Rapid calculation method and system for significant wave height field of tropical cyclones in open sea area
By constructing a spatiotemporal fusion neural network model and fine-tuning it using satellite altimeter and buoy observations, the problems of high computational resource consumption and insufficient forecast accuracy in tropical cyclone wave prediction were solved, achieving fast and accurate wave field prediction and improving the model's efficiency and adaptability.
Patent Information
- Application Number
- CN202511650440.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-12
- Publication Date
- 2026-03-06
AI Technical Summary
Existing technologies in tropical cyclone wave prediction have high computational resource consumption and long running time, making it difficult to meet the requirements of rapid response and high-frequency calls. Furthermore, traditional empirical parameterized models have insufficient forecast accuracy under strong wind disturbance conditions and fail to effectively embed the nonlinear mechanism of the wave system.
By employing a spatiotemporal fusion neural network model, the effective wave height field of tropical cyclones is rapidly calculated through the construction of a priori cyclone sample set, dynamic wind field structure embedding, coordinate system transformation, and neural network training. Fine-tuning is then performed using satellite altimeter and buoy observations to achieve high-precision and low-time-consuming predictions.
It achieves rapid and accurate prediction of tropical cyclone wave fields, improves the quality and efficiency of the model, can complete wave field distribution prediction within seconds, reduces computational resource consumption and time costs, and enhances the reliability and adaptability of the prediction.
Smart Images
Figure CN121615460A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of oceanography, and more particularly to a rapid method for calculating the effective wave height field of tropical cyclones in open sea areas. Background Technology
[0002] With increasingly significant global climate change, the frequency and intensity of tropical cyclones (such as typhoons and hurricanes) are showing a clear upward trend. During their development, tropical cyclones are accompanied by extreme strong winds and severe sea states, especially the generation of large-scale high-energy wave systems, which seriously threaten marine engineering facilities, coastal infrastructure, and personnel safety. Therefore, how to quickly and accurately predict the wave characteristics during tropical cyclones has become one of the research hotspots and key technical issues in the fields of marine engineering, disaster early warning, and disaster prevention and mitigation.
[0003] Currently, the simulation and forecasting of tropical cyclone waves mainly rely on third-generation wave numerical models based on the principle of spectral energy conservation. These models are driven by wind pressure forcing as input and coupled with multiple process terms including nonlinear energy transfer, wave breaking (whitening), bottom friction, and non-uniform topographic feedback. They achieve fine reconstruction of the dynamic evolution of the wave field by solving a system of high-dimensional nonlinear partial differential equations. Although these models have strong physical description capabilities and simulation accuracy, in practical applications, their computational resource consumption is significant and their running time is long, which limits their application in rapid response scenarios such as tropical cyclone emergencies. They also cannot meet the high-frequency call requirements for large-sample long-term climate return analysis and extreme statistical scenarios.
[0004] On the other hand, although traditional empirical parameterized models have high operating efficiency, their modeling logic is often based on linear assumptions between parameters and local statistical regression. They fail to effectively embed the highly nonlinear wave-wave interaction mechanism in wave systems, and do not fully consider the dynamic reconstruction effect of cyclone movement speed, path structure, etc. on equivalent wind area configuration and energy input pattern. Therefore, the forecast accuracy is limited under strong wind disturbance conditions, and there is a significant risk of deviation. Summary of the Invention
[0005] The main objective of this invention is to provide a method and system for rapidly calculating the effective wave height field of tropical cyclones in open sea areas.
[0006] The technical solution adopted in this invention is: A rapid method for calculating the significant wave height field of tropical cyclones in open sea areas is provided, including the following steps: By embedding the dynamic wind field structure into the information of the tropical cyclone event to be measured, a spatial-temporal distribution field is constructed, and key cyclone parameters and relative position coordinates for a preset time period containing the current moment are extracted to form a time series matrix. By inputting the time series matrix into the spatiotemporal fusion neural network model, the spatiotemporal distribution of effective wave height in the open sea area at the target time can be quickly calculated. The construction process of the spatiotemporal fusion neural network model is as follows: Obtain the existing tropical cyclone database and perform systematic perturbation enhancement processing to construct a priori sample set of cyclones; By embedding the dynamic wind field structure into the cyclone prior sample set, a spatial-temporal distribution field of tropical cyclone wind field is constructed, forming a tropical cyclone wind field enhancement dataset. Based on the tropical cyclone wind field enhancement dataset, the third-generation spectral numerical wave model was called to perform high spatiotemporal resolution wave numerical simulations of all tropical cyclone events in open sea areas. The effective wave height response field in the simulation results was extracted to obtain the effective wave height dataset of tropical cyclones in open sea areas. A dynamic covariant coordinate system with the cyclone motion path as the main axis is constructed. The coordinate system transformation and scale normalization are performed on the tropical cyclone significant wave height dataset to obtain a spatially structure-normalized tropical cyclone significant wave height dataset. Extract key cyclone parameters and relative position coordinates of all tropical cyclone events within a preset time period containing the target time from the tropical cyclone wind field enhancement dataset, construct a time series matrix and use it as input, and use the effective wave height at the target location at the target time in the spatially structure-normalized tropical cyclone effective wave height dataset as output to train a pre-constructed spatiotemporal fusion neural network model.
[0007] Following the above technical solution, the spatiotemporal fusion neural network model specifically involves: using a multi-head attention mechanism to identify the cross-correlation between different parameter time series in the time series matrix, the dynamic trends within the parameter time series, and the local jumps and nonlinear perturbation structures; introducing jump connections and feature map splicing for residual learning during feature reconstruction; and using a fully connected neural network as a nonlinear regression mapper to generate a direct estimate of the target effective wave height. The model training objective is to minimize the loss function between the model output and the actual wave height response value in the tropical cyclone effective wave height dataset after spatial structure normalization.
[0008] Following the above technical solution, the training of the spatiotemporal fusion neural network model also includes fine-tuning training, specifically: using the effective wave height information of satellite altimeter and buoy observations in the tropical cyclone database to construct an observation calibration sample set, and then using this observation calibration sample set to fine-tune the spatiotemporal fusion neural network model.
[0009] Following the above technical solution, the specific process for constructing the observation calibration sample set is as follows: It traverses all cyclone events in the tropical cyclone database and retrieves effective wave height data from satellite altimeter and buoy observations within a preset range during the life cycle of each cyclone event. Spatiotemporal registration is performed on the effective wave height data within the preset range, and a fine-tuned sample vector is constructed. The information in each sample includes: the spatial normalized coordinates of the observation point at the current time and the observed effective wave height, the key parameters of the cyclone at the synchronization time, and the time series consisting of the key parameters of the cyclone in the past preset time period and the corresponding cyclone center coordinates. Iterate through all fine-tuned sample vectors that meet the conditions to construct an observation calibration sample set.
[0010] Following the above technical solution, the construction process of the cyclone prior sample set is as follows: For any given tropical cyclone event in the existing tropical cyclone database, based on its original center trajectory, a certain range of discrete translation operations are performed in the meridional and latitudinal dimensions to generate multiple sets of parallel artificial trajectories with small offset characteristics; a multi-dimensional perturbation mechanism is introduced into the time series of key cyclone parameters corresponding to each original and offset trajectory, and a constrained Monte Carlo sampling strategy is adopted to perform multiple rounds of random sampling within the preset range of the original values of each key cyclone parameter, and a smoothing operation is performed on the time series of key cyclone parameters after each sampling to generate a perturbation sample set as the cyclone prior sample set.
[0011] Following the above technical solution, the formation process of the tropical cyclone wind field enhancement dataset is as follows: on the basis of the static centrosymmetric wind field structure, a superposition correction mechanism of the translational velocity vector field is introduced to generate a dynamic wind field structure with asymmetric perturbation characteristics, and the dynamic wind field structure is embedded into the cyclone prior sample set to construct the spatial-temporal distribution field of the tropical cyclone wind field, thus forming the tropical cyclone wind field enhancement dataset.
[0012] Following the above technical solution, the specific process of spatial structure normalization of the tropical cyclone significant wave height dataset is as follows: For each tropical cyclone event in the significant wave height data of tropical cyclones, a dynamic covariant coordinate system is constructed with the cyclone center as the origin, the direction of cyclone movement as the Y-axis, and the direction of the Coriolis force orthogonal to the Y-axis as the X-axis. Project the relative distances between all locations in the tropical cyclone significant wave height data set and the cyclone center onto the original latitude and longitude plane. Δx and Δy Transform to this covariant coordinate system and perform dimensionless normalization on its spatial scale: x = Δx / R max , y = Δy / R max in, x The x-coordinate in the dynamic covariant coordinate system,y The ordinate is the ordinate in the dynamic covariant coordinate system. R max The radius is the radius of maximum wind speed.
[0013] Following the above technical solution, the key parameters of a cyclone include the pressure gradient. ΔP Cyclone propulsion speed V t Maximum wind speed radius R m Maximum wind speed V m And the radius of wind speed of 34 knots R 34 The preset time period is 72 hours.
[0014] Following the above technical solution, the existing tropical cyclone database is the Global Tropical Cyclone Multi-Source Reanalysis Data Set IBTrACS.
[0015] This invention also provides a rapid calculation system for the significant wave height field of tropical cyclones in open sea areas, comprising: The data acquisition module is used to acquire information about the tropical cyclone event to be measured. The data processing module is used to embed the dynamic wind field structure into the information of the tropical cyclone event to be measured to construct a spatial-temporal distribution field, extract key cyclone parameters and relative position coordinates, and select data containing the current time period to construct a time series matrix. The fast calculation module is used to input the time series matrix into a pre-built spatiotemporal fusion neural network model to quickly calculate the spatiotemporal distribution of effective wave height in the open sea area at the target time. The model building module is used to build the spatiotemporal fusion neural network model using the construction method of the above technical solution.
[0016] The beneficial effects of this invention are as follows: This invention realizes the use of neural networks to replace traditional numerical wave models for effective wave height prediction, realizing the transformation from traditional numerical simulation with "low precision and high time consumption" to intelligent modeling with "high precision and low time consumption", which greatly improves the quality and efficiency of tropical cyclone wave field prediction and modeling; by improving the ability to predict extreme tropical cyclone waves, it helps to deploy emergency measures in advance and reduce the risks of offshore operations and coastal economic losses.
[0017] Furthermore, this invention deconstructs the wave generation mechanism dominated by tropical cyclones to determine the underlying mechanisms. Δ P , V t , R m , V m , R34 Five structural parameters serve as core perturbation factors, and are further extended to their time-series trajectories over the past 72 hours, thereby enabling dynamic capture of cyclone intensity evolution, scale structure changes, and historical energy injection processes. These parameters essentially constitute the complete input domain that determines the wave response prediction model, effectively ensuring the physical consistency and structural integrity of the model's input layer.
[0018] Furthermore, a spatial structure normalized coordinate system with the cyclone's direction of travel as the main axis is introduced. By transforming the target grid point position vector to a dimensionless coordinate space, the structural bias caused by cyclone-scale heterogeneity and path differences on the spatial distribution of samples is significantly reduced. This ensures that the samples have higher spatial consistency and statistical stability during training, and provides a generalizable spatial mapping reference system for subsequent deep neural network modeling.
[0019] Furthermore, in order to eliminate the complex modulation effect of land topography on the wind field structure and wave field response characteristics of tropical cyclones, this invention implements an "open sea area constraint strategy" in sample construction and model applicability scope. That is, the research object is limited to oceanic tropical cyclone events that are not affected by land boundary conditions, thus avoiding the disturbance deviation caused by nonlinear boundary effects such as shoreline irregularity, topographic roughness and nearshore tidal-wave coupling from the source.
[0020] Furthermore, the artificial neural network employed incorporates a dual structure of multi-head attention and multi-head self-attention mechanisms to simultaneously capture the intervariate nonlinear correlations between different cyclone disturbance parameters and the temporal dynamic correlation structure within the same parameter. By weighting and focusing the information flow through parallel attention heads, and further combining the nonlinear mapping capability of fully connected networks, a deep functional approximation of the relationship between multi-source historical disturbance parameter sequences and space wave height responses is achieved.
[0021] Furthermore, in the model fine-tuning stage, this invention introduces a freezing strategy and a small amount of observation data migration mechanism: keeping the parameters of each layer of the neural network except the output layer stable, and by loading the existing coarse model parameters and training only the output layer weights, the model can achieve adaptive fine-tuning of the distribution of new samples and the characteristics of observation errors with the support of limited satellite altimeter and buoy observation samples, effectively improving the model's generalization ability and prediction robustness in actual tropical cyclone events.
[0022] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description
[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0024] Figure 1 This is a flowchart of a rapid calculation method for the effective wave height field of tropical cyclones in open sea areas according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating the construction process of the spatiotemporal fusion neural network model according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the spatiotemporal fusion neural network model according to an embodiment of the present invention. Detailed Implementation
[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0026] It should be noted that the illustrations provided in the embodiments of the present invention are only schematic representations of the basic concept of the present invention. Therefore, the illustrations only show the components related to the present invention and are not drawn according to the number, shape and size of the components in actual implementation. In actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.
[0027] In this invention, it should also be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application. Furthermore, the terms "first" and "second" are used only for descriptive and distinguishing purposes and should not be construed as indicating or implying relative importance.
[0028] Furthermore, it should be noted that the features of the various embodiments of the present invention can be combined or integrated in whole or in part, and as those skilled in the art will understand, they can interact and operate in different ways. Each embodiment can be implemented independently of each other or in association with one another.
[0029] This invention constructs a highly efficient, accurate, and adaptively updated statistical model to predict the wave distribution characteristics of various grid points during tropical cyclones, improving the model's generalization performance and prediction accuracy in practical applications. It can rapidly predict the complete wave field distribution under a given combination of tropical cyclone parameters within seconds, significantly reducing computational resource consumption and time costs. Furthermore, the model not only possesses data-driven advantages but also exhibits a certain degree of physical constraints and observational consistency, enhancing the model's credibility in actual predictions.
[0030] like Figure 1 As shown in the embodiment of the present invention, a rapid calculation method for the significant wave height field of a tropical cyclone in open sea includes the following steps: S101. Embed the dynamic wind field structure into the information of the tropical cyclone event to be measured to construct a spatial-temporal distribution field; S102. Extract key cyclone parameters and relative position coordinates for a preset time period containing the current moment from the spatial-temporal distribution field to form a time series matrix; S103. Input the time series matrix into the spatiotemporal fusion neural network model to quickly calculate the spatiotemporal distribution of effective wave height in the open sea area at the target time.
[0031] The construction process of the spatiotemporal fusion neural network model is as follows: Figure 2 As shown, specifically: S201. Obtain the existing tropical cyclone database and perform systematic perturbation enhancement processing to construct a priori sample set of cyclones; S202. Embed the dynamic wind field structure into the cyclone prior sample set to construct the spatial-temporal distribution field of the tropical cyclone wind field and form a tropical cyclone wind field enhancement dataset. S203. Based on the tropical cyclone wind field enhancement dataset, the third-generation spectral numerical wave model is called to perform high spatiotemporal resolution wave numerical simulations of all tropical cyclone events in open sea areas, and the effective wave height response field in the simulation results is extracted to obtain the effective wave height dataset of tropical cyclones in open sea areas. S204. Construct a dynamic covariant coordinate system with the cyclone motion path as the main axis, and perform coordinate system transformation and scale normalization on the tropical cyclone significant wave height dataset to obtain a spatially structure-normalized tropical cyclone significant wave height dataset. S205. Extract the key cyclone parameters and relative position coordinates of all tropical cyclone events within a preset time period containing the target time from the tropical cyclone wind field enhancement dataset, construct a time series matrix and use it as input, and use the effective wave height at the target position at the target time in the spatially structure-normalized tropical cyclone effective wave height dataset as output to train the pre-constructed spatiotemporal fusion neural network model.
[0032] Through the above method, this invention achieves the use of neural networks to replace traditional numerical wave models for effective wave height prediction, realizing a transformation from traditional numerical simulation that is "low-precision and time-consuming" to intelligent modeling that is "high-precision and time-consuming," significantly improving the quality and efficiency of tropical cyclone wave field prediction and modeling. By enhancing the ability to predict extreme tropical cyclone waves, it helps to deploy emergency measures in advance, reducing the risks of offshore operations and coastal economic losses.
[0033] Specifically, in a preferred embodiment of the present invention, the construction and training process of the spatiotemporal fusion neural network model is as follows: Step 1: First, systematic perturbation enhancement processing is performed on the raw cyclone records in the global tropical cyclone multi-source reanalysis dataset IBTrACS to construct a priori cyclone sample set with higher parameter coverage and spatial diversity: 1. For any given tropical cyclone event, based on its central trajectory (latitude and longitude time series), a discrete translation operation within a range of ±3° can be performed in both the longitude and zonation dimensions. The Cartesian products of the longitude offsets {-3°, 0°, +3°} and the zonation offsets {-3°, 0°, +3°} are combined to generate a total of 9 parallel artificial trajectories with small-amplitude shifts. This operation aims to simulate the path uncertainty disturbance characteristics of tropical cyclones under similar meteorological disturbances.
[0034] 2. Based on each of the original and offset trajectories mentioned above, analyze the five core cyclone parameters during their temporal evolution, including pressure gradients. ΔP Cyclone propulsion speed V t Maximum wind speed radius R m Maximum wind speed V m and a wind speed radius of 34 knots R 34 A multi-dimensional perturbation mechanism is introduced. Specifically, a constrained Monte Carlo sampling strategy is adopted, and multiple rounds of random sampling are performed within ±20% of the original values of each parameter to generate a perturbation sample set. After each sampling, a smoothing operation is performed on the parameter time series to avoid non-physical abrupt changes from interfering with the stability of subsequent modeling.
[0035] 3. Perform the above trajectory translation and parameter perturbation process sequentially on all tropical cyclone events in the IBTRASC dataset to finally form an extended cyclone parameter enhancement dataset with dual characteristics of spatial perturbation and dynamic perturbation.
[0036] Step 2: Based on the improved Holland cyclone structure parameterization formula, combined with the key parameters of the above extended cyclone parameter enhancement dataset (i.e. ΔP , Vt , R m , V m , R 34 A wind field structure dataset is constructed using a parameter set. Specifically, based on the traditional static centrosymmetric wind field structure, a superposition correction mechanism of the translational velocity vector field is introduced to generate a dynamic wind field structure with asymmetric perturbation characteristics. This structure is embedded into the wind field in-analysis dataset to construct the spatial-temporal distribution field of the tropical cyclone wind field, i.e., the tropical cyclone wind field enhancement dataset.
[0037] Step 3: Based on the wind field driving data mentioned above, call the third-generation spectral numerical wave model to analyze all tropical cyclone events in open sea areas (e.g., defined as 10x). R m High spatiotemporal resolution wave numerical simulations were conducted in an area with no nearby land. The significant wave height response field was extracted from the simulation results to obtain a dataset of significant wave heights of tropical cyclones in open sea areas, where each sample represents the spatiotemporal distribution of wave heights at different development stages of a specific cyclone.
[0038] It is evident that, in order to eliminate the complex modulation effect of land topography on the wind field structure and wave field response characteristics of tropical cyclones, this invention implements an "open sea area constraint strategy" in sample construction and model applicability scope. That is, the research object is limited to oceanic tropical cyclone events that are not affected by land boundary conditions, thus avoiding the disturbance deviation caused by nonlinear boundary effects such as shoreline irregularity, topographic roughness, and nearshore tidal-wave coupling from the source.
[0039] Step 4: Perform a unified reference frame transformation and scale normalization on all valid wave height fields in the extended sample set to construct a dynamic covariant coordinate system with the cyclone's motion path as the main axis (the cyclone center is the origin, and the cyclone's direction of movement is the y-axis), in order to eliminate spatial structure disturbances caused by differences in path, scale, and direction. Specifically, this includes: 1. For each tropical cyclone event in the sample, a local reference frame is first constructed based on its translation vector (i.e., the direction of cyclone travel), where the direction of cyclone travel is defined as the Y-axis of the normalized coordinate system, and the direction of the Coriolis force, which is orthogonal to it, is defined as the X-axis. 2. The relative distances between all locations in the wave height data and the cyclone center can be projected onto the original latitude and longitude plane. Δx and Δy Transform to this rotating coordinate system and dimensionlessly convert its spatial scale: x = Δx / R max , y = Δy / R max This step yielded a tropical cyclone effective wave height dataset with normalized spatial structure, achieving uniformity of sample spatial structure, effectively removing the influence of cyclone trajectory deviation on wave field distribution patterns, providing a more stable and physically consistent input basis for subsequent model training, and effectively solving the problem of non-uniform data distribution caused by different cyclone scales.
[0040] Furthermore, a spatial structure normalized coordinate system with the cyclone's direction of travel as the main axis is introduced. By transforming the target grid point position vector to a dimensionless coordinate space, the structural bias caused by cyclone-scale heterogeneity and path differences in the spatial distribution of samples is significantly reduced. This ensures that the samples have higher spatial consistency and statistical stability during training, and provides a generalizable spatial mapping reference system for subsequent deep neural network modeling.
[0041] Step 5: Based on the physical mechanism of tropical cyclone wind wave-swell generation and evolution, it is known that the significant wave height at a given time and spatial location... Hs ( t0, x, y This originates from the energy superposition and propagation interference between the newly formed wind-wave system driven by the current cyclone wind field and the long-range surge system generated by the residual excitation of the cyclone in the preceding period (typically chosen to be about 72 hours). Therefore, this invention assumes in the modeling stage that the effective wave height response at the target location at the target time is... Hs ( t0, x, y This can be considered a multivariable parameter of the cyclone, including parameters at different times. ΔP , V t , R m , V m , R 34 Five parameters of tropical cyclones and their corresponding coordinates of the relative positions of the tropical cyclone centers at the corresponding times. x’, y’ The resulting seven-dimensional dynamic input tensor forms a high-order smooth function over the past 72 hours.
[0042] To effectively model this multivariable, multi-delay, and multi-scale coupling mechanism, this invention designs a spatiotemporal fusion neural network model with global perception and temporal dynamic interaction modeling capabilities. For example... Figure 3 As shown, the model adopts the following structural flow: 1. Construct a 7×72 time series matrix from the parameters of each tropical cyclone disturbance and their relative coordinates (a total of 7 variables) over the past 72 hours (including the current time), and perform standardization preprocessing on each parameter dimension to suppress the interference of scale differences on the training process; 2. First, a multi-head attention mechanism is used to model the cross-correlation between different parameter sequences to identify the heterogeneous synergistic effect among cyclone parameters. Then, a multi-head self-attention mechanism is used to mine the dynamic trend, local jumps and nonlinear perturbation structure within the time series of each parameter. The two types of attention mechanisms are used in the encoder stage to model the structure of the multivariate dynamic input tensor in a coordinated manner.
[0043] 3. A decoder module with a symmetrical structure is adopted. Skip connections and feature map splicing are introduced during the feature reconstruction process to construct a residual-like learning path, thereby enhancing the gradient propagation efficiency and representation stability of deep networks.
[0044] 4. After decoding, a fully connected neural network layer is embedded as a nonlinear regression mapper to generate a direct estimate of the target's effective wave height. The training objective is set as: minimizing the model output's correlation with the actual wave height of the spatially structure-normalized tropical cyclone dataset. Hs The loss function between response values.
[0045] The model trained through this process constitutes a coarse model for rapid calculation of the effective wave height field of tropical cyclones in open sea areas. It has the ability to quickly estimate the space wave field under the drive of complex cyclone disturbances, and provides stable initialization parameters for subsequent fine-tuning and generalization stages.
[0046] As can be seen, the artificial neural network employed incorporates a dual structure of multi-head attention and multi-head self-attention mechanisms to simultaneously capture the intervariate nonlinear correlations between different cyclone disturbance parameters and the temporal dynamic correlation structure within the same parameter. By using parallel attention heads to weight and focus the information flow, and further combining this with the nonlinear mapping capability of a fully connected network, a deep functional approximation of the relationship between multi-source historical disturbance parameter sequences and space wave height response is achieved.
[0047] Furthermore, it should be noted that the artificial neural network used does not necessarily have to be the network structure of the above embodiment. It can also be a neural network such as CNN, LSTM, or Transformer, as long as it can be trained to make predictions.
[0048] Step 6: Due to the uncertainties in the structure of tropical cyclone wind fields and the errors caused by the uncertainties in numerical and physical processes in numerical wave models, the tropical cyclone wind field enhancement dataset obtained from the Holland wind field and numerical wave models, although rich in samples, suffers from accumulated biases in accuracy, making it difficult to directly meet the training requirements of high-precision models. To recalibrate the coarse model trained on this dataset in Step 5 in terms of physical consistency and statistical performance, an observation calibration sample set for fine-tuning training is constructed using effective wave height information from satellite altimeters and buoy observations: 1. Traverse all cyclone events in the IBTrACS tropical cyclone database, defining their time-series trajectories and spatially affected areas. Within the lifecycle of each cyclone event, retrieve... x 2 + y 2 Effective wave height data from satellite altimeter and buoy observations within a radius of <100.
[0049] 2. Perform spatiotemporal registration on the above observation data and construct a fine-tuned sample vector. Each sample contains the following information: the spatially normalized coordinates of the observation point at the current time. x , y With the observed effective wave height Hs Five parameters of a tropical cyclone at synchronous time ΔP , V t , R m , V m , R 34 Data, the past 72 hours ΔP , V t , R m , V m , R 34 Data and corresponding cyclone center coordinates x’、 y’ The time series formed 。 This invention deconstructs the wave generation mechanism dominated by tropical cyclones to determine the underlying mechanisms. ΔP , V t , R m , V m , R 34 Five structural parameters serve as core perturbation factors, and are further extended to their time-series trajectories over the past 72 hours, thereby enabling dynamic capture of cyclone intensity evolution, scale structure changes, and historical energy injection processes. These parameters essentially constitute the complete input domain that determines the wave response prediction model, effectively ensuring the physical consistency and structural integrity of the model's input layer.
[0050] 3. Traverse all observation records that meet the conditions, and finally construct a fine-tuned measured calibration sample library for model correction in the subsequent transfer learning stage.
[0051] Step 7: Using the observed significant wave heights from the tropical cyclone wave height sample library obtained in Step 6 as the target output, further refine and train the coarse model for the rapid calculation of the significant wave height field of tropical cyclones in open sea areas obtained in Step 5: 1. Use the entire structure and parameters of the neural network model trained in step 5 as the initial weights; 2. Freeze all encoding and decoding layer parameters except for the output layer at the end of the fully connected network to ensure that the basic structure of the model remains consistent with the previously learned spatiotemporal perturbation relationship of the cyclone; 3. Using the observed effective wave heights in the tropical cyclone wave height sample library obtained in step 6 above as the target output, retrain the parameters of the model's terminal output layer. Without destroying the overall network stability structure, guide the network to adaptively adjust its response to new data distributions and observation biases. 4. Maintain the original training hyperparameter settings, monitor the independent validation set error in real time during training, and determine whether the convergence condition has been met.
[0052] The final model is a rapid calculation model (spatiotemporal fusion neural network model) for the significant wave height of tropical cyclones in open sea areas after observation and calibration. It can quickly, stably and reliably estimate the significant wave height of open sea areas dominated by tropical cyclones in complex environmental scenarios.
[0053] As can be seen, in the model fine-tuning stage, this invention introduces a freezing strategy and a small amount of observation data migration mechanism: keeping the parameters of each layer of the neural network except the output layer stable, and by loading the existing coarse model parameters and training only the output layer weights, the model can achieve adaptive fine-tuning of the distribution of new samples and the characteristics of observation errors with the support of limited satellite altimeter and buoy observation samples, effectively improving the model's generalization ability and prediction robustness in actual tropical cyclone events.
[0054] This invention constructs a highly efficient, accurate, and adaptively updated statistical model for predicting wave distribution characteristics at various grid points during tropical cyclones, improving the model's generalization performance and prediction accuracy in practical applications. It can rapidly predict the complete wave field distribution under a given combination of tropical cyclone parameters within seconds, significantly reducing computational resource consumption and time costs. Furthermore, the model not only possesses data-driven advantages but also exhibits a certain degree of physical constraints and observational consistency, enhancing its reliability in actual predictions.
[0055] To implement the above method embodiments, the rapid calculation system for the effective wave height field of tropical cyclones in open sea areas according to the present invention specifically includes: The data acquisition module is used to acquire information about the tropical cyclone event to be measured. The data processing module is used to embed the dynamic wind field structure into the information of the tropical cyclone event to be measured to construct a spatial-temporal distribution field, extract key cyclone parameters and relative position coordinates, and select data containing the current time period to construct a time series matrix. The fast calculation module is used to input the time series matrix into a pre-built spatiotemporal fusion neural network model to quickly calculate the spatiotemporal distribution of effective wave height in the open sea area at the target time. The model building module is used to build the spatiotemporal fusion neural network model in the above method embodiments.
[0056] Each module is mainly used to implement the various steps of the above method embodiments, which will not be described in detail here.
[0057] This application also provides a computer-readable storage medium, such as flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, server, App store, etc., which stores a computer program. When the program is executed by a processor, it performs corresponding functions. This embodiment implements a method for rapidly calculating the significant wave height field of a tropical cyclone in open sea areas when the computer-readable storage medium is executed by a processor.
[0058] Compared with existing technical solutions, the present invention has the following advantages: 1. Compared with numerical models, this invention has an absolute advantage in computational efficiency, and can complete the wave field prediction of a complete tropical cyclone event in a few seconds, which is far superior to the large computational cost of traditional numerical models.
[0059] 2. Compared with numerical models, interpolation statistical models, and parameterized models, this invention makes full use of reliable observational information on the effective wave height of tropical cyclones from altimeters and buoys, enabling transfer learning to make the model closer to the actual situation and significantly improving the simulation capability for extreme tropical cyclone scenarios.
[0060] 3. Good adaptability: The model can be used across regions and years, has adaptive update capabilities, and is suitable for climate statistics and extreme value analysis tasks.
[0061] It should be noted that, depending on the implementation needs, the various steps / components described in this application can be broken down into more steps / components, or two or more steps / components or parts of the operation of steps / components can be combined into new steps / components to achieve the purpose of this invention.
[0062] The order of the steps in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.
[0063] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.
Claims
1. A method for fast calculation of open ocean tropical cyclone significant wave height field, characterized in that, The method comprises the following steps: The dynamic wind field structure is embedded into the tropical cyclone event information to be measured, a space-time distribution field is constructed, the key parameters of cyclones and relative position coordinates in a preset time period containing the current time are extracted, and a time series matrix is formed; The time series matrix is input into the space-time fusion neural network model to quickly calculate the space-time distribution of the effective wave height in the open sea at the target time; The construction process of the space-time fusion neural network model is as follows: An existing tropical cyclone database is obtained, and systematic disturbance enhancement processing is performed to construct a cyclone prior sample set; The dynamic wind field structure is embedded into the cyclone prior sample set to construct a space-time distribution field of the tropical cyclone wind field, and a tropical cyclone wind field enhanced data set is formed; Based on the tropical cyclone wind field enhanced data set, a third-generation spectral numerical wave model is called to perform high space-time resolution wave numerical simulation on all tropical cyclone events in the open sea, extract the effective wave height response field in the simulation result, and obtain the tropical cyclone effective wave height data set under the open sea; A dynamic coordinate system with the cyclone movement path as the main axis is constructed, and the tropical cyclone effective wave height data set is subjected to coordinate system transformation and scale normalization processing to obtain a space structure normalized tropical cyclone effective wave height data set; The key parameters of cyclones and relative position coordinates in a preset time period containing the target time in the tropical cyclone wind field enhanced data set are extracted, a time series matrix is constructed and taken as input, and the effective wave height at the target position at the target time in the space structure normalized tropical cyclone effective wave height data set is taken as output, and the pre-constructed space-time fusion neural network model is trained.
2. The method for fast computation of open ocean tropical cyclone significant wave height field according to claim 1, characterized in that, Wherein, The space-time fusion neural network model is as follows: the cross correlation between different parameter time series in the time series matrix, the dynamic trend, local jump and nonlinear disturbance structure in the parameter time series are identified by using a multi-head attention mechanism, and residual learning is performed by introducing jump connection and feature map splicing in the feature reconstruction process. Then, a fully connected neural network is used as a nonlinear regression mapper to generate a direct estimated value of the target effective wave height. The model training target is to minimize the loss function between the model output and the actual wave height response value in the space structure normalized tropical cyclone effective wave height data set.
3. The method of claim 1, wherein, The training of the space-time fusion neural network model also includes fine tuning training, specifically: using the effective wave height information observed by the satellite altimeter and the buoy in the tropical cyclone database to construct an observation calibration sample set, and using the observation calibration sample set to fine tune the space-time fusion neural network model.
4. The method of fast computation of open ocean tropical cyclone significant wave height field according to claim 3, characterized in that, The construction process of the observation calibration sample set is as follows: All cyclone events in the tropical cyclone database are traversed, and the effective wave height data observed by the satellite altimeter and the buoy within a preset range are retrieved in the life cycle of each cyclone event; The effective wave height data within the preset range are subjected to space-time registration, and a fine tuning sample vector is constructed, wherein the information in each sample includes: the space normalized coordinates and the observed effective wave height of the observation point at the current time, the key parameters of the cyclone at the synchronous time, and the time series composed of the key parameters of the cyclone in the past preset time period and the corresponding cyclone center coordinates; The observation calibration sample set is constructed by traversing all fine-tuning sample vectors satisfying the condition.
5. The method of fast computation of open ocean tropical cyclone significant wave height field according to claim 1, wherein, The construction process of the cyclone prior sample set is as follows: for any given tropical cyclone event in the existing tropical cyclone database, a certain range of discrete translation operations is performed on the original center trajectory in the meridional and latitudinal dimensions to generate a set of artificial trajectories that are parallel to each other and have a small offset characteristic; a multi-dimensional perturbation mechanism is introduced into the cyclone key parameter time series corresponding to each original and offset trajectory, a constrained Monte Carlo sampling strategy is used to perform multiple rounds of random sampling within the preset range of the original values of the cyclone key parameters, and a smoothing operation is performed on the cyclone key parameter time series after each sampling to generate a perturbed sample set as the cyclone prior sample set.
6. The method of fast computation of open ocean tropical cyclone significant wave height field according to claim 1, wherein, The formation process of the tropical cyclone wind field enhancement data set is as follows: on the basis of the static center-symmetric wind field structure, a superimposed correction mechanism of the translation velocity vector field is introduced to generate a dynamic wind field structure with asymmetric disturbance characteristics, and the dynamic wind field structure is embedded in the cyclone prior sample set to construct the space-time distribution field of the tropical cyclone wind field and form the tropical cyclone wind field enhancement data set.
7. The method of fast computation of open ocean tropical cyclone significant wave height field according to claim 1, wherein, The specific process of spatial structure normalization of the tropical cyclone significant wave height data set is as follows: For each tropical cyclone event in the tropical cyclone significant wave height data set, a dynamic covariant coordinate system is constructed with the cyclone center as the origin, the cyclone travel direction as the Y-axis, and the Coriolis force direction orthogonal to the Y-axis as the X-axis. Projecting all the location points in the tropical cyclone significant wave height dataset with the relative distance of the cyclone center in the original latitude-longitude plane Delta x and Delta y Transform to the kinetic covariant coordinate system and normalize the spatial scale: x = Delta x / R max Delta y / R max wherein, x is the horizontal coordinate in the kinetic covariant coordinate system, y is the vertical coordinate in the kinetic covariant coordinate system, R max is the maximum wind speed radius.
8. The method of fast computation of open ocean tropical cyclone significant wave height field according to any one of claims 1-7, characterized in that, Key parameters of a cyclone include the pressure gradient Delta P , the cyclone propagation speed V t , the maximum wind speed radius R m , the maximum wind speed V m , and the 34 knot wind speed radius R 34 , with a default time period of 72 hours.
9. The method of fast computation of open ocean tropical cyclone significant wave height field according to claim 8, wherein, The existing tropical cyclone database is the global tropical cyclone multi-source reanalysis data set IBTrACS.
10. A system for fast computation of open ocean tropical cyclone significant wave height field, characterized in that, It includes: a data acquisition module for acquiring information of a to-be-measured tropical cyclone event; a data processing module for embedding a dynamic wind field structure into the to-be-measured tropical cyclone event information to construct a space-time distribution field, extracting cyclone key parameters and relative position coordinates, and selecting data of a preset time period including the current time to construct a time series matrix; a fast calculation module for inputting the time series matrix into a pre-constructed spatio-temporal fusion neural network model to quickly calculate the effective wave height spatio-temporal distribution of the open sea at the target time; a model construction module for constructing the spatio-temporal fusion neural network model according to the method of claim 1.