Typhoon wind field model construction method and device based on machine learning

Abstract

The invention provides a typhoon wind field model construction method and device based on machine learning. The method comprises the following steps: acquiring actually observed satellite-borne microwave radiometer data, typhoon data of an American hurricane center and synthetic aperture radar data; related training parameters suitable for a machine learning model are selected, and typhoon asymmetric parameters are introduced; a typhoon wind field is reconstructed according to the obtained typhoon trajectory data by adopting various machine learning models; and verifying the accuracy of the typhoon wind field simulation result of the machine learning model. The method is suitable for obtaining the wind field in the typhoon process through a machine learning method, and improves the simulation precision of the typhoon wind field.

Description

Technical Field

[0001] The present invention relates to the field of typhoon process artificial intelligence technology, and in particular to a method and device for constructing a typhoon wind field model based on machine learning. Background Art

[0002] It is well known that strong winds during tropical cyclones (TCs) are usually associated with extreme sea states and heavy rainfall, which are typical hazards in coastal waters. Therefore, tropical cyclones are an interesting aspect of atmospheric and oceanographic research because wind is the main forcing of ocean surface dynamics. Traditionally, in situ observations from moored buoys are the most reliable source for tropical cyclone-related research. However, there are few moored buoys to choose from in the open ocean, especially in extreme sea conditions during tropical cyclones. A numerical weather prediction model based on nonlinear partial differential equations, namely the Weather Research and Forecasting Model (WRF), is a solution to describe the changes in atmospheric motion. Currently, the WRF model is widely used in reanalysis systems such as the Global Forecast System (GFS) of the National Centers for Environmental Prediction of the United States and the European Center for Medium-Range Weather Forecasts (ECMWF). However, it was found that the ECMWF wind reanalysis data was underestimated during a tropical cyclone. Since the numerical wave models WAVEWA Tropical Cyclone H-III (WW3) and Simulation WaveNearshore (SWAN) use reanalysis wind fields as forcing fields, the forecasted waves after tropical cyclones may also be considered underestimated.

[0003] At present, there are three main microwave frequency remote sensing technologies suitable for monitoring global ocean surface wind fields: scatterometers, microwave radiometers, and altimeters. Since the backscatter signal of the scatterometer has a saturation problem when the wind speed is >30m / s, the wind speed of the microwave radiometer and the altimeter can be used to study the characteristics of tropical cyclones such as motion, intensity, and asymmetry, and verify the simulation effect of the WRF model. However, the spatial resolution of these remote sensing products is relatively low, with a 0.25 strip grid for microwave radiometers and a 10km spatial resolution for altimeters. Synthetic aperture radar has good spatial resolution and a wide swath coverage, that is, Sentinel-1 (S-1) images acquired in the interferometric wide swath (IW) mode with a pixel size of 10m and 200km and the ultra-wide swath (EW) mode with a pixel size of 40m and 400km, which have rich complex sea conditions and polar information.

[0004] However, in the era of artificial intelligence, based on a large amount of sample data, machine learning can be as close to reality as possible, which has been successfully applied to atmosphere and oceanography. In order to solve the problem of inaccurate typhoon wind field simulation in the existing technology, it is necessary to propose a typhoon wind field model development method based on machine learning. Summary of the invention

[0005] In view of the defects in the prior art, the purpose of the present invention is to provide a method and device for constructing a typhoon wind field model based on machine learning. The present invention utilizes machine learning methods to improve the accuracy of typhoon wind field simulation.

[0006] To solve the above problems, the technical solution of the present invention is:

[0007] A method for constructing a typhoon wind field model based on machine learning comprises the following steps:

[0008] Obtain actual observed satellite-borne microwave radiometer data, US Hurricane Center typhoon data, and synthetic aperture radar data;

[0009] Select training parameters suitable for machine learning models and introduce typhoon asymmetry parameters;

[0010] Using a variety of machine learning models, typhoon wind fields are reconstructed based on the acquired typhoon track data;

[0011] Verify the accuracy of typhoon wind field simulation results of the machine learning model.

[0012] Preferably, the step of obtaining actually observed satellite-borne microwave radiometer data, U.S. Hurricane Center typhoon data, and synthetic aperture radar data specifically includes: collecting satellite-borne microwave radiometer data, U.S. Hurricane Center data, and synthetic aperture radar data of more than 200 typhoons from 2018 to 2022.

[0013] Preferably, the step of selecting training parameters applicable to the machine learning model and introducing typhoon asymmetric parameters specifically includes: selecting training parameters applicable to the machine learning model according to the Holland 10 model and introducing typhoon asymmetric parameters, the Holland 10 model is used to simulate the typhoon sea surface wind field profile at a height of 10m V10 with a radius of r, and the formula is:

[0014]

[0015] Where v m is the maximum wind speed, R m is the maximum wind speed radius, b s , x is the proportional parameter calculated by the equation and is defined as follows:

[0016]

[0017] Among them, ρ a is the air density constant, P c is the typhoon center pressure, x n With radius r nThe observed value is tuned; since the actual typhoon wind structure is asymmetric and the maximum wind is often located in the right front quadrant, an asymmetric term v needs to be added to the typhoon wind field add ,as follows:

[0018] v add =0.4v t sinβ

[0019] In the formula, v t is the typhoon moving speed, β is the angle relative to the typhoon moving direction;

[0020] The model training parameters finally selected include pressure radius r, typhoon center pressure, maximum wind speed, maximum wind speed radius, typhoon moving speed, typhoon moving direction and position, distance from the typhoon eye, and the average value of the maximum wind speed radius of 34kt, 50kt, and 64kt.

[0021] Preferably, in the step of using multiple machine learning models to reconstruct the typhoon wind field according to the acquired typhoon trajectory data, four machine learning models are used, including LightGBM, RF, XGBoost, and MLP.

[0022] Preferably, the XGBoost model is the optimal machine learning model for reconstructing typhoon wind fields.

[0023] Preferably, the step of verifying the accuracy of the typhoon wind field simulation results of the machine learning model specifically includes: comparing the typhoon wind speed reconstructed based on the machine learning method and the H10 model with the wind speed of the SMAP wind, performing time and space matching on the CyclObs and SMAP data, setting the time less than 2 hours and the space less than 10 kilometers as conditions for screening, and using the screening results as a training set for verification of the model simulated wind field and the SMAP wind field.

[0024] Furthermore, the present invention also provides a typhoon wind field model construction device based on machine learning, the device comprising a processor and a memory for storing executable instructions of the processor, and the processor is configured to execute the typhoon wind field model construction method based on machine learning as described above by executing the executable instructions.

[0025] Compared with the prior art, the present invention trains a variety of machine learning methods for reconstructing simulated wind fields during typhoons, thereby improving the accuracy of typhoon wind field simulation through machine learning methods and solving the problem of inaccurate typhoon wind field simulation in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0027] Figure 1 A flowchart of a method for constructing a typhoon wind field model based on machine learning provided in an embodiment of the present invention;

[0028] Figure 2 It is the best path and maximum wind speed diagram of TC;

[0029] Figure 3 This is the CyclObs wind map during Typhoon Morakot;

[0030] Figure 4 This is the SMAP wind map collected during Typhoon Shanshan;

[0031] Figure 5a , 5b , 5c, and 5d are the fitting result graphs of four machine learning models;

[0032] Figure 6a , 6b , 6c, 6d, 6e, and 6f are TC wind speed maps based on machine learning methods;

[0033] Figure 7a , 7b ,7c, 7d are the density plots of the comparison results between SMAP and four machine learning;

[0034] Figure 8 This is the density map of the wind speed comparison results between the H10 model and SMAP;

[0035] Figure 9a , 9b , 9c, and 9d are wind speed diagrams in four radial directions for the CyclObs product, the four machine learning models, and the H10 model. DETAILED DESCRIPTION

[0036] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0037] Specifically, the present invention provides a method for constructing a typhoon wind field model based on machine learning, such as Figure 1 As shown, the method comprises the following steps:

[0038] S1: Obtain actual observed satellite-borne microwave radiometer data, US Hurricane Center typhoon data, and synthetic aperture radar data;

[0039] Specifically, a total of Spaceborne Microwave Radiometer (SMAP) data of more than 200 typhoons from 2018 to 2022, typhoon data from the U.S. Hurricane Center (NHC) and synthetic aperture radar data were collected. The advantage of SMAP is that its spatial coverage can reach 100 kilometers, the time resolution is 1 hour, and the spatial resolution is 0.5°, which can effectively verify and analyze the typhoon wind field structure and intensity. Figure 2 This is the TC best track and wind speed map from NHC data. Figure 3 This is an example of wind speed for CyclObs products. Figure 4 This is an example of wind speed from the SMAP product.

[0040] The discontinuity of the SAR inverted wind field is significant, but the combination of VV and VH polarization NRCSs can improve the SAR wind field inversion of typhoons. On this basis, the CyclObs wind field with a spatial resolution of 500m was inverted from the dual-polarization S-1 image using the prior wind vector provided by ECMWF.

[0041] S2: Select training parameters related to the machine learning model and introduce typhoon asymmetry parameters;

[0042] Specifically, the key feature of using machine learning methods is to select appropriate parameters as input during the training process. The input variables include pressure at radius r, typhoon center pressure, maximum wind speed radius (RMWS), maximum wind speed (MWS), typhoon moving speed, the angle between the typhoon moving direction and position, the distance from the typhoon eye, and the average 34kt, 50kt, and 64kt maximum wind speed radius from the NHC. The output parameter is the wind speed in the CyclObs product. More than 300,000 samples were collected using the CyclObs wind speed product as the training process and fitting process. The fitting results of the four machine learning methods are shown in Figure 2. Figure 5a , 5b , 5c, and 5d, overall, the three machine learning models are based on the gradient boosting decision tree (Light Gradient Boosting Machine, LightGBM) ( Figure 5a As shown), extreme gradient boosting tree (eXtremeGradient Boosting, XGBoost) ( Figure 5c As shown), Multi-Layer Perceptron (MLP) ( Figure 5d The root mean square error (RMSE) of the random forest (RF) model is less than 5m / s, which is better than that of the random forest (RF) model ( Figure 5b ) is 5.17m / s. And the fitting effect of XGBoost model under high wind speed conditions is better than that of LightGBM and MLP models.

[0043] According to the Holland 10 model, the training parameters related to the machine learning model are selected, and the typhoon asymmetric parameters are introduced. The Holland 10 model is used to simulate the typhoon sea surface wind field profile at a height of 10m V10 with a radius of r. The formula is:

[0044]

[0045] Where v m is the maximum wind speed, R m is the maximum wind speed radius, b s , x is the proportional parameter calculated by the equation and is defined as follows:

[0046]

[0047] Among them, ρ a is the air density constant, generally 1.15kg / m 3 , p n Assume that the outer pressure of the typhoon is 1015hPa, P c is the typhoon center pressure, which can be obtained from NHC data, x n With radius r n Since the actual typhoon wind structure is asymmetric and the maximum wind is often located in the right front quadrant, an asymmetric term v needs to be added to the typhoon wind field. add ,as follows:

[0048] v add =0.4v t sinβ

[0049] In the formula, v t is the typhoon moving speed, and β is the angle relative to the typhoon moving direction.

[0050] With the foundation of the Holland 10 model, the model training parameters finally selected include pressure radius r, typhoon center pressure, maximum wind speed, maximum wind speed radius, typhoon moving speed, typhoon moving direction and position, distance from the typhoon eye, and the average value of the maximum wind speed radius of 34kt, 50kt, and 64kt.

[0051] S3: Use multiple machine learning models to reconstruct the typhoon wind field based on the acquired typhoon track data;

[0052] Specifically, in order to compare the machine learning-based approach with the reconstructed TC wind speed from the H10 model, Typhoon Morakot is selected as an example to demonstrate the algorithm results. Figure 6a , 6b , 6c, 6d, 6e, and 6f are TC wind speed maps based on machine learning methods. Figure 6a CyclObs products, Figure 6b LightGBM, Figure 6c RF, Figure 6d XGBoost, Figure 6e MLPs, Figure 6f The H10 model shows the wind speed map at 10:00 UTC on October 27, 2020. In each TC wind speed map, the TC eye is clear and the coverage is consistent. It is worth noting that the machine learning-based method ( Figure 6b~6e ) cannot reach the accuracy of CyclObs products ( Figure 6a ). The asymmetry of the typhoon structure needs to be further improved through machine learning-based methods. However, compared with the traditional H10 model ( Figure 6f ), the machine learning method effectively improves the stratification effect of the typhoon wind field and is closer to the typhoon process. Figure 6b~6e ) of the TC wind speed results, MLP( Figure 6e ) and XGBoost( Figure 6d ) model is better than LightGBM( Figure 6b ) and RF( Figure 6c ) models are more similar to the CyclObs product results, especially when the typhoon wind field is asymmetric. The reason is that the MLP and XGBoost models include an iterative training process to improve the training results. And the LightGBM and RF models do not have an iterative training process.

[0053] S4: Verify the accuracy of the typhoon wind field simulation results of the machine learning model.

[0054] Specifically, to further analyze the comparison between TC wind speed based on machine learning methods and SMAP measurement results, four machine learning methods were implemented in 50 TCs. The wind speed of the verification results can reach 60m / s. Figure 7a LightGBM, Figure 7b RF Figure 7c XGBoost and Figure 7d Density plot of the comparison results between the MLP machine learning models. The results show that the XGBoost model ( Figure 7c The wind speed root mean square error (RMSE) of the model (shown) is 3.89m / s, the correlation coefficient (COR) is 0.84, and the scatter index (SI) is 0.27, which is better than the other three machine learning methods, such as the LightGBM model ( Figure 7a The RMSE of the RF model (shown) is 4.85 m / s, the COR is 0.76, and the SI is 0.32; Figure 7b The RMSE of the MLP model (shown) is 4.17 m / s, the COR is 0.83, and the SI is 0.27; Figure 7dThe RMSE of the results is 5.25, COR is 0.71, and SI is 0.33. Figure 8 The H10 model was also implemented for the same TC case to compare with the SMAP model in Figure 8 As shown in Figure 2, the analysis results are RMSE of 12.29 m / s, COR of 0.56, and SI of 0.43, which are inferior to the results of the XGBoost model. Based on the above results, it is believed that the machine learning method is the optimal model for reconstructing TC wind speed, especially using the XGBoost model.

[0055] In addition to comparing the model results, we also compared the model wind speed errors at different directions and distances using four survey lines starting from the typhoon eye. Figure 9a , 9b , 9c, 9d give the wind speeds on four radial sections of the CyclObs product of TC Morafi; Figure 9a For the 45° direction wind diagram, Figure 9b This is the 135° direction wind diagram. Fig.9c For the 225° direction wind map, Figure 9d The wind map at 315° is shown in the figure, where the black line represents the CyclObs product wind, the red line represents the XGBoost model wind, the blue line represents the MLP model wind, the yellow line represents the RF model wind, the light blue line represents the LightGBM model wind, and the pink line represents the H10 model wind. According to the results, the pattern of wind profile inversion by the XGBoost method performs well in all four directions, especially in the 135° direction. However, the wind profile of the XGBoost model at 315° is significantly lower than that of the CyclObs product. In addition, the wind profiles of the MLP model and the H10 model are significantly higher than those of the CyclObs product. Therefore, the XGBoost model is an optional machine learning model for reconstructing typhoon wind fields.

[0056] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A method for constructing a typhoon wind field model based on machine learning, characterized in that: The method comprises the following steps: Obtain actual observed satellite-borne microwave radiometer data, US Hurricane Center typhoon data, and synthetic aperture radar data; Select training parameters suitable for machine learning models and introduce typhoon asymmetry parameters; Using a variety of machine learning models, typhoon wind fields are reconstructed based on the acquired typhoon track data; Verify the accuracy of typhoon wind field simulation results of the machine learning model.

2. The method for constructing a typhoon wind field model based on machine learning according to claim 1, characterized in that: The steps of obtaining the actually observed satellite-borne microwave radiometer data, US Hurricane Center typhoon data and synthetic aperture radar data specifically include: collecting satellite-borne microwave radiometer data, US Hurricane Center data and synthetic aperture radar data of more than 200 typhoons from 2018 to 2022.

3. The method for constructing a typhoon wind field model based on machine learning according to claim 1, characterized in that: The step of selecting training parameters applicable to the machine learning model and introducing typhoon asymmetric parameters specifically includes: selecting training parameters applicable to the machine learning model according to the Holland 10 model and introducing typhoon asymmetric parameters, wherein the Holland 10 model is used to simulate the typhoon sea surface wind field profile at a height of 10m V10 with a radius of r, and the formula is: Where v m is the maximum wind speed, R m is the maximum wind speed radius, b s , x is the proportional parameter calculated by the equation and is defined as follows: Among them, ρ a is the air density constant, P c is the typhoon center pressure, x n With radius r n The observed value is tuned; since the actual typhoon wind structure is asymmetric and the maximum wind is often located in the right front quadrant, an asymmetric term v needs to be added to the typhoon wind field add ,as follows: v add =0.4v t sinβ In the formula, v t is the typhoon moving speed, β is the angle relative to the typhoon moving direction; The model training parameters finally selected include pressure radius r, typhoon center pressure, maximum wind speed, maximum wind speed radius, typhoon moving speed, typhoon moving direction and position, distance from the typhoon eye, and the average value of the maximum wind speed radius of 34kt, 50kt, and 64kt.

4. The method for constructing a typhoon wind field model based on machine learning according to claim 1, characterized in that: In the step of using multiple machine learning models to reconstruct the typhoon wind field according to the acquired typhoon track data, four machine learning models are used, including LightGBM, RF, XGBoost, and MLP.

5. The method for constructing a typhoon wind field model based on machine learning according to claim 4, characterized in that: The XGBoost model is the optimal machine learning model for reconstructing typhoon wind fields.

6. The method for constructing a typhoon wind field model based on machine learning according to claim 1, characterized in that: The step of verifying the accuracy of the typhoon wind field simulation results of the machine learning model specifically includes: comparing the typhoon wind speed reconstructed based on the machine learning method and the H10 model with the wind speed of the SMAP wind, performing time and space matching on the CyclObs and SMAP data, setting the time less than 2 hours and the space less than 10 kilometers as conditions for screening, and using the screening results as a training set for verification of the model simulated wind field and the SMAP wind field.

7. A typhoon wind field model construction device based on machine learning, characterized in that: The device includes a processor and a memory for storing executable instructions of the processor, and the processor is configured to execute the machine learning-based typhoon wind field model construction method as described in any one of claims 1 to 6 by executing the executable instructions.

Images