter of the wind field, is the position angle of asymmetry; the persistence method is described using a linear regression model, expressed as , in, and Represent the current and future wind circle radii, Indicates the current wind speed. , , represents the regression coefficient, represents the error term.

[0020] The preset rule scales the input and output of the variables in the statistical-dynamic model; then applies a penalty term to the error function, multiplies the out-of-range amount in the vortex model parameter by a large coefficient, which is 10^6, to increase the root mean square error (RMSE); finally, uses the RMSE with the penalty term as the loss function of the model, finds the minimum mean square error, and obtains the adjusted vortex model parameters.

[0021] The steps of the genetic algorithm include: s1, introduce the parameters of the statistical-dynamic model as the target, and randomly generate a set of parameter combinations from the target as the initial population; s2, the difference between the prediction result of the statistical-dynamic model and the actual data is used as the fitness function, which uses the mean square error MSE formula, expressed as , in, is the predicted wind circle radius, is the actual observed value; the fitness function is used to calculate the fitness of each individual in the population, and the size of the fitness is inversely proportional to the mean square error; s3, using a selection strategy to select individuals with higher fitness from the population as parents; this selection strategy includes roulette wheel selection and tournament selection; s4, randomly select two parent individuals and generate new offspring through crossover operation, which includes single-point crossover and multi-point crossover; s5, randomly select some offspring individuals to make small parameter changes to increase the diversity of the population and prevent falling into the local optimum; s6, replace individuals with lower fitness with newly generated offspring to form a new population; s7, when the fitness of individuals in the population reaches the preset threshold, or the number of iterations reaches the upper limit, the algorithm is terminated.

[0022] The machine learning model selects a model suitable for processing time series and nonlinear relationships for training, and the selected objects include random forest, XGBoost and LSTM models.

[0023] The meteorological data include the temperature, humidity, wind speed, moving speed and central air pressure of the cyclone, and the data output by the statistical-dynamic model include the wind circle radius, storm path and intensity change of the cyclone; the historical data trends include the temperature at historical moments, the humidity at historical moments, the wind speed at historical moments and the central air pressure at historical moments, and the environmental variables include the current temperature, the current humidity, the current wind speed, the current moving speed and the current central air pressure.

[0024] The ensemble learning method obtains a comprehensive prediction result by weighted averaging the output results of machine learning and statistical-dynamic models, which can be expressed as: ,in, and is the weight, is the prediction result of the machine learning model, is the prediction result of the statistical-dynamic model; the adjustment formula of the dynamic update model is , where w is the learning rate, L is the loss function, is the parameter value updated at time step t+1 after adjustment based on the model parameter at time t. Indicates that at time t, the loss function L has an effect on the parameter gradient.

[0025] The machine learning model monitors and collects system performance indicator data and uses feedback to optimize the machine learning model. The deviation formula is: ,in, represents the predicted value, Represents the true value; after optimizing the model, A / B testing is used to evaluate the model effect; the machine learning model also regularly uses reinforcement learning methods to collect new meteorological data to form an updated data set to improve the performance of the machine learning model.

[0026] The prediction of the target cyclone radius is first obtained through the statistical-dynamic model, and then the meteorological data and the data output by the statistical-dynamic model are input into the machine learning model. The trained machine learning model outputs the predicted cyclone radius, identifies and captures the complex nonlinear relationship in the meteorological data, and has higher accuracy in predicting the wind circle radius; based on the ensemble learning method, it is combined with the output results of the statistical-dynamic model to form a comprehensive prediction result, which can reduce the restrictions of the model assumptions on the prediction results and further optimize the estimation accuracy of the wind circle radius. When calculating the comprehensive prediction results, weights w1 and w2 are respectively assigned to indicate the importance of the model prediction results in the overall prediction.

[0027] By collecting real-time data, using preset rules to automatically adjust the parameters of the statistical-dynamic model, and utilizing iterative optimization technology genetic algorithms, the optimal parameter settings are obtained during the model operation process, and the statistical-dynamic model is dynamically updated; the system's performance indicator data is monitored and collected through machine learning models, and feedback measurement is used to optimize the machine learning model. Real-time observation data is effectively used for forecast updates to improve the timeliness and adaptability of the forecast; and reinforcement learning methods are also used regularly to collect new meteorological data to form an updated data set to improve the performance of the machine learning model, so that the model can reflect the latest environmental and cyclone changes and adapt to the dynamic development process of cyclones.

[0028] By using an integrated learning method, the output results of the machine learning model and the statistical-dynamic model are combined to form a comprehensive prediction result, reducing the uncertainty and bias of a single model, ensuring the stability and accuracy of the model's prediction of cyclone radius changes, and providing more reliable support for meteorological forecasting and emergency response.

[0029] By continuously optimizing parameter combinations through genetic algorithms and automating parameter adjustments, it is possible to quickly respond to real-time data changes, improve the adaptability and real-time performance of statistical-dynamic models, and effectively explore a vast parameter space to avoid falling into local optimal solutions, thereby ensuring the robustness of the model under different meteorological conditions, significantly improving prediction accuracy, and reducing prediction errors.

Claims

1. A tropical cyclone wind radius prediction system based on a statistical-dynamic and machine learning fusion model, characterized by: There are statistical-dynamic models and machine learning models. The statistical-dynamic model is based on the improved Rankine vortex model, and the wind speed V is described as a function of the radius r and the azimuth angle θ; by collecting historical cyclone data and climate statistical data, the basic wind speed estimation is obtained, and then combined with the historical cyclone development trend, the persistence method is used to predict the short-term change of the cyclone radius; Then, real-time data is collected and model parameters are automatically adjusted using preset rules; finally, the iterative optimization technique genetic algorithm is used to obtain the best parameter settings during the model operation; The machine learning model collects meteorological data and data output by the statistical-dynamic model, processes the data using a feature engineering method, obtains key features of historical data trends, environmental variables, and data output by the statistical-dynamic model, and generates an input feature vector; After the features are input, historical cyclone data is used for training. During the process, the cross-validation method is used to optimize the model parameters, and the mean square error of the indicator is calculated to evaluate the model performance. After the training is completed, the integrated learning method is used to combine it with the output results of the statistical-dynamic model to form a comprehensive prediction result. Finally, by inputting real-time data, the model is dynamically updated and the comprehensive prediction result is corrected.

2. A tropical cyclone wind radius prediction system based on a statistical-dynamic and machine learning fusion model according to claim 1, characterized in that: The function describing the wind speed V is , , in, Indicates the maximum wind speed, represents the radius of maximum wind speed, x represents the shape parameter that controls the radial attenuation of wind speed, a is the asymmetry parameter of the wind field, is the position angle of asymmetry; the persistence method is described using a linear regression model, expressed as , in, and Represent the current and future wind circle radii, Indicates the current wind speed. , , represents the regression coefficient, represents the error term.

3. The tropical cyclone wind radius prediction system based on the statistical-dynamic and machine learning fusion model according to claim 1, characterized in that: The preset rule scales the input and output of the variables in the statistical-dynamic model; then applies a penalty term to the error function, multiplies the out-of-range amount in the vortex model parameter by a large coefficient, which is 10^6, to increase the root mean square error (RMSE); finally, uses the RMSE with the penalty term as the loss function of the model, finds the minimum mean square error, and obtains the adjusted vortex model parameters.

4. The tropical cyclone wind radius prediction system based on the statistical-dynamic and machine learning fusion model according to claim 1, characterized in that: The steps of the genetic algorithm include: s1, introduce the parameters of the statistical-dynamic model as the target, and randomly generate a set of parameter combinations from the target as the initial population; s2, the difference between the prediction result of the statistical-dynamic model and the actual data is used as the fitness function, which uses the mean square error MSE formula, expressed as , in, is the predicted wind circle radius, is the actual observed value; the fitness function is used to calculate the fitness of each individual in the population, and the size of the fitness is inversely proportional to the mean square error; s3, using a selection strategy to select individuals with higher fitness from the population as parents; this selection strategy includes roulette wheel selection and tournament selection; s4, randomly select two parent individuals and generate new offspring through crossover operation, which includes single-point crossover and multi-point crossover; s5, randomly select some offspring individuals to make small parameter changes to increase the diversity of the population and prevent falling into the local optimum; s6, replace individuals with lower fitness with newly generated offspring to form a new population; s7, when the fitness of individuals in the population reaches the preset threshold, or the number of iterations reaches the upper limit, the algorithm is terminated.

5. The tropical cyclone wind radius prediction system based on the statistical-dynamic and machine learning fusion model according to claim 1, characterized in that: The machine learning model selects a model suitable for processing time series and nonlinear relationships for training, and the selected objects include random forest, XGBoost and LSTM models.

6. The tropical cyclone wind radius prediction system based on the statistical-dynamic and machine learning fusion model according to claim 1, characterized in that: The meteorological data include the temperature, humidity, wind speed, moving speed and central air pressure of the cyclone, and the data output by the statistical-dynamic model include the wind circle radius, storm path and intensity change of the cyclone; the historical data trends include the temperature at historical moments, the humidity at historical moments, the wind speed at historical moments and the central air pressure at historical moments, and the environmental variables include the current temperature, the current humidity, the current wind speed, the current moving speed and the current central air pressure.

7. A tropical cyclone wind radius prediction system based on a statistical-dynamic and machine learning fusion model according to claim 6, characterized in that: The ensemble learning method obtains a comprehensive prediction result by weighted averaging the output results of machine learning and statistical-dynamic models, which can be expressed as: ,in, and is the weight, is the prediction result of the machine learning model, is the prediction result of the statistical-dynamic model; the adjustment formula of the dynamic update model is , where w is the learning rate, L is the loss function, is the parameter value updated at time step t+1 after adjustment based on the model parameter at time t. Indicates that at time t, the loss function L has an effect on the parameter gradient.

8. A tropical cyclone wind radius prediction system based on a statistical-dynamic and machine learning fusion model according to claim 7, characterized in that: The machine learning model monitors and collects system performance indicator data and uses feedback to optimize the machine learning model. The deviation formula is: ,in, represents the predicted value, Represents the true value; after optimizing the model, A / B testing is used to evaluate the model effect; the machine learning model also regularly uses reinforcement learning methods to collect new meteorological data to form an updated data set to improve the performance of the machine learning model.

Images