TC path and intensity prediction method and device based on AI and physical characteristics, and medium

Abstract

The invention discloses a TC path and intensity prediction method and device based on AI and physical characteristics and a medium, and aims to improve the short time scale prediction precision of the TC path and intensity, and through integration of a plurality of factors with definite physical significance, such as guide airflow, ventilation flow, vorticity, ocean surface temperature and the like, the TC path and intensity can be predicted. The physical factors are subjected to correlation analysis, regression analysis and the like, and it is ensured that the input factors have significant influences on TC paths and intensity changes. A plurality of AI models such as a convolutional neural network (CNN), a recurrent neural network (RNN), a multi-layer perceptron (MLP), a random forest (RF), a limit gradient boost (XGBoost) and a long short-term memory network (LSTM) are adopted to capture a complex nonlinear relationship between meteorological factors. In addition, by adopting a real-time feedback mechanism, new observation data are allowed to be dynamically input, model parameters are optimized, and the real-time performance of a prediction result is improved. The method is particularly suitable for the northwest Pacific Ocean region, has wide application prospects, and provides scientific support in the fields of meteorology, aviation, marine transportation and the like.

Description

Technical Field

[0001] The present invention relates to the field of meteorological forecasting and artificial intelligence (AI) intersection technology, and relates to the prediction of tropical cyclone paths and intensities. The present invention specifically relates to a prediction method for tropical cyclone (TC) paths and intensities based on AI technology and meteorological ocean physical characteristics, a computer device and a readable storage medium. The accuracy, timeliness and interpretability of tropical cyclone path and intensity predictions are improved by integrating AI technology and ocean-atmospheric physical processes. Background Art

[0002] Tropical cyclone (TC) is an important and complex phenomenon in meteorology. It can cause extreme weather such as strong winds, heavy rains, floods and storm surges. Therefore, improving the accuracy and timeliness of TC path and intensity forecasts has always been an important issue in meteorological disaster forecasting. With the intensification of global climate change, the frequent occurrence of extreme weather events makes the accuracy and timeliness of TC forecasts more important.

[0003] Traditional TC forecasts mainly rely on Numerical Weather Prediction (NWP) models, which simulate the evolution of meteorological systems by solving atmospheric dynamics and thermodynamics equations. However, although NWP models show high accuracy in long-term forecasts, they often face many challenges in short-term forecasts (such as 6-24 hours). First, although data assimilation technology is becoming more and more mature, the quality and quantity of real-time observation data often cannot meet the needs, thus affecting the performance of the model. Secondly, NWP models require a lot of computing resources, especially in high-resolution conditions, which limits their application. In addition, the evolution of TC is jointly affected by multiple meteorological factors, such as vertical wind shear of the atmosphere, ocean surface temperature, ambient humidity, etc. Traditional models are not capable of handling these complex nonlinear relationships, which often leads to deviations in prediction results.

[0004] In recent years, the rapid development of artificial intelligence (AI) technology has provided new possibilities for weather forecasting. Using machine learning and deep learning algorithms, AI models are able to extract useful features from large-scale meteorological data and capture the complex relationships between meteorological factors. For example, advanced AI models such as convolutional neural networks (CNN) and recurrent neural networks (RNN) have been widely used in the prediction of TC paths and intensities. The advantages of these models lie in their powerful feature extraction capabilities and adaptability to complex patterns, which can surpass traditional numerical models in some cases. However, although AI models perform well in some occasions, their prediction results often lack interpretability, especially when it comes to the interaction of multiple meteorological factors.

[0005] Most of the current research involving AI models has failed to fully consider factors that have obvious physical significance for TC paths and intensity, such as guiding air currents that have a greater impact on typhoon paths, ocean surface and subsurface temperatures, and mesoscale ocean eddy information that have a greater impact on typhoon intensity. These factors play a vital role in influencing the movement and intensity of TCs, but in existing AI models, because their focus is more on data-driven algorithm optimization, the direct connection between these physical factors and tropical cyclone behavior is often ignored. If they are included in the AI ​​model input, it is expected to significantly improve the accuracy and reliability of the forecast. Secondly, although existing AI models have powerful feature extraction capabilities, they often lack a deep understanding of physical processes, which limits the generalization ability of the model and its performance in extreme situations.

[0006] In summary, how to effectively combine physical processes with artificial intelligence technology, and fully consider the synergistic effect of multiple influencing factors, to build an innovative TC path and intensity forecast model with high efficiency, high accuracy and good interpretability, so as to improve the accuracy and timeliness of TC forecasts, and provide more effective support for meteorological services, disaster management and related decision-making, is a technical problem that needs to be urgently solved in the current research field of TC path and intensity forecasting. Summary of the invention

[0007] 1. Purpose of the invention In view of the defects and shortcomings of existing tropical cyclone path and intensity forecasts, such as insufficient accuracy and timeliness, poor physical interpretability of AI models, and insufficient consideration of important meteorological physical factors, in order to solve at least one of the above and other technical problems in the prior art, the present invention aims to provide a TC path and intensity prediction method, computer equipment and medium based on AI and physical characteristics. By identifying and integrating multiple influencing factors with clear physical significance, such as guiding airflow, ventilation flow, vorticity, ocean surface temperature and vertical wind shear, combined with AI models and introducing a real-time feedback mechanism to dynamically update the prediction data, efficient and accurate TC path and intensity prediction can be achieved. It is particularly suitable for responding to climate change and extreme weather events, which not only significantly improves the accuracy, timeliness and interpretability of the forecast, but also provides reliable support for meteorological services, disaster management and related industries.

[0008] (II) Technical solution In order to achieve the purpose of the invention and solve the technical problems, the present invention adopts the following technical solutions: The first invention object of the present invention is to provide a TC path and intensity prediction method based on AI and physical characteristics, which is used to predict the path and intensity changes of tropical cyclones on a short time scale (6-24 hours). The prediction method at least includes the following steps when implemented: SS1. Data Collection At least five consecutive years of historical meteorological data from multiple meteorological and oceanographic data sources, including at least TC Best Track data that records track and intensity information, Final Analysis Data (FNL) that provides six-hourly updated meteorological variables, and ocean reanalysis data generated by the Global Ocean Data Assimilation System (GODAS); SS2. Data preprocessing Based on the historical meteorological and oceanographic data collected in step SS1, extract the relevant data of each TC occurrence period in the target prediction area (e.g., the northwest Pacific region) and perform data quality control on it, including identifying and eliminating missing or abnormal data, standardizing and normalizing the data to ensure the temporal and spatial consistency between different data sources, and then divide the processed data set into a training set and a validation set, and ensure that the training data and the validation data are independent in terms of time (different years or different periods) and space (different geographical regions); SS3. Selection and calculation of model output factors Based on the historical TC Best Track data preprocessed in step SS2 and the characteristics of the target forecast area, multiple variables closely related to TC paths and intensity changes are selected as model output factors, and the model output factors at least include TC position changes (longitude and latitude changes) and intensity changes (central minimum pressure and maximum wind speed changes). The time series of these variables are calculated to extract characteristic changes and establish model output targets; SS4. Selection and calculation of model input factors Based on the physical mechanism of TC movement and intensity change, and combined with the FNL atmospheric reanalysis data and GODAS ocean reanalysis data preprocessed in step SS2, multiple physical factors related to TC movement and intensity change are selected and calculated. Then, the correlation analysis between the selected physical factors and TC path and intensity change is carried out, and the correlation coefficient between each physical factor and the model output factor is calculated. Finally, the model input factor set is determined based on the analysis results. SS5. Construction and training of TC track and intensity prediction model Based on the model output factors and model input factors determined in steps SS3 and SS4, at least one AI model selected from convolutional neural network (CNN), recurrent neural network (RNN), multi-layer perceptron (MLP), long short-term memory network (LSTM), random forest (RF) and extreme gradient boosting (XGBoost) is constructed to build a TC path and intensity prediction model, and the constructed TC path and intensity prediction model is trained using the training set data in step SS2; SS6. Verification and optimization of TC path and intensity prediction model Using the validation set data in step SS2, the TC track and intensity prediction model trained in step SS5 is validated, and the prediction performance of the model at different forecast time periods (e.g., 6 hours, 12 hours, and 24 hours) is evaluated. The evaluation indicators include at least the track forecast error and the intensity forecast error. Based on the validation results, the model is optimized and adjusted by at least adjusting hyperparameters, increasing regularization, and / or using a cross-validation method to improve the prediction accuracy and generalization ability of the model. SS7. Real-time forecast and dynamic updates Establish a real-time feedback mechanism to dynamically input new observed meteorological data into the optimized model in step SS6 after preprocessing, continuously adjust and optimize model parameters through online learning and / or regular retraining triggered by real-time changes in new data, and regularly evaluate model performance and update input factor sets and model structure to improve the timeliness and accuracy of TC track and intensity forecasts; SS8. Output of prediction results Generate and output the forecast results of TC path change (longitude and latitude) and intensity change (minimum air pressure and maximum wind speed) in the next 6 hours, 12 hours and 24 hours, calculate and output the forecast error, including but not limited to path forecast error and intensity forecast error, provide credibility assessment of the forecast results, and provide decision support information for meteorological departments and disaster prevention and mitigation agencies.

[0009] The second invention object of the present invention is to provide a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the steps of the above-mentioned TC path and intensity prediction method based on AI and physical characteristics.

[0010] The third invention object of the present invention is to provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-mentioned TC path and intensity prediction method based on AI and physical characteristics.

[0011] (III) Technical Effect Compared with the prior art, the TC path and intensity prediction method, device and medium based on AI and physical properties of the present invention have the following beneficial and significant technical effects: (1) The present invention significantly improves the accuracy of tropical cyclone (TC) path and intensity forecasts by integrating traditional meteorological physics with modern AI technology. By selecting physical factors related to TC movement and intensity changes, such as guiding airflow, ventilation flow, vertical wind shear and sea surface temperature, and combining the powerful nonlinear fitting ability of the AI ​​model, the evolution process of TC can be captured more accurately. At the same time, the prediction method of the present invention is suitable for short-term forecasts (6 to 24 hours), which greatly improves the response speed to sudden meteorological events and provides effective technical support for disaster prevention and mitigation.

[0012] (2) The core of this invention is to identify and integrate 35 influencing factors with clear physical meanings, such as guiding airflow, ventilation flow, vorticity, sea surface temperature and vertical wind shear. These factors have been identified through correlation analysis and regression analysis to determine their significant impact on TC path and intensity changes. This method not only enhances the interpretability of the model, but also makes the prediction results more consistent with meteorological principles, thereby improving the accuracy and reliability of the prediction.

[0013] (3) The present invention designs a variety of artificial intelligence models, including convolutional neural networks (CNN), recurrent neural networks (RNN), multi-layer perceptrons (MLP), random forests (RF), extreme gradient boosting (XGBoost) and long short-term memory networks (LSTM). These models are designed to capture the complex nonlinear relationships between meteorological factors and provide accurate TC path and intensity forecasts. In addition, these models are trained using TC data from 2015 to 2020, and the data from 2021 to 2022 are used for verification to ensure the effectiveness and adaptability of the models.

[0014] (4) The present invention also introduces a real-time feedback mechanism, which enables new observation data to be dynamically input into the model, thereby continuously adjusting and optimizing the prediction results. This mechanism not only improves the real-time performance of TC forecasts, but also enables online learning and regular retraining based on the latest data, ensuring that the model can always maintain a high prediction accuracy.

[0015] (5) The forecast model of the present invention can not only output the TC path and intensity changes in the next 6 hours, 12 hours and 24 hours, but also provide error assessment and credibility analysis of the path and intensity forecast. This function helps meteorological experts make decisions based on the reliability of the forecast results, thereby improving the use value and reliability of the forecast information.

[0016] (6) The TC forecast model of the present invention can significantly improve the accuracy and timeliness of path and intensity forecasts, providing an important basis for decision-making in the fields of meteorology, aviation, shipping, agriculture, etc. This technology has broad application prospects in the field of meteorological forecasting and can effectively enhance the ability to respond to extreme weather. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1 FIG. 4 is a flowchart of the implementation of the TC path and intensity prediction method based on AI and physical properties of the present invention; Figure 2 The figure shows the schematic diagram of variable factor statistics, where (A) is the schematic diagram of the calculation area of ​​atmospheric background field variables, and (B) is the schematic diagram of the calculation area of ​​ocean variables and geographic variables; Figure 3 The figure shows a comparison of the forecast errors of different models for all TC paths and intensities in 2021-2022, where (a) is the 6-24h forecast error of TC paths, and (b) is the 6-24h forecast error of TC intensity; Figure 4 The figure shows the comparison of the path forecast results of TC No. 10 in 2021 by various models, where (a) to (d) correspond to the path forecast results of Models 1-4, and (e) to (j) correspond to the path forecast results of Models 5-10; Figure 5 The figure shows the comparison of the intensity forecast results of each model for TC No. 10 in 2021, where (a) to (d) correspond to the intensity forecast results of Models 1-4, and (e) to (j) correspond to the intensity forecast results of Models 5-10; Figure 6 The figure shows the path return results of the AI ​​model RF for all TCs from August to October 2021, where (a) to (l) correspond to the path return results of TC202109 to TC202120 respectively; Figure 7 Shown is a schematic diagram of the intensity return results of the AI ​​model RF for all TCs from August to October 2021, where (a) to (l) correspond to the intensity return results of TC202109 to TC202120, respectively. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical scheme and advantages of the implementation of the present invention clearer, the technical scheme in the embodiment of the present invention will be described in more detail below in conjunction with the drawings in the embodiment of the present invention. The described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. The embodiments described below with reference to the drawings are exemplary and are intended to be used to explain the present invention, and should not be understood as limiting the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0020] The present invention aims to provide a TC path and intensity prediction method, computer equipment and medium based on AI and physical characteristics. By identifying and integrating multiple influencing factors with clear physical meanings, such as guiding airflow, ventilation flow, vorticity, ocean surface temperature and vertical wind shear, combined with AI models and introducing real-time feedback mechanisms to dynamically update prediction data, efficient and accurate TC path and intensity predictions can be achieved. This method is particularly suitable for responding to climate change and extreme weather events, which not only significantly improves the accuracy, timeliness and interpretability of forecasts, but also provides reliable support for meteorological services, disaster management and related industries.

[0021] As a specific example, Figure 1 As shown, the TC path and intensity prediction method based on AI and physical characteristics of the present invention is used to predict the path and intensity changes of tropical cyclones on a short time scale (6-24 hours). The implementation mainly includes the following steps:

[0022] At least five consecutive years of historical meteorological data are obtained from multiple meteorological and oceanographic data sources, including TC data (TC Best Track) that records path and intensity information, atmospheric reanalysis data (FNL) that provides meteorological variables updated every six hours, and ocean reanalysis data (Global Ocean Data Assimilation System, GODAS) generated by the Global Ocean Data Assimilation System.

[0023] The data and related information used in this invention are all commonly used data sets in the process of studying TC. The TC BestTrack data set can have multiple sources. The data of the Japan Meteorological Agency (JMA) is updated the earliest, so the BestTrack data set of JMA is selected, and the data is downloaded directly from its website. FNL data and GODAS data can be downloaded from the website of the Information Center of the China Meteorological Administration. The time period for all data acquisition is from January 1, 2015 to December 31, 2022. The information of the period of TC occurrence in the northwest Pacific is used, and some data are missing.

[0024] Based on the historical meteorological and oceanographic data collected in step SS1, extract the relevant data of each TC occurrence period in the target prediction area (e.g., the northwest Pacific region) and perform data quality control on it, including identifying and eliminating missing or abnormal data, standardizing and normalizing the data to ensure the temporal and spatial consistency between different data sources, and then divide the processed data set into a training set and a validation set, and ensure that the training data and the validation data are independent in terms of time (different years or different periods) and space (different geographical regions).

[0025] Preferably, in steps SS1 and SS2, information on the occurrence period of TC in the northwest Pacific from January 1, 2015 to December 31, 2022 is selected from multiple meteorological and oceanographic data sources, and the atmospheric and oceanic background field information in the report is counted once every 6 hours for each TC case, and missing data records are removed. The TC occurrence period from January 1, 2015 to December 31, 2020 is used as training data, of which there are 148 TC cases and 2,660 report statistics. All TCs from 2021 to 2011 are test data, with a total of 47 TC cases and 621 report statistics.

[0026] Preferably, in step SS2, when performing data quality control, the missing parts in the data can be processed by using mean filling, interpolation or deleting records with missing values, the outliers in the data can be identified and removed by using the statistical analysis method of Z-score or box plot, and the data of different scales can be processed by using Z-score standardization or Min-Max normalization, and the collected historical meteorological data can be synchronized to ensure that the time resolution of all data sources is consistent (such as one data point every 6 hours). In addition, data quality control can also use an automated data cleaning method based on machine learning to identify potential abnormal data through historical data learning, and automatically perform data correction or filling to improve the efficiency and accuracy of data preprocessing.

[0027] Based on the historical TC Best Track data preprocessed in step SS2 and the characteristics of the target forecast area, multiple variables closely related to TC path and intensity changes are selected as model output factors, and the model output factors at least include TC position changes (longitude and latitude changes) and intensity changes (central minimum pressure and maximum wind speed changes). The time series of these variables are calculated to extract characteristic changes and establish model output targets.

[0028] Preferably, in this step, the output factors of the model are the hourly position changes of TC delta(LON), delta(LAT), and hourly intensity changes of TC delta(SLP) and delta(MWS) calculated from the TC Best Track data set. The intensity changes of TC are generally represented by the minimum central pressure SLP of TC and the maximum central wind speed MWS of TC. The position and intensity changes of TC are calculated by the formula X_n=X(t+n)-X(t), where X is LON, LAT, SLP and MWS, t represents the current time, and n represents the hour. In the present invention, n is 6h, 12h and 24h respectively. There are 12 output factors in total, as shown in OutputTarget in Table 1 below.

[0029] Based on the physical mechanism of TC movement and intensity change, and combined with the FNL atmospheric reanalysis data and GODAS ocean reanalysis data preprocessed in step SS2, multiple physical factors related to TC movement and intensity change are selected and calculated. Then, the correlation analysis between the selected physical factors and the TC path and intensity change is carried out, and the correlation coefficient between each physical factor and the model output factor is calculated. Finally, the model input factor set is determined based on the analysis results.

[0030] Preferably, in the above step SS4, the multiple physical factors selected that are related to the TC motion and intensity change should at least include guiding airflows in the middle layer (500hPa-700hPa) and high layer (300hPa-500hPa), Beta drift force and ventilation flow, airflow vorticity in the middle and high layers, vertical wind shear at high and low altitudes, lower atmospheric temperature, lower atmospheric humidity, diagnostic CAPE value, sea surface temperature SST, average temperature of the upper 200m of the ocean, and LANDMASK of the TC underlying surface, so as to comprehensively consider the key physical processes that affect TC motion and intensity change.

[0031] Further preferably, in the above step SS4, the calculation of the Beta drift force and the ventilation flow at least includes: SS41. Extract the U and V wind field data of the three pressure layers of 500hPa, 600hPa and 700hPa within the range of 0-500km around TC from the FNL atmospheric reanalysis data, where U is the wind speed component in the x direction and V is the wind speed component in the y direction. The U and V wind fields of the three extracted pressure layers are vertically averaged to obtain the average wind field representing the middle atmosphere. SS42. Calculate the moving speed of TC using the current position and the position of TC in the previous 6 hours after obtaining the vertically averaged U and V wind field data, and subtract this speed from the averaged U and V wind fields to obtain the environmental wind field relative to TC, so as to eliminate the error caused by the movement of TC itself; SS43. The wind field within a radius of 500 km from the TC center is selected as the calculation area and decomposed into tangential wind and radial wind, where the tangential wind represents the wind speed component along the TC rotation direction, and the radial wind represents the wind speed component perpendicular to the tangential direction; SS44. Perform Fourier decomposition on radial wind and tangential wind respectively, extract the first-order wave component to describe the main asymmetric component in the wind field, and transform the first-order wave component obtained after decomposition into U and V wind speed components in Cartesian coordinate system through inverse Fourier transform; SS45. The back-calculated U and V wind speed components are averaged over the entire calculation area to obtain the magnitude and direction of the Beta drift force and ventilation flow, where the Beta drift force represents the deviation of the TC relative to the atmospheric environment guide airflow, and the ventilation flow represents the influence of the asymmetric secondary airflow.

[0032] Preferably, in the above step SS4, when determining the final model input factor, at least the following steps are included: First, the correlation coefficient between each physical factor and the model output factor is calculated, and the correlation coefficient is tested for significance level. According to the preset significance level (such as 1% or 5%), the correlation between each physical factor is evaluated to see whether it is statistically significant. Physical factors with statistically significant correlation are screened out, and factors that fail the significance level test are excluded. Secondly, multiple linear regression was performed on the factors that passed the significance test to evaluate the relative contribution of each physical factor to the changes in TC paths and intensities, so as to further screen out the input factors with higher importance. Afterwards, principal component analysis (PCA) was performed on the physical factors that passed the significance test and preliminary regression analysis. The principal components whose cumulative explained variance reached a preset threshold (e.g., 95%) were extracted by the PCA method to reduce the data dimension and eliminate the remaining multicollinearity. Finally, the results of correlation analysis, significance level test, regression analysis and principal component analysis are comprehensively considered to finally determine the input factor set of the model to ensure the prediction accuracy and stability of the model.

[0033] More specifically, there are many factors that affect TC movement and intensity. First, we need to rely on the relevant theories of TC movement and change to select input factors: (1) Steering Flow: From the perspective of vortex dynamics, the movement of TC mainly depends on the "steering effect" of large-scale circulation on TC. The large-scale environmental airflow around TC is called steering flow. There is a clear relationship between the movement of TC and the surrounding large-scale airflow. The average wind at an altitude of 500hPa-700hPa and a latitude radius of 5-7° from the center of TC can be defined as the steering flow of TC, which has the highest correlation with the movement direction of TC; in the northern hemisphere, TC always moves 10-20° to the left of the direction of the steering flow. Generally speaking, the low-level steering flow is not consistent with the direction of TC movement, mainly because the low-altitude weather system is greatly affected by the terrain and boundary layer. The influence of the middle and high-level steering flow on TC is relatively stable. Therefore, the present invention first calculates the middle-level (500hPa-700hPa) steering flow and the high-level (300hPa-500hPa) steering flow of TC as the primary input factors of the model.

[0034] (2) Beta drift and ventilation flow: Beta drift also significantly affects the movement of TCs, representing most of the deviation between TC motion and environmental guiding airflow. Beta drift is mainly caused by beta vortex pairs, which are mainly composed of a wave of wind field in the tangential direction of the TC and appear as a pair of counter-rotating vortices, also known as ventilation flow. Beta drift can be regarded as an asymmetric secondary guiding airflow induced by the interaction between cyclones and beta. Its initial distribution is the result of the dispersion of linear Rossby waves causing the symmetric circulation to deform. The maintenance and development of beta vortex pairs are mainly understood from the perspective of energetics. They can be strengthened by obtaining energy from the axisymmetric cyclonic circulation.

[0035] (3) Vorticity: From the perspective of physical equations, TC can be regarded as a vortex, and the movement of the vortex obeys the vorticity equation, so vorticity is one of the factors that must be considered. And when TC has abnormal path and intensity, it is inseparable from the influence of the nearby high-altitude cold vortex, which is why the vorticity factor must be considered. Therefore, the present invention also uses the airflow vorticity of the middle and high layers as input factors.

[0036] (4) Other factors: In addition to the above factors, which have a clear theoretical effect on the TC path, some other factors also have a great influence on the TC intensity in theory. For example, vertical wind shear at high and low altitudes can also cause TC to spin, which obviously affects the intensity of TC. Therefore, it is also one of the factors that have a very important influence on TC. In addition, the temperature of the lower atmosphere, the humidity of the lower atmosphere and the diagnostic CAPE value can indicate the potential for convective activity in a region. As a strong convective system, TC is also sensitive to this. As the main source of TC energy, the ocean continuously provides water vapor and heat to TC to promote the occurrence, development and movement of TC. Therefore, SST is very important for the movement and intensity of TC. At the same time, TC will also cause significant sea surface cooling, inhibiting the occurrence and development of TC. The average temperature of the upper ocean can, to a certain extent, indicate the maximum sea surface cooling that TC can cause. The effect of TC on the ocean can generally reach about 200m in the ocean, so the average temperature of the upper 200m of the ocean is introduced. When a TC encounters land, the intensity of the TC will change rapidly due to the huge friction of the land, so the LANDMASK of the TC underlying surface is also an important factor affecting the intensity of the TC.

[0037] Based on the consideration of the above physical mechanisms, physical variables are directly extracted from atmospheric and ocean reanalysis data and geographic data: U, V, T, VOR, hum2m, CAPE and To. The subscript o represents ocean variables, and the absence of subscript o represents atmospheric variables by default. After simple calculation, the various physical factors mentioned above can be obtained.

[0038] The guiding airflow is referred to as flow in the following text, the component in the x direction is called u_flow, and the component in the y direction is called v_flow. The vertical wind shear is referred to as vws, the x component is u_vws, and the y component is v_vws. The sea surface temperature is abbreviated as sst. Since TC exists in the atmosphere, when calculating the atmospheric background field variables, the variables within 500km of the TC center will be removed first, and only the values ​​within 500-1000km of the TC center will be counted to prevent the structural field of the TC itself from being mistakenly counted into the background field. Since TC only affects the ocean but does not exist in the ocean, this step is not performed. In addition, the humidification and heating of the ocean for TC directly affects the TC center, so when counting the ocean variable values, the sea temperature variable values ​​within 500km of the TC center are counted. Since the effect of landmask on TC also mainly acts on the TC center, when counting landmask, the landmask within 500km of the TC center is also counted.

[0039] The calculation diagram is as follows Figure 2As shown in the figure: (A) is the calculation of atmospheric background field variables. The figure shows an annular area with an inner radius of R=0-5° (about 500km) and an outer radius of R=5-10° (about 500-1000km). During the calculation, only the values ​​within 500-1000km of the TC center are counted (i.e., the blue annular area) to avoid mistakenly counting the structural field of the TC itself into the background field. The annular area is divided into four sectors (1, 2, 3, 4) for calculating the differences in variables in each direction. (B) is the calculation of ocean variables and geographic variables. The figure shows a circular area with a radius of R=0-5° (about 500km). During the calculation, the sea temperature variable values ​​and landmask within 500km of the TC center (i.e., the blue circular area) are counted. The circular area is also divided into four sectors (1, 2, 3, 4) for calculating the differences in variables in each direction. For both cases, Figure 2 The formulas for calculating variables are given in the table. Var-mean represents the average value of the four sectors, Var-xdiff calculates the difference in the x direction (east-west direction), and Var-ydiff calculates the difference in the y direction (north-south direction).

[0040] For the calculation of the variable mean field, the average value of the relevant variables within the statistical range of the TC center is used, and it is hereinafter represented by the variable name plus mean. For the calculation of the variable gradient field, the relevant variables within the statistical range of the TC center are used, and the right average value minus the left average value, or the upper average value minus the lower average value, are used. It is hereinafter represented by the variable name plus diff. Add subscript x in the latitude direction and subscript y in the longitude direction. Different layers in the vertical direction of the atmosphere will also have different effects on TC. The subscript high represents the average value of 300-500hPa, mid represents the average value of 500-700hPa, and low represents the average value of 700-850hPa. Vertical wind shear is the difference between 200hPa and 850hPa.

[0041] In addition, the current latitude of the TC is considered to be an important deterministic internal force of the TC, so in the consideration of input factors, the current longitude and latitude are directly selected as one of the model input variables. And considering the existence of TC inertia, the position and intensity change value pre of the TC 6 hours before, as well as the change trend value trend, are added. The calculation formula is as follows: X_pre=X(t)-X(t-6h) X_trend=[X(t)-X(t-6h)]-[X(t-6h)-X(t-12h)] When calculating Xpre and Xtrend of MWS and SLP, the SLP value is used uniformly for calculation, because in TCBest Track data, the quantitative value of MWS appears much later than the SLP value, and there will be a sudden change from 0 to a high value, and from a high value to 0, which is not in line with reality. When the maximum wind speed is 0m / s, the SLP value is already given, so SLP is used uniformly for calculation.

[0042] All the variables and all the calculation methods mentioned above are used to calculate 35 influencing factors. The influencing factors are shown in Input Factor in Table 1. Note that the input factors in Table 1 are not arranged according to the size of the correlation, but the factors with similar physical relationships are arranged together to facilitate the use of the subsequent AI model. Factor 1-6 represents the upper atmospheric information, Factor 7-12 represents the middle atmospheric information, Factor 13-15 represents the lower atmospheric information, Factor 16-20 represents the difference in the vertical or horizontal direction of the atmosphere, Factor 21-26 represents the sea-air interface information, and Factor 27-35 represents the current and past 6h position and intensity information of the TC. Table 2 shows the meaning of each variable name and abbreviation. Table 3 shows the linear correlation coefficient between the model input factors and output factors.

[0043] According to the physical relationship of these factors, they are divided into several categories, among which the time series factors are Factor 27-35, hereinafter referred to as Time Only; the guiding airflow and secondary guiding airflow factors are Factor (1-3) & (7-12), hereinafter referred to as Flow; the vortex correlation factors are Factor (4-6) & (13-15), hereinafter referred to as Vor; and the remaining factors. The factors that use all factors are called All Vars.

[0044] Table 1 Model input factors and output factors

[0045] Table 2 Variable names and abbreviations

[0046] Table 3 Linear correlation coefficients between model input factors and output factors **Indicates that the test passed the 1% significance level *Indicates that the 5% significance level test was passed No * indicates that the test failed at the 5% significance level.

[0047] Based on the model output factors and model input factors determined in steps SS3 and SS4, select at least one AI model from among convolutional neural network (CNN), recurrent neural network (RNN), long short-term memory network (LSTM), random forest (RF) and extreme gradient boosting (XGBoost) to build a TC path and intensity prediction model, and use the training set data in step SS2 to train the constructed TC path and intensity prediction model to capture the complex nonlinear relationship between meteorological and ocean factors; Preferably, in the above step SS5, for the CNN model, a combination of multiple convolutional layers and pooling layers is used, each convolution layer is followed by batch normalization and ReLU activation function, and finally the prediction result is output through the fully connected layer; for the RNN and LSTM models, a bidirectional network structure suitable for time series data is designed to capture the front and back dependencies of the time series; for the RF and XGBoost models, the complexity and generalization ability of the model are balanced by optimizing the hyperparameters including at least the depth and number of trees.

[0048] Use the validation set data in step SS2 to validate the TC path and intensity prediction model trained in step SS5, and evaluate the prediction performance of the model at different forecast time periods (e.g., 6 hours, 12 hours, and 24 hours). The evaluation indicators include at least the path forecast error and the intensity forecast error. According to the validation results, optimize the model by at least adjusting hyperparameters, increasing regularization, and / or using cross-validation methods to improve the prediction accuracy and generalization ability of the model.

[0049] Preferably, in the above steps SS5 and SS6, the prediction model training and optimization strategy includes: using batch normalization technology to improve the convergence speed and stability of the model; using dropout technology to prevent overfitting; applying a learning rate decay strategy to achieve more refined parameter adjustment in the later stage of training; and determining the optimal hyperparameter combination through k-fold cross validation.

[0050] More specifically, in the above steps SS5 and SS6, the present invention constructs four linear regression models (Model 1-4) to evaluate the relative contribution of different meteorological physical factors to the changes in tropical cyclone paths and intensities. These linear regression models are used to quantify the impact of each input factor in path and intensity prediction by capturing the linear relationship between the input factor and the prediction target, thereby providing a benchmark for the selection of input factors for the AI ​​model. In addition, Model 1-4, as a benchmark performance model, can quickly provide linear prediction results for path and intensity. These results provide an important reference for subsequent AI model performance evaluation, especially when capturing complex nonlinear relationships. The improvement of the AI ​​model compared to the linear regression model can be clearly reflected through performance comparison.

[0051] Models 1-4 are constructed using different combinations of input factors: Model 1 only uses the time series factor Time Only to evaluate the time evolution characteristics of path and intensity changes; Model 2 uses Time+Flow, and adds a guided airflow factor on the basis of Model 1 to evaluate the impact of atmospheric guided airflow on TC path; Model 3 uses Time+Flow+Vor, and adds a vorticity factor on the basis of Model 2 to analyze the effect of cyclone vorticity on path and intensity changes; Model 4 uses all factors All Vars to comprehensively analyze the common impact of different physical processes on TC path and intensity.

[0052] In order to further improve the accuracy of the forecast, the present invention also constructs 6 artificial intelligence models (Model 5-10), which are particularly suitable for scenarios where the TC path and intensity change drastically in the short term by capturing the complex nonlinear relationship between meteorological factors. The specific models and their functions are as follows: Model 5: Multi-Layer Perceptron (MLP) Multilayer Perceptron (MLP) is a classic feedforward neural network model, which consists of an input layer, one or more hidden layers, and an output layer. It can learn complex nonlinear mapping relationships and is suitable for a wide range of regression and classification tasks. The core of the MLP model lies in its hidden layer. By using nonlinear activation functions (such as ReLU), the model can capture the nonlinear features in the input data. The MLP model used in the present invention includes multiple layers of fully connected layers (nn.Linear), each of which is followed by a ReLU activation function (nn.ReLU). This structure helps the model capture the nonlinear features in the input data. The architecture of the model includes a series of fully connected layers, and the number of neurons in each layer increases and decreases in sequence, and finally outputs the required dimension. Through this structure, the model can fit complex data distributions to a certain extent.

[0053] Model 6: Convolutional Neural Network (CNN) Convolutional neural network (CNN) is a deep learning model specially designed to process data with grid structure, such as images and sequence data. It has shown excellent capabilities in fields such as computer vision and signal processing. The key components of the CNN model include convolutional layers, activation functions, pooling layers, and fully connected layers. The CNN model implemented in the present invention adopts a one-dimensional convolutional layer (nn.Conv1d), which is suitable for processing one-dimensional sequence data. The model contains three convolutional layers, each followed by a ReLU activation function (nn.ReLU) and a maximum pooling layer (nn.MaxPool1d). The convolutional layer is used to capture local features, while the pooling layer is used to reduce the spatial dimension of the features and improve computational efficiency. Finally, the model makes the final prediction by flattening the convolutional features and passing them to the fully connected layer (nn.Linear).

[0054] Model 7: Improved Recurrent Neural Network Model (RNN) The improved recurrent neural network model uses a long short-term memory unit (Long Short-Term Memory, LSTM) to process sequence data. LSTM can effectively capture long-term dependencies and is particularly useful for time series analysis tasks. The model in the present invention uses a multi-layer bidirectional LSTM (nn.LSTM) that can process input sequences in both forward and reverse directions to better utilize contextual information. In addition, the model also includes two fully connected layers (nn.Linear) for further processing the output of the LSTM. During the forward propagation process, the model takes the output of the last time step of the sequence and makes the final prediction through the fully connected layer.

[0055] Model 8: Improved bidirectional LSTM model The improved bidirectional LSTM model also uses long short-term memory units, but processes the input sequence in both the forward and reverse directions. Bidirectional LSTM can obtain information from both the beginning and the end of the sequence, which is essential for many natural language processing and time series prediction tasks. The model uses a multi-layer bidirectional LSTM (nn.LSTM) and makes the final prediction through two fully connected layers after the outputs in the forward and reverse directions are concatenated. This structure allows the model to make full use of the contextual information in the sequence data.

[0056] Model 9: Extreme Gradient Boosting Regressor (XGBoost Regressor) Extreme Gradient Boosting (XGBoost) is an efficient gradient boosting framework designed to quickly and accurately solve regression and classification problems. XGBoost gradually reduces the prediction error of training data by building multiple weak learners (usually decision trees). The XGBoost regressor used in the present invention builds a model by setting specific parameters. XGBoost is suitable for processing large-scale data sets, can handle missing values, and has good generalization ability. XGBoost supports parallel computing and can effectively use multi-core processors to accelerate the training process.

[0057] Model 10: Random Forest Regressor (RF) Random forest is an ensemble learning method that improves prediction performance by building multiple decision trees and averaging their results, which can reduce the risk of overfitting and obtain better results without too much parameter adjustment. The random forest regressor used in the present invention builds a model by setting the number of trees n_estimators=500 and the random seed random_state=42. Random forest is very suitable for processing data sets with high-dimensional feature spaces and can handle missing values ​​and non-numerical data.

[0058] Figure 3 The figure shows a comparison diagram of the forecast errors of different models for all TC paths and intensities in 2021-2022, where (a) is the forecast error of TC paths for 6-24 hours, and (b) is the forecast error of TC intensity for 6-24 hours. Through performance comparison analysis, the AI ​​model shows stronger adaptability when dealing with complex nonlinear relationships. In particular, Model 10 (Random Forest Regressor) performs outstandingly when dealing with multiple combinations of physical factors, and can effectively reduce the forecast errors of paths and intensity. In contrast, although the CNN and MLP models can extract complex spatial and nonlinear features, their performance is slightly inferior to the random forest model in some complex scenarios. In addition, Model 8 (improved bidirectional LSTM model) performs particularly well in long-term (such as 24 hours) path and intensity predictions, and is particularly suitable for dealing with scenarios where TC intensity changes dramatically.

[0059] Figure 4 and Figure 5 The results of the various models for the path and intensity forecast of TC No. 10 in 2021 are shown. It can be seen that the regression model has significantly improved after adding guiding airflow and ventilation flow to its path. Its intensity forecast results have significantly improved after adding factors related to intensity. This result is consistent with Figure 3The results are consistent with those in the previous section. For several AI models, there are several models with obvious result deviations, such as CNN and MLP. The results of other models do not show obvious intuitive differences.

[0060] Figure 6 and Figure 7 The best AI model RF predicts the path and intensity of all TCs from August to October 2021. It can be seen that this model performs well on most TCs, such as TC202109, 202110, 202112, 202116 and 202120. Its predicted path and intensity are consistent with the actual situation. When the TC path turns, most of them can be reported, and only a few cases have obvious errors, such as the turning points of 202111, 202114 and 202119. There are also some individual forecast results, such as TC202117 and TC202118, which have obvious forecast deviations.

[0061] Establish a real-time feedback mechanism to dynamically input new observed meteorological data into the optimized model in step SS6 after preprocessing, continuously adjust and optimize model parameters through online learning and / or regular retraining triggered by real-time changes in new data, and regularly evaluate model performance and update input factor sets and model structure to improve the timeliness and accuracy of TC track and intensity forecasts; Preferably, in the above step SS7, the implementation method of the real-time feedback mechanism includes: updating the real-time forecast at intervals of 6 hours, setting a sliding time window (such as the last 7 days) for online learning of the model, regularly (such as monthly) using accumulated new data to retrain the model to update model parameters, and dynamically adjusting the prediction weight of the model by comparing the deviation between the model prediction results and the actual observation results to maintain the model's adaptability to the latest meteorological patterns.

[0062] Generate and output the forecast results of TC path change (longitude and latitude) and intensity change (minimum air pressure and maximum wind speed) in the next 6 hours, 12 hours and 24 hours, calculate and output the forecast error, including but not limited to path forecast error and intensity forecast error, provide credibility assessment of the forecast results, and provide decision support information for meteorological departments and disaster prevention and mitigation agencies.

[0063] Preferably, in the above steps SS7 and SS8, when the prediction error exceeds a preset threshold, the model retraining process is triggered; it also includes an automatic identification and manual review mechanism for abnormal prediction results, and makes necessary adjustments to the model output in combination with the experience of meteorological experts, and generates forecast products in a standard format to ensure the reliability and practicality of the prediction results.

[0064] Through the above embodiments, the purpose of the present invention is fully and effectively achieved. Those skilled in the art can understand that the present invention includes but is not limited to the contents described in the drawings and the above specific embodiments. Although the present invention has been described with respect to the most practical and preferred embodiments currently considered, it should be understood that the present invention is not limited to the disclosed embodiments, and any modification that does not deviate from the functional and structural principles of the present invention will be included in the scope of the claims.

Claims

1. A TC path and intensity prediction method based on AI and physical properties, characterized in that: The prediction method comprises at least the following steps when implemented: SS1. Data Collection Obtain at least five consecutive years of historical meteorological data from multiple meteorological and oceanographic data sources, including at least TCBest Track data, FNL atmospheric reanalysis data, and GODAS ocean reanalysis data; SS2. Data preprocessing Based on the historical meteorological data and ocean data collected in step SS1, extract the relevant data of each TC occurrence period in the target prediction area and perform data quality control on it, including identifying and eliminating missing or abnormal data, standardizing and normalizing the data to ensure the temporal and spatial consistency between different data sources, and then divide the processed data set into a training set and a validation set, and ensure that the training data and the validation data are independent in terms of temporal and spatial characteristics; SS3. Selection and calculation of model output factors Based on the historical TC Best Track data preprocessed in step SS2 and the characteristics of the target forecast area, multiple variables closely related to TC path and intensity changes are selected as model output factors, and the model output factors at least include TC position changes and intensity changes. The time series of these variables are calculated to extract characteristic changes and establish model output targets; SS4. Selection and calculation of model input factors Based on the physical mechanism of TC movement and intensity change, combined with the FNL atmospheric reanalysis data and GODAS ocean reanalysis data preprocessed in step SS2, multiple physical factors related to TC movement and intensity change are selected and calculated. Then, the correlation analysis between the selected physical factors and TC path and intensity change is carried out, and the correlation coefficient between each physical factor and the model output factor is calculated one by one. Finally, the model input factor set is determined based on the analysis results. SS5. Construction and training of TC track and intensity prediction model Based on the model output factors and model input factors determined in steps SS3 and SS4, at least one AI model from CNN, RNN, MLP, LSTM, RF and XGBoost is selected to build a TC path and intensity prediction model, and the training set data in step SS2 is used to train the constructed TC path and intensity prediction model to capture the complex nonlinear relationship between meteorological factors; SS6. Verification and optimization of TC path and intensity prediction model Using the validation set data in step SS2 to validate the TC path and intensity prediction model trained in step SS5, evaluating the prediction performance of the model at different forecast time periods, where the evaluation indicators include at least path prediction error and intensity prediction error, and optimizing the model by at least adjusting hyperparameters, increasing regularization, and / or using a cross-validation method based on the validation results; SS7. Real-time forecast and dynamic updates Establish a real-time feedback mechanism to dynamically input new observed meteorological data into the optimized model in step SS6 after preprocessing, continuously adjust and optimize model parameters through online learning and / or regular retraining triggered by real-time changes in new data, and regularly evaluate model performance and update input factor sets and model structure to improve the timeliness and accuracy of TC track and intensity forecasts; SS8. Output of prediction results Generate and output the prediction results of TC path change and intensity change in the next 6 hours, 12 hours and 24 hours, calculate and output the prediction error, and provide a credibility assessment of the prediction results.

2. The TC path and intensity prediction method based on AI and physical characteristics according to claim 1, characterized in that: In the above steps SS1 and SS2, information on the period of TC occurrence in the northwest Pacific from January 1, 2015 to December 31, 2022 is obtained from multiple meteorological and oceanographic data sources. For each TC case, the atmospheric and ocean background field information in the report is statistically reported every 6 hours, and missing data records are removed. The data from 2015 to 2020 are used for model training, and the data from 2021 to 2022 are used for model verification.

3. The TC path and intensity prediction method based on AI and physical characteristics according to claim 1, characterized in that: In the above step SS2, when performing data quality control, the missing parts in the data are processed by using mean filling, interpolation or deleting records with missing values. The outliers in the data are identified and removed through statistical analysis methods such as Z-score or box plots. Z-score standardization or Min-Max normalization is used to process data of different scales. The collected historical meteorological data are synchronized in time to ensure that the time resolution of all data sources is consistent.

4. The TC path and intensity prediction method based on AI and physical characteristics according to claim 3 is characterized in that: In the above step SS2, data quality control also uses an automated data cleaning method based on machine learning to identify potential abnormal data through historical data learning, and automatically perform data correction or filling to improve the efficiency and accuracy of data preprocessing.

5. The TC path and intensity prediction method based on AI and physical characteristics according to claim 1, characterized in that: In the above step SS3, each output factor is calculated using X_n=X(t+n)-X(t) to calculate the hourly position and intensity changes of TC, where X represents the longitude LON, latitude LAT, central minimum pressure SLP and central maximum wind speed MWS, t represents the current time, and n is 6h, 12h and 24h, respectively, forming a total of 12 model output factors.

6. The TC path and intensity prediction method based on AI and physical characteristics according to claim 1, characterized in that: In the above step SS4, the multiple physical factors selected that are related to the TC motion and intensity change should at least include guiding airflows at the middle and upper levels, Beta drift force and ventilation flow, airflow vorticity at the middle and upper levels, vertical wind shear at high and low altitudes, lower atmospheric temperature, lower atmospheric humidity, diagnostic CAPE value, sea surface temperature SST, average temperature of the upper 200 m of the ocean, and LANDMASK of the TC underlying surface.

7. The TC path and intensity prediction method based on AI and physical characteristics according to claim 6, characterized in that: The calculation of Beta drift force and ventilation flow includes at least the following sub-steps: SS41. Extract the U and V wind field data of the three pressure layers of 500hPa, 600hPa and 700hPa within the range of 0-500km around TC from the FNL atmospheric reanalysis data, where U is the wind speed component in the x direction and V is the wind speed component in the y direction. The U and V wind fields of the three extracted pressure layers are vertically averaged to obtain the average wind field representing the middle atmosphere. SS42. Calculate the moving speed of TC using the current position and the position of TC in the previous 6 hours after obtaining the vertically averaged U and V wind field data, and subtract this speed from the averaged U and V wind fields to obtain the environmental wind field relative to TC, so as to eliminate the error caused by the movement of TC itself; SS43. The wind field within a radius of 500 km from the TC center is selected as the calculation area and decomposed into tangential wind and radial wind, where the tangential wind represents the wind speed component along the TC rotation direction, and the radial wind represents the wind speed component perpendicular to the tangential direction; SS44. Perform Fourier decomposition on radial wind and tangential wind respectively, extract the first-order wave component to describe the main asymmetric component in the wind field, and transform the first-order wave component obtained after decomposition into U and V wind speed components in Cartesian coordinate system through inverse Fourier transform; SS45. The back-calculated U and V wind speed components are averaged over the entire calculation area to obtain the magnitude and direction of the Beta drift force and ventilation flow, where the Beta drift force represents the deviation of the TC relative to the atmospheric environment guide airflow, and the ventilation flow represents the influence of the asymmetric secondary airflow.

8. The TC path and intensity prediction method based on AI and physical characteristics according to claim 1, characterized in that: In the above step SS4, when determining the final model input factors, at least the following steps are included: First, the correlation coefficient between each physical factor and the model output factor is calculated, and the correlation coefficient is tested for significance level. According to the preset significance level, the correlation between each physical factor is evaluated to see whether it is statistically significant. Physical factors with statistically significant correlation are screened out, and factors that fail the significance level test are excluded. Secondly, multiple linear regression was performed on the factors that passed the significance test to evaluate the relative contribution of each physical factor to the changes in TC paths and intensities, so as to further screen out the input factors with higher importance. Afterwards, principal component analysis was performed on the physical factors that passed the significance test and preliminary regression analysis, and the principal components whose cumulative explained variance reached the preset threshold were extracted by the PCA method to reduce the data dimension and eliminate the remaining multicollinearity. Finally, the results of correlation analysis, significance level test, regression analysis and principal component analysis are comprehensively considered to finally determine the input factor set of the model to ensure the prediction accuracy and stability of the model.

9. The TC path and intensity prediction method based on AI and physical characteristics according to claim 1, characterized in that: In the above step SS5, for the CNN model, a combination of multiple convolutional layers and pooling layers is used. Each convolution layer is followed by batch normalization and ReLU activation function, and finally the prediction result is output through the fully connected layer. For the RNN and LSTM models, a bidirectional network structure suitable for time series data is designed to capture the dependencies between the time series. For RF and XGBoost models, the complexity and generalization ability of the model are balanced by optimizing hyperparameters including the depth and number of trees.

10. The TC path and intensity prediction method based on AI and physical characteristics according to claim 1, characterized in that: In the above steps SS5 and SS6, the prediction model training and optimization strategies include: using batch normalization technology to improve the convergence speed and stability of the model; using dropout technology to prevent overfitting; applying a learning rate decay strategy to achieve more refined parameter adjustments in the later stages of training; and determining the optimal hyperparameter combination through k-fold cross validation.

11. The TC path and intensity prediction method based on AI and physical characteristics according to claim 1, characterized in that: In the above step SS7, the implementation method of the real-time feedback mechanism includes: updating the real-time forecast at intervals of 6 hours, setting a sliding time window for online learning of the model, regularly using accumulated new data to retrain the model to update the model parameters, and dynamically adjusting the prediction weight of the model by comparing the deviation between the model prediction results and the actual observation results to maintain the model's adaptability to the latest meteorological patterns.

12. The TC path and intensity prediction method based on AI and physical characteristics according to claim 1, characterized in that: In the above steps SS7 and SS8, when the prediction error exceeds the preset threshold, the model retraining process is triggered; it also includes an automatic identification and manual review mechanism for abnormal prediction results, and makes necessary adjustments to the model output based on the experience of meteorological experts, and generates forecast products in a standard format to ensure the reliability and practicality of the prediction results.

13. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the TC path and intensity prediction method based on AI and physical properties described in any one of claims 1 to 12 are implemented.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the TC path and intensity prediction method based on AI and physical properties described in any one of claims 1 to 12 are implemented.

Images