Method and system for analyzing marine storm elements based on a physical model

Abstract

The application discloses a marine storm element analysis method and system based on a physical model, and relates to the field of marine storm element analysis. The method comprises the following steps: acquiring marine environment initial field data and atmospheric forcing field data of a target sea area; constructing a neural network model embedded with physical constraints, wherein the physical constraints comprise Navier-Stokes equations and storm surge control equations; and dynamically adjusting the weight coefficients of the physical constraints in the neural network loss function according to different stages of storm evolution, wherein the storm evolution stages comprise a generation stage, a mature stage and a decay stage. By constructing the neural network model embedded with the physical constraints such as the Navier-Stokes equations and the storm surge control equations, the problem of poor physical consistency of a pure data-driven model is effectively solved, and reliable prediction capability can still be maintained in a sparse observation data area.

Description

Technical Field

[0001] This invention relates to the field of marine storm element analysis, specifically to a method and system for marine storm element analysis based on a physical model. Background Technology

[0002] Marine storms (such as typhoons and extratropical cyclones) are among the major hazardous weather systems affecting coastal safety and marine engineering activities. The accompanying strong winds, giant waves, and storm surges exhibit strong spatiotemporal evolution characteristics. Accurate analysis and forecasting of the evolution patterns of marine elements during storms are of great significance for maritime navigation, marine resource development, coastal protection, and disaster emergency response.

[0003] Currently, various physical model-based numerical forecasting methods are widely used in the analysis of marine storm elements. For example, regional ocean models based on the finite difference or finite volume method (such as ROMS and FVCOM) coupled with atmospheric models can simulate changes in key elements such as flow fields and water levels during storm processes. Some research has attempted to introduce data assimilation techniques, integrating observational data (such as buoy station and satellite remote sensing data) into the model's initial field or boundary conditions to improve forecast accuracy. In recent years, with the development of deep learning methods, purely data-driven neural network models have also been preliminarily explored for rapid prediction of storm evolution processes.

[0004] Existing technologies still suffer from the following shortcomings: purely data-driven neural network models lack physical constraints, making it difficult to guarantee the physical rationality and consistency of forecast results, and they perform poorly in ocean areas with sparse observational data; traditional numerical models have high computational costs, making it difficult to meet the needs of real-time or near-real-time storm forecasting; existing methods have varying degrees of dependence on physical mechanisms at different stages of storm evolution (formation, maturity, and decay), lacking targeted dynamic adjustment mechanisms; most existing methods only provide deterministic forecast results, making it difficult to provide the uncertainty range of forecast results, and lacking the ability to correct forecast errors online, resulting in insufficient predictive reliability during the rapid evolution of storms. Therefore, there is a need for a marine storm element analysis method and system that can integrate physical mechanisms, adapt to the characteristics of storm evolution stages, and possess uncertainty quantification and online correction capabilities. Summary of the Invention

[0005] Based on this, the purpose of this invention is to provide a method and system for analyzing marine storm elements based on a physical model, in order to solve the technical problems in the prior art where pure data-driven models lack physical constraints, resulting in poor physical consistency of forecast results and poor performance in sparse observation areas; traditional numerical models have high computational costs, making it difficult to meet real-time forecast requirements; lack dynamic adjustment mechanisms for different stages of storm evolution; and cannot provide forecast uncertainty ranges and online error correction capabilities.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method and system for analyzing marine storm elements based on a physical model.

[0007] The first aspect of this application provides a physical model-based method for analyzing marine storm elements, comprising: acquiring initial marine environmental field data and atmospheric forcing field data of a target sea area; constructing a neural network model embedded with physical constraints, the physical constraints including the Navier-Stokes equations and storm flood control equations; dynamically adjusting the weight coefficients of the physical constraints in the neural network loss function according to different stages of storm evolution, wherein the storm evolution stages include the formation period, maturity period, and decay period; inputting the initial field data and atmospheric forcing field data into the neural network model to predict the spatiotemporal evolution of at least one marine element during the storm; and outputting the prediction results and their corresponding uncertainty intervals, the uncertainty intervals being generated by the Monte Carlo dropout method built into the neural network model.

[0008] The first aspect of this application provides a method for analyzing marine storm elements based on a physical model, comprising: a data acquisition module for acquiring initial marine environmental data and atmospheric forcing field data of a target sea area; a model construction module for constructing a neural network model embedded with physical constraints, the physical constraints including the Navier-Stokes equations and storm surge control equations; a dynamic weight adjustment module for dynamically adjusting the weight coefficients of the physical constraints in the neural network loss function according to different stages of storm evolution, wherein the storm evolution stages include the formation period, maturity period, and decay period; a storm forecasting module for inputting the initial field data and atmospheric forcing field data into the neural network model to forecast the spatiotemporal evolution of at least one marine element during the storm; and an uncertainty output module for outputting the forecast results and their corresponding uncertainty intervals, the uncertainty intervals being generated by the Monte Carlo dropout method built into the neural network model.

[0009] In summary, the present invention has the following main beneficial effects:

[0010] This invention effectively solves the problem of poor physical consistency in purely data-driven models by constructing a neural network model that embeds physical constraints such as the Navier-Stokes equations and storm surge control equations, and maintains reliable forecasting capabilities even in regions with sparse observational data. Simultaneously, by dynamically adjusting the weight coefficients of physical constraints in the loss function according to the storm's formation, maturity, and attenuation stages, the model can adapt to the dominant physical mechanisms at different evolutionary stages, significantly improving forecast accuracy. Furthermore, this invention employs the Monte Carlo dropout method to output the uncertainty range of the forecast results and combines it with an online correction module to correct forecast errors in real time, overcoming the shortcomings of existing methods in lacking uncertainty quantification and error correction capabilities. This achieves efficient, accurate, and reliable spatiotemporal evolution forecasting of marine storm elements. Attached Figure Description

[0011] Figure 1 This is the control flowchart of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0013] The embodiments of the present invention will now be described.

[0014] A physical model-based method and system for analyzing marine storm elements, such as Figure 1 As shown, it includes:

[0015] I. Overall System Composition.

[0016] The physical model-based marine storm element analysis system of this invention includes the following functional modules: a data acquisition module, a model construction module (further including a sparse constraint training unit), a dynamic weight adjustment module, a storm forecasting module, an uncertainty output module, and an online correction module. These modules can be integrated into one or more computer servers, and their respective functions are achieved by the processor executing computer programs stored in the memory.

[0017] II. Specific implementation steps of the method.

[0018] Step 100: Obtain initial marine environmental field data and atmospheric forcing field data for the target sea area.

[0019] The initial marine environmental field data includes, but is not limited to, the vertical distribution of sea surface height (in meters, m) and current velocity (in meters per second, m / s) in the target sea area at the initial forecast time. This data can be obtained through: global ocean reanalysis datasets (such as HYCOM, GLORYS12v1), with a spatial resolution typically of 1 / 12° (approximately 9 kilometers) and a temporal resolution of 1 hour; or through on-site observations from buoy stations deployed in the target sea area. For example, in one embodiment, the northern part of the South China Sea (110° to 120° E, 18° to 25° N) was selected as the target sea area, with the initial time being during the passage of Typhoon "Hagupit" (September 12, 2014, 00:00 UTC). Initial current velocity field data for this area was obtained with a horizontal grid spacing of 5 km × 5 km and a vertical division into 40 layers.

[0020] The atmospheric forcing field data includes: wind speed (m / s), wind direction (degrees), sea level pressure (Pascals), atmospheric temperature (°C), relative humidity (%, dimensionless percentage), downward shortwave radiation (W / m²), and downward longwave radiation (W / m²) at a height of 10 meters above the sea surface. This data was acquired from the ERA5 reanalysis dataset of the European Centre for Medium-Range Weather Forecasts (ECMWF), with a temporal resolution of 1 hour and a spatial resolution of 0.25° × 0.25° (approximately 28 km). In this embodiment, the atmospheric forcing field data spans 72 hours before and after the typhoon's impact, with a time step of 10 minutes (600 seconds), used to drive the subsequent neural network model.

[0021] Step 200: Construct a neural network model with embedded physical constraints.

[0022] The physical constraints include the Navier-Stokes equations and the storm surge control equations. In this embodiment, all physical constraints are embedded simultaneously.

[0023] Specifically, a physical information neural network is constructed. This neural network model uses spatial coordinates ( , , (Unit: meters), Time coordinate () The input variables are atmospheric forcing field data (wind speed and air pressure) and target ocean features (velocity component). , , Unit: m / s; water level (Unit: m) is used as the output variable.

[0024] The neural network uses a fully connected deep neural network structure, comprising an input layer, six hidden layers, and an output layer. Each hidden layer contains 128 neurons, and the activation function is the hyperbolic tangent function. The network weights are initialized using the Xavier initialization method.

[0025] The physical constraints are embedded in the form of residual terms in the loss function. Specifically, the physical equation residuals are defined as follows:

[0026] (1) Incompressible Navier-Stokes equations (conservation of momentum):

[0027] ;

[0028] Where (using Einstein's summation convention, subscripts are used) , The value range is 1, 2, 3):

[0029] For the flow rate at Components in direction, , , These correspond to eastward, northward, and vertical flow velocities, respectively, in m / s.

[0030] Time, unit: seconds;

[0031] For spatial coordinates, , , These correspond to east, north, and vertical coordinates, respectively, in meters (m).

[0032] For reference, the density of seawater is taken as 1025 kg / m³;

[0033] Pressure, unit: Pa, approximated by hydrostatic pressure. Calculation, where m / s is the acceleration due to gravity. This is the still water depth (unit: m, measured from the chart datum). Storm surge height (unit: m);

[0034] The kinematic viscosity coefficient has a value of [value missing]. ;

[0035] For Coriolis force terms, among which , , Coriolis parameters , The Earth's rotation speed, Latitude (unit: rad, obtained through GPS coordinate conversion).

[0036] Momentum equation residual Defined as the degree to which the output of a neural network does not satisfy the above equation, specifically the square of the difference between the left and right sides of the equation.

[0037] (2) Storm surge control equation:

[0038] ;

[0039] in:

[0040] Storm rise height (relative to mean sea level), unit: m;

[0041] For the total water depth, The still water depth (unit: m, measured from the chart datum and obtained through bathymetry data);

[0042] , The vertical average velocity (unit: m / s) is obtained by integrating and averaging the three-dimensional velocity output by the neural network along the vertical direction.

[0043] , Horizontal coordinates, unit: m.

[0044] Storm surge residual Thus, the square of the value on the left-hand side of the equation is defined.

[0045] Step 300: Dynamically adjust the weight coefficients of physical constraints in the neural network loss function according to different stages of storm evolution.

[0046] The storm evolution process includes a formation phase, a maturity phase, and a decay phase. In this embodiment, the rate of change of the minimum sea surface pressure in the atmospheric forcing field is analyzed. (Unit: Pa / s) and maximum wind speed change rate (Unit: m / s²) Automatically determine the storm stage. When and The time is determined to be the generation period; when and (The decline in air pressure slows down) is then considered the maturation period; when Furthermore, the decay period is defined as lasting for more than 6 hours.

[0047] loss function Includes observation data-driven loss and physical constraint residual loss :

[0048] ;

[0049] in Further decomposed into momentum equation residual loss Storm rise equation residual loss :

[0050] ;

[0051] in , , , , The number of calculation points for each residual term (dimensionless integer).

[0052] The method for dynamically adjusting the physical constraint weight coefficients is as follows: During the storm generation period: increase the weight of the momentum equation residuals. , , All weighting coefficients are dimensionless.

[0053] During the mature stage: Increase the weight of the storm surge equation residuals. (Settings...) , , .

[0054] During the decay period: Increase the weight of the constraints related to the energy dissipation term. The energy dissipation term is represented by the viscous dissipation term in the Navier-Stokes equations. (Unit: m² / s³). At this point, the settings are... and in Internally, the dissipative term residuals are assigned sub-weights. The remaining sub-weights ,at the same time , 2.

[0055] Step 400: Input the initial field data and atmospheric forcing field data into the neural network model to predict the spatiotemporal evolution of at least one marine element during the storm.

[0056] The initial marine environmental field (temperature, salinity, and current velocity) obtained in step 100 is used as input, while time-series atmospheric forcing field data (wind speed and air pressure every 10 minutes) are fed into the neural network model hourly. Automatic Differentiation (AD) is used to calculate the physical residuals. The network is trained using the Adam optimizer with an initial learning rate of... (Dimensionless, representing the learning step size ratio), the number of training iterations is 50,000 (dimensionless integer). During training, the loss function... The value dropped to Stop training when the value is below the dimensionless level.

[0057] The forecast output includes hourly ocean elements for each grid point (horizontal resolution 5 km × 5 km, vertical 40 layers) within the target sea area over the next 72 hours, including the following: three-dimensional current field ( , , (Unit: m), Sea surface height ( (Unit: m, i.e., storm surge height).

[0058] Step 500: Output the forecast results and their corresponding uncertainty intervals.

[0059] The uncertainty interval is generated by the Monte Carlo Dropout method built into the neural network model.

[0060] Specifically, a Dropout layer is added after each hidden layer of the neural network model, with the Dropout rate set to [value missing]. (Dimensionless, representing the random discarding of 10% of neurons). During model training, Dropout maintains activation; during prediction, for the same input sample, [the process is repeated]. 100 forward propagations (dimensionless integers) are performed, with a different subset of neurons randomly discarded in each forward propagation. This results in 100 sets of predicted output values. ,in Indicates the first The arithmetic mean of the 100 outputs of a certain ocean element from the second forward propagation is taken as the final forecast result. :

[0061] ;

[0062] Take the standard deviation of these 100 outputs As a measure of uncertainty:

[0063] ;

[0064] The uncertainty interval at the 95% confidence level is: The coefficient 1.96 is the 0.975th quantile (dimensionless) of the standard normal distribution. For example, in forecasting storm surge height at a certain grid point... meters, standard deviation If the value is in meters, then the uncertainty interval is... .

[0065] III. Sparse constraint training methods for neural network models.

[0066] During the training of the neural network model, the data-driven term loss is calculated only at the spatiotemporal locations corresponding to the actual observation stations, while the physical residual term loss is calculated only at other locations.

[0067] In practice, a real-world dataset was obtained from the target sea area: 10 fixed buoy stations, with each station recording water level and current velocity every hour. A three-dimensional spatiotemporal mask matrix was then constructed. ,in:

[0068] like If there is actual measured observation data for the location, then (Dimensionless).

[0069] otherwise .

[0070] The data-driven loss is defined as:

[0071] ;

[0072] in:

[0073] The total number of all spatiotemporal location points (dimensionless integer);

[0074] For the neural network in the first The predicted values ​​for each spatiotemporal location are given, with units consistent with the corresponding physical quantities (water level in meters, flow velocity in meters per second).

[0075] This represents the observation value at the k-th spatiotemporal location, in the same units as above;

[0076] For the mask matrix Total number of positions (dimensionless integer), used for normalization.

[0077] All other unobserved spatiotemporal locations ( (Not calculated) Only calculate the loss of the physical residual term. (i.e., the aforementioned) , (The sum of these). Therefore, this method significantly reduces the reliance on dense observation data and is suitable for environments with sparse observations in open sea areas.

[0078] IV. Steps for real-time tracking of storm tracks and online correction of forecast errors.

[0079] The present invention also includes a real-time tracking of storm trajectories and an online correction step for forecast errors.

[0080] The specific implementation method is as follows:

[0081] (1) Real-time acquisition of newly arrived observation data: Assume the current time is (For example, 6 hours before the typhoon makes landfall) (The index is a dimensionless integer representing the time step; the model has given future times.) The predicted value. At this time, the system receives the new time. Actual observation data (water level measured at a certain buoy station).

[0082] (2) Calculate the residual between the current forecast value and the observed value:

[0083] ;

[0084] in:

[0085] For a moment The observed water level values, in meters;

[0086] For a moment Forecast water level values, unit: meters;

[0087] The water level residual is expressed in meters.

[0088] (3) The residual mapping is calculated as the correction amount of the hidden state at the current time through the temporal convolutional network, and superimposed on the hidden state of the neural network model at the current time to correct the prediction output at subsequent times.

[0089] In this embodiment, the Temporal Convolutional Network (TCN) is used. The input to this TCN is the most recent... Forecast residual sequence at (dimensionless integer) time steps The output is a correction vector. Its dimension is related to the hidden state of the neural network model at the current moment. The dimensions are the same. Hidden state It is a dimension A vector of dimensionless integers representing the number of neurons, where each component is dimensionless (due to the dimensionless output of the neural network activation function). The corrected hidden state is:

[0090] ;

[0091] Vector addition is performed component-wise. The corrected hidden state is used as the initial state to re-unfold the neural network model for subsequent time steps. , , The forecast.

[0092] Through the above online correction, the model can absorb the latest observations in real time, dynamically correct subsequent forecasts, and significantly improve the forecast accuracy during the rapid evolution of storms.

[0093] V. System Implementation Examples.

[0094] Corresponding to the above method, the present invention also provides a marine storm element analysis system based on a physical model, comprising:

[0095] Data Acquisition Module: Used to execute step 100, acquiring initial marine environmental field data and atmospheric forcing field data for the target sea area. This module communicates with the data server and supports reading NetCDF standard meteorological and oceanographic data format.

[0096] Model building module: This module is used to execute step 200, building a neural network model embedded with physical constraints. This module further includes a sparse constraint training unit, used to calculate the data-driven loss only at the spatiotemporal locations corresponding to the actual observation stations during training, and only the physical residual loss at other locations.

[0097] The dynamic weight adjustment module is used to execute step 300, dynamically adjusting the weight coefficients of physical constraints in the neural network loss function according to different stages of storm evolution (generation, maturity, and decay). Specifically, this module increases the weight of the momentum equation residual during the generation stage, increases the weight of the storm water gain equation residual during the maturity stage, and increases the weight of the energy dissipation term-related constraints during the decay stage.

[0098] Storm forecasting module: used to execute step 400, inputting initial field data and atmospheric forcing field data into the neural network model to forecast the spatiotemporal evolution of at least one marine element during the storm.

[0099] Uncertainty output module: used to execute step 500, output the forecast result and its corresponding uncertainty interval, which is generated by the Monte Carlo dropout method built into the neural network model.

[0100] Online correction module: This module is used to perform real-time tracking of storm trajectories and online correction of forecast errors. It acquires newly arriving observation data in real time, calculates the residual between the current forecast value and the observed value, and uses a temporal convolutional network to map the residual into the correction amount of the hidden state at the current moment. This correction is then added to the hidden state of the neural network model at the current moment to correct the forecast output at subsequent moments.

[0101] The data flow and control flow between the above modules are implemented through software interfaces and deployed on a single high-performance computing server.

[0102] VI. Parameter Sources and Feasibility Explanation.

[0103] Seawater density The value is 1025 kg / m³, which is the standard oceanographic reference value.

[0104] kinematic viscosity coefficient The value is This is a typical value for seawater at 20℃;

[0105] gravitational acceleration m / s, which is the standard value of the International Committee for Weights and Measures;

[0106] Earth's rotation speed International Astronomical Union standard value;

[0107] Coriolis parameters :latitude It is calculated from GPS coordinates in radians (rad).

[0108] The values ​​mentioned above (such as 128 neurons in the hidden layer, Dropout rate of 0.1, forward propagation times of 100, and time window length of 10) are all exemplary parameters. Those skilled in the art can make reasonable adjustments based on factors such as actual sea area scale, storm intensity, and computing resources, without the need for creative effort.

[0109] This application, through its specific embodiments and in conjunction with example data and parameter definitions, fully and clearly discloses the physical model-based marine storm element analysis method and system claimed in this invention. Those skilled in the art, based on the above description, can implement this invention without excessive experimentation and achieve the aforementioned beneficial effects: improved physical consistency of forecasts, adaptation to storm evolution stages, reduced dependence on dense observation data, provision of uncertainty intervals, and online error correction.

[0110] The steps of the methods or algorithms described in this application can be directly embedded in hardware, software units executed by a processor, or a combination of both. Exemplarily, a storage medium can be connected to a processor so that the processor can read information from and write information to the storage medium. Optionally, the storage medium can also be integrated into the processor. Optionally, the processor and the storage medium can also be located in different components within a terminal. These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0111] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely illustrative examples of this application and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Thus, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A method for analyzing marine storm elements based on a physical model, characterized in that, include: Acquire initial marine environmental data and atmospheric forcing field data for the target sea area; A neural network model embedded with physical constraints is constructed, including the Navier-Stokes equations and storm surge control equations. The weighting coefficients of the physical constraints in the neural network loss function are dynamically adjusted according to different stages of storm evolution. These stages include the formation, maturity, and decay phases, defined by analyzing the rate of change of the minimum sea surface pressure in the atmospheric forcing field. and maximum wind speed change rate Automatically determine the storm phase, when and The time is determined to be the generation period; when and Later determined to be in the mature stage; when And the decay period is defined as lasting for more than 6 hours. At maximum wind speed, This is the lowest air pressure at sea level. This represents the minute change in the lowest atmospheric pressure at sea level. This represents a tiny change in the maximum wind speed. For a small change over time; The initial field data and atmospheric forcing field data are input into the neural network model to predict the spatiotemporal evolution of at least one marine element during the storm. The forecast results and their corresponding uncertainty intervals are output, which are generated by the Monte Carlo dropout method built into the neural network model.

2. The method for analyzing marine storm elements based on a physical model according to claim 1, characterized in that: The method for dynamically adjusting the weight coefficients of physical constraints is as follows: during the storm generation period, increase the weight of the momentum equation residuals, specifically by setting the weight coefficient of the momentum equation residual terms to more than twice the weight of the data-driven term loss, while reducing the weight of the storm water gain equation residual terms. During the storm maturity phase, the weight of the residuals in the storm water gain equation is increased. Specifically, the weight coefficient of the residual terms in the storm water gain equation is set to more than 1.5 times the weight of the loss of the data-driven terms, while the weight of the residual terms in the momentum equation is reduced. During the storm decay period, the weight of the energy dissipation term related constraints is increased. Specifically, the sub-weight of the residual of the viscous dissipation term in the Navier-Stokes equation is increased to more than 0.6 of the total weight of the residual of the equation, and the weight of other physical constraint terms is reduced accordingly. The weight coefficients are smoothly transitioned using continuous functions or piecewise constant functions during the storm evolution process to avoid model training instability caused by sudden weight changes.

3. The method for analyzing marine storm elements based on a physical model according to claim 1, characterized in that: During the training process of the neural network model, the data-driven term loss is calculated only at the spatiotemporal location corresponding to the actual observation station or satellite observation trajectory, while only the physical residual term loss is calculated at other locations, in order to reduce the dependence on dense observation data.

4. The method for analyzing marine storm elements based on a physical model according to claim 1, characterized in that: It also includes a step-by-step real-time tracking of storm tracks and online correction of forecast errors: Acquire newly arriving observation data in real time; Calculate the residual between the current forecast and the observed values: ； in, For a moment Observed water level values, unit: meters. For a moment Forecast water level values, unit: meters. Water level residual, unit: meters; The residual mapping is calculated as a correction amount for the hidden state at the current time step using a temporal convolutional network, and then superimposed on the hidden state of the neural network model at the current time step to correct the prediction output at subsequent time steps.

5. The method for analyzing marine storm elements based on a physical model according to claim 4, characterized in that: The temporal convolutional network takes the predicted residual sequence as input, performs correction, and obtains the corrected hidden state: ； in, This is the corrected implicit state. Given a vector with dimension D=128, This is a correction vector output by the temporal convolutional network.

6. A marine storm element analysis system based on a physical model, characterized in that: The data acquisition module is used to acquire initial marine environmental data and atmospheric forcing field data of the target sea area; The model building module is used to build a neural network model embedded with physical constraints, including the Navier-Stokes equations and storm surge control equations. The dynamic weight adjustment module is used to dynamically adjust the weight coefficients of the physical constraints in the neural network loss function according to different stages of storm evolution. The storm evolution stages include the generation stage, maturity stage, and decay stage. The generation stage is the stage when the storm initially forms, its intensity has not yet reached the typhoon level, and the maximum wind speed is on the rise. The maturity stage is the stage when the storm intensity reaches or exceeds the typhoon level, the central pressure decreases slowly, and the maximum wind speed remains at a high level. The decay stage is the stage when the maximum wind speed of the storm continues to decrease, and this downward trend is maintained for a certain period of time. The storm forecasting module is used to input the initial field data and atmospheric forcing field data into the neural network model to predict the spatiotemporal evolution of at least one marine element during the storm. The uncertainty output module is used to output the forecast results and their corresponding uncertainty intervals, which are generated by the Monte Carlo dropout method built into the neural network model.

7. The marine storm element analysis system based on a physical model according to claim 6, characterized in that: The dynamic weight adjustment module is specifically used to: increase the weight of the momentum equation residual during the generation period, increase the weight of the storm surge equation residual during the maturity period, and increase the weight of the energy dissipation term related constraints during the decay period.

8. The marine storm element analysis system based on a physical model according to claim 6, characterized in that: The model building module also includes a sparse constraint training unit, which is used to calculate the data-driven term loss only at the spatiotemporal locations corresponding to the actual observation stations or satellite observation trajectories during the training process, and only calculate the physical residual term loss at other locations.

9. A marine storm element analysis system based on a physical model according to claim 6, characterized in that: It also includes an online calibration module for: Acquire newly arriving observation data in real time; Calculate the residual between the current forecast and the observed values; The residual mapping is calculated as a correction amount for the hidden state at the current time step using a temporal convolutional network, and then superimposed on the hidden state of the neural network model at the current time step to correct the prediction output at subsequent time steps.

10. A marine storm element analysis system based on a physical model according to claim 6, characterized in that: The temporal convolutional network takes the predicted residual sequence as input and outputs a correction vector of the hidden state.

Images