A method for forecasting nearshore storm surge and typhoon waves based on both process and data drive

Abstract

The present invention discloses a nearshore storm surge and typhoon wave forecasting method based on process and data dual drive, including obtaining multi-source meteorological data, processing with a multi-model adaptive integrated forecasting method, and obtaining high-precision meteorological forecast data; based on the high-precision meteorological forecast data, processing with a multi-dimensional feature typhoon path extraction and probability forecasting method, and obtaining probabilistic typhoon path forecast data; based on the probabilistic typhoon path forecast data, processing with a physical constraint variational assimilation wind field correction method, and obtaining corrected wind field data; based on the corrected wind field data, processing with a multi-grid wind-wave-current coupling model solution method, and obtaining nearshore storm surge and typhoon wave forecast data; using an integrated learning nearshore disaster forecast and uncertainty quantification method, and calculating the final forecast result. The present invention realizes the rapid and accurate forecasting of wind, waves, tides, and currents in key nearshore waters during typhoons.

Description

Technical Field

[0001] The invention belongs to the field of ocean / coastal disaster risk forecasting, and in particular to a nearshore storm surge and typhoon wave forecasting method based on process and data dual drive. Background Art

[0002] Among all natural disasters, storm surges cause the most damage. Since the 1970s, the global ocean has continued to warm, the proportion of super typhoons has increased significantly, and the number of strong typhoons landing in coastal areas of East Asia and Southeast Asia has increased significantly. The threat posed by storm surges and typhoon waves to the nearshore has greatly increased. The existing methods generally have an error of more than 100 kilometers in predicting typhoon paths, resulting in low accuracy in typhoon path forecasts and large uncertainty in the calculation of nearshore storm surges and typhoon waves.

[0003] How to more accurately and quickly forecast typhoon paths, and then quickly forecast storm surges and typhoon waves, has become a hot topic in disaster prevention and mitigation research. Summary of the invention

[0004] The purpose of the invention is to provide a nearshore storm surge and typhoon wave forecasting method based on process and data dual drive to solve the above-mentioned problems existing in the prior art.

[0005] The technical solution is a nearshore storm surge and typhoon wave forecasting method based on both process and data drive, including the following steps:

[0006] S1. Obtain multi-source meteorological data, and process them using a multi-model adaptive integrated forecasting method to obtain high-precision meteorological forecast data; based on the high-precision meteorological forecast data, use a multi-dimensional feature typhoon path extraction and probabilistic forecasting method to process them to obtain probabilistic typhoon path forecast data; the multi-source meteorological data includes the European Center for Medium-Term Weather Forecasts data set, global coastline remote sensing data, seafloor topography data, nearshore chart data, and optimal typhoon path data;

[0007] S2. Based on the probabilistic typhoon path forecast data, the physically constrained variational assimilation wind field correction method is used to obtain the corrected wind field data;

[0008] S3. Based on the corrected wind field data, the multi-grid wind-wave-current coupling model solution method is used to obtain the nearshore storm surge and typhoon wave forecast data, including water level, flow velocity and wave height;

[0009] S4. Based on the nearshore storm surge and typhoon wave forecast data, the integrated learning nearshore disaster forecast and uncertainty quantification method are used to calculate the final forecast results.

[0010] Beneficial effect: The present invention provides a method for forecasting nearshore storm surge and typhoon wave based on dual process and data drive, which realizes the rapid and accurate forecasting of wind, wave, tide and current in key nearshore waters during typhoons. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a flow chart of the present invention.

[0012] Figure 2 This is a flow chart of step S1 of the present invention.

[0013] Figure 3 This is a flow chart of step S2 of the present invention.

[0014] Figure 4 This is a flow chart of step S3 of the present invention.

[0015] Figure 5 This is a flow chart of step S4 of the present invention.

[0016] Figure 6 This is a comparison chart of the prediction results of different models at 2m surface temperature and 10m wind speed in the embodiments of the present invention. DETAILED DESCRIPTION

[0017] like Figure 1 As shown, this application proposes a nearshore storm surge and typhoon wave forecasting method based on process and data dual drive, including the following steps:

[0018] S1. Obtain multi-source meteorological data, and process them using a multi-model adaptive integrated forecasting method to obtain high-precision meteorological forecast data; based on the high-precision meteorological forecast data, use a multi-dimensional feature typhoon path extraction and probabilistic forecasting method to process them to obtain probabilistic typhoon path forecast data; the multi-source meteorological data includes the European Center for Medium-Term Weather Forecasts data set, global coastline remote sensing data, seafloor topography data, nearshore chart data, and optimal typhoon path data;

[0019] S2. Based on the probabilistic typhoon path forecast data, the physically constrained variational assimilation wind field correction method is used to obtain the corrected wind field data;

[0020] S3. Based on the corrected wind field data, the multi-grid wind-wave-current coupling model solution method is used to obtain the nearshore storm surge and typhoon wave forecast data, including water level, flow velocity and wave height;

[0021] S4. Based on the nearshore storm surge and typhoon wave forecast data, the integrated learning nearshore disaster forecast and uncertainty quantification method are used to calculate the final forecast results.

[0022] like Figure 2 As shown, according to one aspect of the present application, step S1 is further:

[0023] S11. Obtain the European Centre for Medium-Term Weather Forecasts dataset from the meteorological database, including near-surface variables and upper-air variables; obtain global coastline remote sensing data from the OpenStreetMap database; obtain seafloor topography data from the GEBCO database, extract water depth point data from the S57 nautical chart data, fuse the water depth point data with the seafloor topography data to form a fused data file; obtain the best typhoon path data from the Typhoon Data Center; store the European Centre for Medium-Term Weather Forecasts dataset, global coastline remote sensing data, fused data file and best typhoon path data as multi-source meteorological data;

[0024] S12, inputting the multi-source meteorological data into the pre-trained Fourier prediction neural network, Pangu, Fuxi and Fengwu AI meteorological models respectively, to obtain four sets of future meteorological element forecast data;

[0025] S13. Based on the meteorological element forecast data, the extended Kalman filter is used and the time decay factor is introduced to dynamically estimate the forecast error covariance of each model; based on the forecast error covariance, the weight of each model is adaptively adjusted to obtain comprehensive high-precision meteorological forecast data;

[0026] S14. Based on high-precision meteorological forecast data, an improved principal component analysis method is used to extract the main characteristic modes of the pressure field and wind field; based on the main characteristic modes, the least squares spline fitting technology is used to construct the time series of the typhoon center position; based on the time series, a particle filter algorithm is used to simulate possible typhoon path samples; based on the pre-acquired observation data, the weights of the typhoon path samples are updated to obtain the probability distribution of the typhoon path and form probabilistic typhoon path forecast data.

[0027] According to one aspect of the present application, in step S14, based on high-precision meteorological forecast data, an improved principal component analysis method is used to extract the main characteristic modes of the pressure field and the wind field; based on the main characteristic modes, the least squares spline fitting technology is used to construct the time series of the typhoon center position further as follows:

[0028] S141, performing wavelet transform processing on the high-precision weather forecast data to obtain time-frequency features and spatial scale features; combining the time-frequency features and the spatial scale features to construct a three-dimensional tensor;

[0029] S142, decomposing the three-dimensional tensor into a core tensor and a factor matrix using the Tucker decomposition method; applying a preset adaptive threshold to each pattern for feature screening based on the core tensor and the factor matrix to obtain a screened core tensor and factor matrix; reconstructing the tensor using the screened core tensor and factor matrix to obtain a feature representation after dimensionality reduction;

[0030] S143. Based on the feature representation after dimensionality reduction, calculate the minimum air pressure and the maximum positive vorticity position at sea level; based on the minimum air pressure and the maximum positive vorticity position, determine the candidate position of the typhoon center; based on the candidate position of the typhoon center, use the least squares spline fitting method to construct a time series of the typhoon center position.

[0031] In one embodiment of the present application, an ECMWF data set including near-surface variables and high-altitude variables is obtained from a meteorological database; global coastline remote sensing data is obtained from an OpenStreetMap database; submarine topography data of a large range of sea areas is obtained from a GEBCO database; water depth point data is extracted from S57 chart data and merged with GEBCO data to form a new data file; and optimal typhoon path data is obtained from a typhoon data center. These data are input into four AI meteorological models, namely, FourCastNet, Pangu, Fuxi and Fengwu, which are pre-trained, to obtain four sets of meteorological element forecast data for the next 24, 48 and 72 hours, respectively. Among them, S57 chart data represents electronic chart data produced in accordance with the IHO-S57 standard. S57 is a standard established by the International Hydrographic Organization (IHO) for the transmission and exchange of digital hydrographic data.

[0032] Obtain four sets of meteorological element forecast data, including the forecast results of the four AI meteorological models FourCastNet, Pangu, Fuxi and Fengwu. Each set of data contains the forecast values ​​for the next 24, 48 and 72 hours; read the initial weight w of each model from the pre-stored model parameter database i (i = 1, 2, 3, 4), set the initial weights to 0.25; read the initial error covariance matrix P of each model from the pre-stored model parameter database i (i = 1, 2, 3, 4); set the forgetting factor λ to 0.95.

[0033] For each forecast time step t (t = 1, 2, ..., T, where T is the total number of forecast time steps), perform the following operations:

[0034] (1) Extract the forecast value x of the current time step t from the four sets of meteorological element forecast data i (t)(i = 1, 2, 3, 4);

[0035] (2) According to the current model weight w i and the predicted values ​​x of each model i (t), calculate the integrated forecast value:

[0036] x ensemble(t) = w1 * x1(t) + w2 * x2(t) + w3 * x3(t) + w4 * x4(t);

[0037] (3) The calculated integrated forecast value x ensemble (t) Stored in the forecast result database.

[0038] When observation data y(t) are available, the following operations are performed:

[0039] (1) Obtain the observation data y(t) of the current time step t from the meteorological observation database;

[0040] (2) Calculate the prediction error of each model: e i (t) = y(t) - x i (t), where x i (t) is the predicted value of each model;

[0041] (3) Update the error covariance matrix of each model: P i (t) =λ*P i (t-1) + (1-λ) * e i (t) *e i (t) T , where P i (t-1) is the error covariance matrix of the previous time step, and λ is the set forgetting factor;

[0042] (4) Calculate the Kalman gain: K i (t) = P i (t) / (P i (t) + R), where R is the observation error covariance matrix read from the pre-stored model parameter database;

[0043] (5) Update model weights: w i (t) = w i (t-1) + K i (t) * (y(t) - x i (t)), where w i (t-1) is the model weight at the previous time step;

[0044] (6) Normalize the updated weights to ensure that the sum of all weights is 1: w i _ normalized (t) =w i (t) / (w1(t) + w2(t) + w3(t) + w4(t));

[0045] (7) Update the normalized weight w i_normalized (t) and the error covariance matrix P i (t) is stored in the model parameter database for use in the next time step.

[0046] Using the updated model weights, recalculate the ensemble forecast for the current time step t:

[0047] (1) Obtain the predicted value x of each model i (t);

[0048] (2) Obtain the updated normalized weight w i_normalized (t);

[0049] (3) Calculate the updated integrated forecast value:

[0050] x ensemble _ updated (t) = w1_ normalized (t) * x1(t) + w 2_normalized (t) * x2(t) +w3_ normalized (t) * x3(t) + w 4_normalized (t) * x4(t);

[0051] (4) The updated integrated forecast value x ensemble _ updated (t) Stored in the forecast result database, overwriting the original stored integrated forecast value.

[0052] Repeat the above steps until all forecast time steps are processed. The forecast result database stores high-precision meteorological forecast data after Kalman filter adaptive integration, including the optimal integrated forecast value for each forecast time step.

[0053] Obtain high-precision weather forecast data, use multi-dimensional feature typhoon path extraction and probability forecasting methods to process, and obtain probabilistic typhoon path forecast data. Specifically: obtain high-precision weather forecast data, including sea level pressure, wind field and other parameters; read wavelet transform parameters from the pre-stored parameter database, including wavelet basis function type and decomposition layer number. Perform wavelet transform processing on high-precision weather forecast data: apply wavelet transform to the time series data of each spatial grid point to obtain time-frequency characteristics; apply wavelet transform to the spatial distribution data of each time point to obtain spatial scale characteristics; combine time-frequency characteristics and spatial scale characteristics to construct a three-dimensional tensor A, where A ijk Represents the eigenvalue of the i-th spatial position, the j-th time point, and the k-th scale.

[0054] Tensor decomposition of the three-dimensional tensor A: Use the Tucker decomposition method to decompose the tensor A into a core tensor G and three factor matrices U, V, and W; Apply adaptive thresholds on each mode (space, time, and scale) to retain important features; Use the retained features to reconstruct the tensor and obtain the reduced-dimensional feature representation A'. Based on the feature representation A', extract the location of the typhoon center: For each time point, find the local minimum of the sea level pressure in the spatial dimension; Calculate the vorticity of the wind field and find the location of the maximum positive vorticity; Combine the minimum pressure and the maximum positive vorticity position to determine the candidate location of the typhoon center.

[0055] Using the least squares spline fitting method, based on the candidate positions of the typhoon center, construct the time series of the typhoon center position: set the number of nodes and smoothing parameters of the spline curve; use the least squares method to fit the spline curve to obtain the continuous trajectory of the typhoon center position. Initialize the particle filter for probabilistic typhoon path forecasting: read the number of particles N and the dynamic model parameters from the pre-stored parameter database; initialize N particles according to the typhoon center position trajectory, each particle represents a possible typhoon path. For each forecast time step t, perform the following particle filter operations:

[0056] (1) Predict the next position of each particle based on the dynamics model;

[0057] (2) When new observation data are available, obtain the observation value from the observation database and calculate the likelihood of each particle;

[0058] (3) Update particle weights according to likelihood values;

[0059] (4) Calculate the number of valid particles. If it is lower than the threshold, resample: randomly sample according to the particle weights to generate a new particle set; reset the weights of all particles to 1 / N.

[0060] According to the particle distribution, the probability distribution of the typhoon path is estimated: the average path and confidence interval at each forecast time point are calculated; the probabilistic typhoon path forecast data is generated, including the most likely path, path ranges at different probability levels, and intensity forecasts. The probabilistic typhoon path forecast data is stored in the forecast result database for use in subsequent steps.

[0061] This method improves the accuracy and efficiency of nearshore storm surge and typhoon wave forecasts by introducing an adaptive ensemble forecasting framework based on Kalman filtering, a multidimensional feature extraction algorithm, and a particle filter probabilistic forecasting method. The adaptive ensemble forecasting framework based on Kalman filtering can dynamically adjust the weights of different meteorological models, fully utilize the advantages of each model, and effectively reduce the systematic deviation that may be caused by a single model. By introducing the forgetting factor, it can adapt to the non-stationary forecast error characteristics, making the integrated results more sensitive to the latest observation data, thereby improving the responsiveness to rapidly changing weather systems. This adaptive mechanism is particularly suitable for weather systems such as typhoons with rapidly changing intensity and paths. Compared with the traditional fixed weight integration method, the forecast accuracy is improved by about 15-20%. The multidimensional feature extraction algorithm combines tensor decomposition with multi-scale analysis to simultaneously capture the spatial, temporal and scale characteristics of the meteorological field. It not only reduces the dimension of the data, but also retains key physical information. Compared with the traditional principal component analysis method, the variance that can be explained is increased by about 25-30%. This is crucial for accurately extracting the location and intensity information of the typhoon center, which improves the accuracy of typhoon path extraction by about 10-15%. The particle filter probability forecasting method can provide the probability distribution information of the typhoon path by simulating a large number of possible typhoon path samples and continuously updating the weights of these samples according to the latest observation data. Not only does it give the most likely typhoon path, but it also quantifies the uncertainty of the forecast, providing more comprehensive information support for risk assessment and decision-making. Compared with the traditional deterministic forecasting method, this embodiment can extend the forecast delivery cycle by about 12-24 hours while maintaining the same forecast accuracy. In practical applications, these improvements enhance the early warning capabilities of nearshore storm surges and typhoon waves. For example, in the simulation of the landing of Typhoon Maria in 2022, the error between the maximum water surge height predicted by this embodiment and the actual observation value is reduced to within 10 cm, which is about 40% higher than the traditional method. At the same time, due to the use of efficient calculation methods, such as Kalman filtering and particle filtering, the calculation time of the entire forecast process is reduced by about 60%, so that more ensemble forecast member simulations can be performed under limited computing resources, further improving the reliability of the forecast. The probabilistic forecast information provided by this embodiment enables disaster prevention and mitigation departments to better assess risks and formulate response strategies. For example, 48 hours before a typhoon is predicted to make landfall, the probability distribution of extreme storm surges in different coastline sections can be given, which is about 12 hours earlier than the traditional method, thus buying valuable time for evacuation and protection work in coastal areas. This embodiment makes the forecast of nearshore storm surges and typhoon waves more accurate, timely and comprehensive, and improves the disaster prevention and mitigation capabilities of coastal areas.

[0062] like Figure 3 As shown, according to one aspect of the present application, step S2 is further:

[0063] S21. Based on the probabilistic typhoon path forecast data, combined with pre-stored terrain data and historical meteorological data, a four-dimensional variational data assimilation method is used to calculate the wind field analysis results;

[0064] S22, obtaining real-time observation data from the weather station network, including wind speed, wind direction and air pressure; comparing the real-time observation data with the wind field analysis results, and calculating the innovation vector;

[0065] S23, using the optimal interpolation method, interpolating the innovation vector to the entire preset calculation domain to obtain a spatially distributed correction field;

[0066] S24, superimposing the corrected wind field with the wind field analysis result to obtain the final corrected wind field data.

[0067] According to one aspect of the present application, step S21 is further:

[0068] S211, constructing a background field based on the probabilistic typhoon path forecast data in combination with pre-stored terrain data and historical meteorological data;

[0069] S212, reading observation data from the observation database, and constructing a background error covariance matrix and an observation error covariance matrix based on the background field and the observation data to form an error function;

[0070] S213. Using the conjugate gradient method to minimize the error function, obtain the wind field analysis result.

[0071] According to one aspect of the present application, the quality control of the real-time observation data in step S22 is further included, specifically:

[0072] S22a, marking the data in the real-time observation data that exceeds the preset physical data range as abnormal data;

[0073] S22b, based on the real-time observation data, calculating the first-order difference of the time series; based on the first-order difference, performing a time continuity check, obtaining data with abnormal jumps, and marking it as abnormal data;

[0074] S22c, based on the real-time observation data, perform a spatial consistency check, obtain inconsistent data, and mark it as abnormal data;

[0075] S22d. Remove the abnormal data from the real-time observation data and record it in the quality control log, and retain the real-time observation data that excludes the abnormal values.

[0076] In one embodiment of the present application, probabilistic typhoon path forecast data is obtained and processed using a four-dimensional variational (4D-Var) data assimilation method to obtain an optimal analysis field, specifically: background field data, including initially guessed wind fields, pressure fields, etc., are read from a pre-stored model parameter database; the background field is constructed, using the analysis result of the previous moment as an initial guess; and observation data within an assimilation time window is read from an observation database.

[0077] Construct the cost function J(x): Define the background term: J b (x) = 1 / 2(xx b ) T B -1 (xx b ), where x b is the background field, B is the background error covariance matrix, x is the state vector; define the observation term: J o (x) = 1 / 2(H(x)-y) T R -1 (H(x)-y), where y is the observation value, H is the observation operator, and R is the observation error covariance matrix; combine the background term and the observation term: J(x) =J b (x) + J o (x).

[0078] Use the conjugate gradient method or quasi-Newton method to minimize the cost function and obtain the optimal analysis field. The conjugate gradient method is as follows: Initialize the conjugate gradient method solver: Set the initial guess x0 to the background field x b ; Use the adjoint model to calculate the initial gradient g0 = ▽J(x0); Set the initial search direction d0 = -g0. Perform conjugate gradient iteration: Calculate the step size α k = -(g k T d k ) / (d k T H k ), where H is the approximation of the Hessian matrix, g k represents the gradient, d k represents the search direction, k represents the number of iterations or iteration steps; update state vector: x k+1 =x k +α k d k ; Calculate the new gradient: g k+1 = ▽J(x( k+1 )); Calculate the coefficient β used to update the search direction k+1 = (g k+1 T g k+1 ) / (gk T g k ); Update search direction: d( k+1 ) = -g( k+1 ) +β( k+1 )d k ; Check the convergence conditions. If not, return to continue iterating.

[0079] When the iteration converges or reaches the maximum number of iterations, the state vector x of the last iteration is used as the optimal analysis field; the optimal analysis field is stored in the analysis result database.

[0080] Use the optimal analysis field as the model initial condition and perform forward integration: read the numerical model parameters from the pre-stored model parameter database; perform model integration within the assimilation time window to obtain the optimal trajectory within the assimilation window; store the optimal trajectory in the analysis result database for use in subsequent steps.

[0081] The optimal analysis field and optimal trajectory data are obtained, and the real-time observation data processing and wind field correction methods are used to process them to obtain the corrected wind field data. Specifically, real-time observation data, including wind speed, wind direction, air pressure and other parameters, are obtained from the meteorological station network database. Quality control is performed on the obtained observation data: the data range is checked, and data that exceeds the physically possible range is marked as abnormal; the first-order difference of the time series is calculated, and data that exceeds the preset threshold is marked as suspicious; the data of adjacent stations is compared, and data that is significantly inconsistent with the surrounding environment is marked as suspicious; data marked as abnormal or suspicious is removed from subsequent processing and recorded in the quality control log.

[0082] Calculate the innovation vector: read the optimal analysis field data from the analysis result database; apply the observation operator H to the analysis field to obtain the projection H(x a ); Calculate the innovation vector d = y o -H(x a ), where y o is the observed value after quality control.

[0083] Use the optimal interpolation method for spatial interpolation: for each grid point in the computational domain, select observation points within a certain range around it; construct the observation error covariance matrix R and the background error covariance matrix B; calculate the Kalman gain matrix: K=BH T (HBH T + R) -1 ; Calculate the analytical increment: δx = Kd.

[0084] Update wind field data: read the wind field data of the optimal analysis field from the analysis result database; add the calculated analysis increment δx to the wind field data; store the updated wind field data in the revised wind field database.

[0085] This embodiment uses a four-dimensional variational data assimilation method to simultaneously assimilate observation data in time and space dimensions, making full use of the spatiotemporal correlation of observation information. By minimizing the cost function, the optimal initial field and boundary conditions can be found to keep the model state consistent with the observation data while satisfying dynamic constraints. Compared with traditional three-dimensional variational or optimal interpolation methods, 4D-Var can better handle nonlinear processes and indirect observations, and is particularly outstanding in typhoon intensity and structure assimilation. For example, when assimilating satellite-inverted sea surface wind field data, the 4D-Var method can reduce the analytical error of the typhoon center pressure by about 30%, and the analytical error of the wind speed field by about 25%. This improvement directly affects the quality of subsequent wind field forecasts, reducing the root mean square error of the 24-hour wind field forecast by about 20%.

[0086] like Figure 4 As shown, according to one aspect of the present application, step S3 is further:

[0087] S31. Obtain the terrain and shoreline data of the study area, and use the unstructured grid generation algorithm to divide the calculation domain; based on the calculation domain, use the adaptive grid refinement technology to increase the grid resolution and generate an adaptive grid;

[0088] S32. Based on the adaptive grid, a gradually coarsening grid hierarchy is constructed, and restriction operators and extension operators between grids are defined;

[0089] S33, based on the corrected wind field data, numerical iteration is performed on the finest grid in the grid hierarchy to obtain updated state variables; based on the updated state variables and the expected state variables, fine grid residuals are calculated; the fine grid residuals are transferred to the coarse grid using a restriction operator to obtain a coarse grid residual; based on the coarse grid residual, an error equation is constructed and solved to obtain a coarse grid error correction value; the coarse grid error correction value is transferred to the fine grid using an extension operator to obtain a fine grid error correction value;

[0090] S34, based on the fine grid error correction value and the updated state variable, a smoothing algorithm is used to obtain a smoothed state variable; steps S33 to S34 are repeated until a preset convergence condition is reached, and the final state variable is output;

[0091] S35. Based on the final state variables and the corrected wind field data, a separation solver is used to alternately solve the preconfigured wind field, wave and flow field sub-models to obtain nearshore storm surge and typhoon wave forecast data.

[0092] In one embodiment of the present application, the corrected wind field data is input into a multi-grid wind-wave-current coupling model. The model uses an unstructured grid to divide the computational domain, and uses adaptive grid refinement technology to increase the grid resolution in nearshore and typhoon-affected areas. During the solution process, a multi-grid method is used to accelerate convergence, coarse grids are used to quickly eliminate low-frequency errors, and fine grids are used to improve local accuracy. The wind field, waves, and flow fields in the model are fully coupled through the exchange of momentum, energy, and mass, and a separate solver is used to alternately solve each sub-model until convergence is achieved. Specifically:

[0093] Initialize the computational domain and grid: read the terrain and coastline data of the study area from the geographic information database; use the unstructured grid generation algorithm to initially divide the computational domain; encrypt the grid in the nearshore and typhoon-affected areas to generate an adaptive grid.

[0094] Construct multiple grid levels: Based on the adaptive grid, construct a series of gradually coarsening grid levels; define restrictions and prolongation operators between grids. Initialize the wind-wave-current coupling model: read the model parameters from the pre-stored model parameter database; interpolate the corrected wind field data to the finest grid; initialize state variables such as sea surface height, current velocity and wave spectrum.

[0095] For each time step, the following multigrid solution process is performed: several iterations are performed on the finest grid; the residual is calculated and constrained to the coarse grid; the error equation is solved on the coarse grid; the error correction is carried forward to the fine grid; smoothing is performed on the fine grid; the above steps are repeated until the coarsest grid is reached, and then back to the finest grid; convergence conditions are checked and if not converged, the whole cycle is repeated.

[0096] On the finest grid, the wind field, wave and flow field sub-models are solved alternately: the atmospheric boundary layer is updated using the corrected wind field data, taking into account the influence of the sea surface roughness; the wave model is solved, taking into account the influence of wind stress and flow field; the flow field model is solved, taking into account wave-induced stress and wind stress; the sea surface roughness is updated and fed back to the atmospheric model; the above steps are repeated until the preset number of coupling iterations or the convergence threshold is reached.

[0097] Extract key forecast results: extract parameters such as water level, flow velocity and wave height from the coupled model solution; calculate key indicators such as maximum water surge height, maximum wave height and maximum flow velocity; store the forecast results in the coupled model result database.

[0098] This embodiment solves the problem at different scales simultaneously through the multi-grid wind-wave-current coupling model solving technology, thereby improving the computational efficiency and accuracy. The adaptive grid refinement technology can dynamically increase the grid resolution in the nearshore and typhoon-affected areas, while maintaining a relatively coarse grid in the offshore area. This strategy not only ensures the computational accuracy of key areas, but also saves computing resources. For example, when simulating nearshore waves caused by typhoons, this method can increase the resolution to 1 km in the typhoon-affected area while maintaining a global 1-degree resolution, which improves the computational efficiency by about 5 times compared to the traditional fixed grid method, while reducing the root mean square error of the nearshore wave height forecast by about 30%. In addition, the fully coupled simulation of wind-wave-current takes into account the complex nonlinear interactions between the three, and can more accurately describe the ocean dynamic process under typhoon conditions. For example, when simulating the seawater intrusion caused by storm surges, the forecast error of the intrusion distance is reduced by about 20% after considering the wave-current interaction.

[0099] According to one aspect of the present application, step S3 may also be:

[0100] S3a, read the coastline data and water depth data of the study area from the geographic information database; based on the coastline data and water depth data, use SMS software to generate the boundary and unstructured triangular mesh of the calculation domain;

[0101] S3b, interpolating the corrected wind field data onto the unstructured triangular grid to generate a wind stress file;

[0102] S3c. Based on the wind stress file, a finite volume community ocean model is constructed, including a wind-wave-current coupling module. The finite volume community ocean model is used to calculate the nearshore storm surge and typhoon wave forecast data, including water level, current velocity and wave height.

[0103] S3d, based on the nearshore storm surge and typhoon wave forecast data, calculates the flooding range and depth and generates a visual chart.

[0104] In another embodiment of the present application, the coastline data and water depth data of the study area are read from the geographic information database; the SMS (Surface-water Modeling System) software is used to generate an unstructured triangular mesh: a) import the coastline and water depth data; b) define the boundary of the computational domain; c) set the mesh resolution and encrypt the mesh in the nearshore and typhoon-affected areas; d) generate an initial triangular mesh; e) perform mesh quality inspection and optimization to ensure that the mesh meets the requirements of the finite volume community ocean model (FVCOM). The generated mesh data is exported to a format that can be read by FVCOM and stored in the model mesh database.

[0105] Read wind field data from the modified wind field database; interpolate wind field data onto the FVCOM computational grid to generate the wind stress file required by the model; extract outer boundary tidal harmonic constant data from the tidal database; extract initial field and boundary field data, including water level, flow velocity, temperature and salinity, from the ocean circulation model database; read probabilistic typhoon path forecast data from the forecast result database; convert all input data into the NetCDF format required by FVCOM and store them in the model input database.

[0106] Read the basic configuration parameters of FVCOM from the pre-stored model parameter database; set the simulation time step and output time interval; select the appropriate turbulence closure mode and bottom friction coefficient model; configure the wind-wave-current coupling module, including: a) enabling the SWAN (Simulating WAves Nearshore) module for wave simulation; b) setting the wave-current interaction parameters; c) configuring the wind-wave interaction parameters. Write the configuration parameters to the FVCOM run control file.

[0107] Read the generated computational grid; import the prepared input data; load the configured model parameters; initialize the model state variables, including water level, flow velocity, temperature and salinity; for each time step, perform the following operations: a) calculate external forces, including wind stress, tidal force and atmospheric pressure gradient force; b) solve the hydrodynamic equations and update the water level and flow velocity; c) execute the SWAN module to calculate the wave field; d) calculate the wave-current interaction terms and update the momentum equation; e) update the temperature and salinity fields; f) perform numerical stability checks and adjust the time step if necessary. Write the calculation results to the output file at the preset output time interval.

[0108] Read the raw calculation results from the model output file; extract key forecast variables, including water level, flow velocity, significant wave height, and wave period; calculate the maximum flood height: subtract the astronomical tide level from the water level time series; calculate the maximum flood value and occurrence time for each grid point. Calculate the maximum wave height and maximum flow velocity: for each grid point, find the maximum significant wave height and maximum flow velocity during the simulation; record the maximum value and its occurrence time and location. Calculate the flooding range and depth: combine the water level data with the terrain data; determine whether each land grid point is flooded; calculate the flooding depth and range.

[0109] Use professional visualization software (such as Matlab or Python's plotting library) to generate the following charts: a) Spatial distribution of maximum water height; b) Spatial distribution of maximum wave height; c) Vector diagram of maximum flow velocity; d) Inundation range and depth diagram; e) Time series diagram of water level, wave height and flow velocity at key locations. Export the processed data to a common format (such as CSV or NetCDF): create a multidimensional array containing spatial coordinates, time and forecast variables; add metadata information such as variable name, unit and coordinate system; write the data to a file and compress it to reduce the file size. Generate a summary report of forecast results, including extreme value statistics and key findings.

[0110] Store raw model outputs, processed data, and visualization results into the FVCOM results database; add indexes and metadata to the data to facilitate subsequent queries and analysis; and transmit key forecast results to the integrated forecast system for integrated learning and uncertainty quantification.

[0111] Through the above steps, the multi-scale wind-wave-current coupling model constructed by FVCOM has completed the whole process from data preparation, model calculation to result processing. It makes full use of the high-precision weather forecast and corrected wind field data provided by the previous steps, combined with the advantages of FVCOM, to achieve accurate simulation of nearshore storm surges and typhoon waves.

[0112] like Figure 5 As shown, according to one aspect of the present application, step S4 is further:

[0113] S41. Segment the nearshore storm surge and typhoon wave forecast data into time series to obtain segmented data, and divide the segmented data into a training set and a validation set; perform time-frequency analysis based on the nearshore storm surge and typhoon wave forecast data using wavelet transform to obtain wavelet coefficients of different scales; construct feature vectors based on the wavelet coefficients;

[0114] S42, initializing hyperparameters, optimizing hyperparameters by cross-validation method based on feature vectors, and obtaining optimal hyperparameters; training a preconfigured ensemble learning model based on the optimal hyperparameters and training set, and obtaining a trained ensemble learning model; wherein sub-models of the ensemble learning model include a random forest model, a gradient boosting tree model, and a support vector regression model;

[0115] S43. Based on the trained ensemble learning model, the marginal contribution of each feature is calculated using the Shapley value method; the marginal contribution is weighted averaged to calculate the feature importance of each feature; based on the feature importance, the weight of the trained ensemble learning model is optimized to obtain an optimized ensemble learning model;

[0116] S44. Based on the validation set, an optimized integrated learning model is used for prediction to obtain the prediction results of each sub-model; the Bayesian model averaging method is used to integrate the prediction results of each sub-model to obtain the final forecast results, including the maximum water surge height, maximum wave height, maximum flow velocity and their temporal and spatial distribution.

[0117] According to one aspect of the present application, step S43 further includes:

[0118] S43a, based on the optimized ensemble learning model, the Markov chain Monte Carlo method is used to sample the model weights to obtain the sampling results;

[0119] S43b, based on the sampling results, calculate the probability distribution of the prediction; based on the probability distribution, estimate the standard deviation and quantile of the prediction, and quantify the uncertainty of the prediction;

[0120] S43c. Based on uncertainty, calculate the prediction interval at different risk levels.

[0121] In one embodiment of the present application, the calculation results of the nearshore area, including parameters such as water level, flow velocity and wave height, are extracted from the multi-grid wind-wave-current coupling model. The time-frequency analysis of these parameters is performed using wavelet transform to separate the change characteristics of different scales. The separated features are input into an integrated learning model, which integrates algorithms such as random forests, gradient boosting trees and support vector regression. The hyperparameters of each sub-model are optimized by cross-validation, and the feature importance is analyzed using the Shapley value method. Finally, the prediction results of the sub-models are integrated using the Bayesian model averaging method to obtain the final nearshore storm surge and typhoon wave forecast data, including key indicators such as maximum water surge height, maximum wave height, maximum flow velocity and their spatiotemporal distribution, while providing forecast uncertainty estimates for these indicators. Specifically:

[0122] Read forecast data from the coupled model result database; segment the time series of parameters such as water level, flow velocity and wave height to create training and validation sets. Use wavelet transform to decompose the time series of each parameter to obtain wavelet coefficients of different scales; calculate statistical features such as mean, variance, skewness, kurtosis, etc.; construct feature vectors, including wavelet coefficients of different scales, statistical features and original physical parameters.

[0123] Initialize random forest, gradient boosted tree, and support vector regression models; optimize the hyperparameters of each model using grid search and cross validation: define the hyperparameter search space; for each set of hyperparameters, perform K-fold cross validation; select the hyperparameter combination with the best performance on the validation set. Train each model on the full training set using the best hyperparameters.

[0124] For each trained model, calculate the feature importance score; use the Shapley value method to analyze the contribution of the feature to the prediction result: for each feature, calculate its marginal contribution in all possible feature subsets; take a weighted average of the marginal contributions to obtain the Shapley value of each feature; sort the features according to the Shapley value to identify the most important predictors.

[0125] Make predictions for each sub-model (random forest, gradient boosted tree, support vector regression) on the validation set; initialize the Bayesian model averaging method: assume that the posterior probability of each model follows the Dirichlet distribution; set the prior parameters of the Dirichlet distribution. Use the Markov Chain Monte Carlo (MCMC) method to sample model weights: initialize the Markov chain; for each iteration, propose new model weights; calculate the acceptance probability and decide whether to accept the new weights; repeat until the preset number of iterations is reached. Use the sampled weights to calculate the weighted average prediction result: for each prediction point, calculate the weighted average of the sub-model predictions; calculate the confidence interval of the prediction.

[0126] Based on the MCMC sampling results, calculate the probability distribution of the prediction; estimate the standard deviation and quantile of the prediction to quantify the uncertainty of the prediction; calculate the prediction interval under different risk levels (such as 90%, 95%, 99%).

[0127] Integrate the forecast results and uncertainty quantification results; generate forecast values, spatiotemporal distribution and uncertainty estimates of key indicators, including: maximum flood height and its occurrence time and location; maximum wave height and its distribution; maximum flow rate and its distribution; flooding range and depth. Create visualization charts, such as time series, spatial distribution and uncertainty range charts of forecast results; write forecast text descriptions to explain forecast results and potential impacts; store the generated reports in the final forecast results database.

[0128] Determine the users and systems that need to receive forecast results based on the predefined distribution list; generate customized forecast products for different user types (such as emergency management departments, maritime departments, and the public): generate technical detailed reports; generate simplified reports and alerts; generate machine-readable format data. Distribute forecast products through pre-defined communication channels (such as email, API, message queues); record distribution logs, including recipients, distribution time, and distribution content.

[0129] Through the above steps, the full process of nearshore storm surge and typhoon wave forecasting from multi-dimensional feature extraction, data assimilation, coupled model solution to integrated learning and uncertainty quantification is realized. It makes full use of multi-source data and advanced numerical calculation technology to provide high-precision and reliable forecast results and their uncertainty estimation, providing strong support for disaster prevention and mitigation decision-making.

[0130] The integrated learning of the nearshore disaster forecast and uncertainty quantification method in this embodiment not only improves the accuracy of the forecast, but also quantifies the uncertainty of the forecast by integrating the prediction results of multiple sub-models. The introduction of wavelet transform enables the model to capture the changing characteristics on different time scales, especially for the superposition effect of tides and storm surges, and the forecast accuracy is improved by about 15%. The application of the Shapley value method enables this embodiment to analyze the importance of different characteristics to the forecast results, which provides an important basis for further optimizing the observation network and improving the numerical model. For example, it was found through analysis that within 24 hours before the typhoon landed, the contribution of the nearshore wind field to the storm surge height exceeded 60%, which emphasized the importance of improving the observation and forecast accuracy of the nearshore wind field. The Bayesian model averaging method not only gives the optimal forecast results, but also provides a reliable confidence interval, which is crucial for risk assessment and decision-making. In practical applications, this probabilistic forecasting method can extend the warning lead time of extreme events (such as a storm surge that occurs once in 50 years) by about 6-12 hours, while reducing the false alarm rate by about 30%. This embodiment improves the accuracy, efficiency and reliability of nearshore storm surge and typhoon wave forecasts. For example, during the super typhoon landing in 2023, the error between the maximum storm surge height predicted by this embodiment and the actual observation value was controlled within 5 cm, which is about 50% higher than the traditional method. At the same time, due to the use of efficient calculation methods, such as multi-grid technology, the calculation time of the entire forecast process is reduced by about 70%, enabling higher resolution and longer-term forecasts with existing computing resources. In addition, the uncertainty information provided by this embodiment enables decision makers to formulate differentiated response strategies based on different risk tolerances, thereby improving the efficiency and effectiveness of disaster prevention and mitigation. This embodiment improves the accuracy and reliability of nearshore storm surge and typhoon wave forecasts.

[0131] In another embodiment of the present application, a nearshore storm surge and typhoon wave forecasting method based on dual process and data drive is specifically:

[0132] Step S1, obtain multi-source meteorological data, use a dynamic integrated learning framework to process it, and obtain high-precision meteorological forecast data; based on the high-precision meteorological forecast data, use a graph convolutional neural network to process it to obtain typhoon path data; based on the typhoon path data, use a Bayesian deep learning model to process it to obtain probabilistic typhoon path forecast data.

[0133] Step S11, obtain ECMWF data set from meteorological database, including near-surface variables and high-altitude variables; obtain global coastline remote sensing data from OpenStreetMap database; obtain seafloor topography data of a large range of sea areas from GEBCO database; extract water depth point data from S57 nautical chart data, and merge with GEBCO data to form a new data file; obtain the best typhoon path data from the Typhoon Data Center. Input these data into the pre-trained FourCastNet, Pangu, Fuxi and Fengwu four AI meteorological models, and obtain four sets of meteorological element forecast data for the next 24, 48 and 72 hours respectively.

[0134] Step S12: Input the four sets of meteorological element forecast data obtained in step S11 into a dynamic ensemble learning framework based on the attention mechanism. The framework uses the Transformer architecture, takes the output of each model as a "token", and learns the relationship and importance between them through the self-attention mechanism. According to different weather patterns, geographical locations and forecast time periods, the weights of each model are automatically adjusted to obtain comprehensive high-precision meteorological forecast data.

[0135] Step S13, input the high-precision weather forecast data obtained in step S12 into a typhoon path extraction algorithm based on a graph convolutional neural network (GCN). The algorithm regards the meteorological field as a dynamic graph structure, with each grid point as a node in the graph, and the edges between nodes represent spatial and temporal relationships. The spatial structure and temporal evolution of the typhoon system are captured by GCN, and typhoon path data is extracted. The extracted typhoon path data is input into a probabilistic typhoon path forecast model based on Bayesian deep learning, which uses a Bayesian neural network to not only give a point forecast, but also provide an estimate of the uncertainty of the forecast, and finally obtain probabilistic typhoon path forecast data.

[0136] Step S2, obtaining the probabilistic typhoon path forecast data obtained in step S1, and processing it using a deep learning wind field correction model guided by physical information to obtain corrected wind field data;

[0137] Step S21, obtain the probabilistic typhoon path forecast data from step S1, combine the pre-stored terrain data and historical meteorological data, and input them into the deep learning wind field correction model guided by physical information. The model integrates the physical equations of the numerical weather forecast model (such as the Navier-Stokes equation) as constraints, and uses a deep learning network to learn complex nonlinear relationships. The model output results are input into a variational autoencoder (VAE) to process the uncertainty of the observed data and obtain preliminary corrected wind field data.

[0138] Step S22, real-time observation data is obtained from the weather station network, including parameters such as wind speed, wind direction, and air pressure. These real-time observation data are compared with the preliminary revised wind field data obtained in step S21 to calculate the difference value. The difference value is interpolated to the entire calculation domain using the Kriging spatial interpolation algorithm to obtain a correction factor for the spatial distribution. The correction factor is applied to the preliminary revised wind field data to obtain further revised wind field data.

[0139] S3. Based on the corrected wind field data, a multi-scale wind-wave-current coupling model is used to process the data and obtain the nearshore storm surge and typhoon wave forecast data.

[0140] Step S31: Input the further corrected wind farm data obtained in step S22 into a multi-scale wind-wave-current coupling model based on deep learning. The model first uses a convolutional neural network to extract the spatial features of the wind farm data, and then uses a long short-term memory network (LSTM) to capture the time series features. The extracted features are input into a multi-layer perceptron to learn the nonlinear mapping relationship between models of different scales (global, regional, local). At the same time, a graph neural network (GNN) is used to process data on unstructured grids to improve computational efficiency and accuracy.

[0141] Step S32, extract the calculation results of the nearshore area from the multi-scale wind-wave-current coupling model of step S31, including parameters such as water level, flow velocity and wave height. Use wavelet transform to perform time-frequency analysis on these parameters and separate the change characteristics of different scales. Input the separated features into an integrated learning model, which integrates algorithms such as random forest, gradient boosting tree and support vector machine. Through integrated learning, the final nearshore storm surge and typhoon wave forecast data are obtained, including key indicators such as maximum water surge height, maximum wave height, maximum flow velocity and their spatiotemporal distribution.

[0142] In another embodiment of the present application, a method for rapid forecasting of nearshore storm surge and typhoon wave based on dual-drive of process and data, the specific steps are as follows:

[0143] Step 1: Collect information.

[0144] ① Collect ECMWF data sets, including near-surface variables and upper-air variables. Among them, near-surface variables include: sea level pressure (MSLP), 10m zonal wind (U10), 10m meridional wind (V10) and 2m temperature (T2M), with a resolution of 0.25°; upper-air variables include: geopotential (Z), specific humidity (Q), temperature (T), zonal wind (U) and meridional wind (V), with a resolution of 0.25°, and are divided into 13 altitude layers, namely 1000hPa, 925hPa, 850hPa, 700hPa, 600hPa, 500hPa, 400hPa, 300hPa, 250hPa, 200hPa, 150hPa, 100hPa and 50hPa.

[0145] ② Collect the latest OpeanStreetMap global coastline remote sensing data;

[0146] ③ Collect seabed topography data GEBCO (General Bathymetric Chart of the Oceans) in a large area of ​​sea; collect nearshore chart data, extract water depth point data from S57 chart data, and merge it with GEBCO data to form a new data file *.xyz.

[0147] ④ Collect the best typhoon path data;

[0148] Step 2: Forecast the meteorological elements for the next 24, 48 and 72 hours based on different AI meteorological models (such as FourCastNet, Pangu, Fuxi, Fengwu, etc.).

[0149] Four different AI meteorological models, FourCastNet, Pangu, Fuxi and Fengwu, are used to forecast meteorological elements for the next 24, 48 and 72 hours (output results every 6 hours), and the prediction results of the four models on meteorological elements are obtained.

[0150] FourCastNet (Fourier Prediction Neural Network) is a data-driven global weather forecast model that can achieve a week-long high-resolution (0.25 degree) weather forecast in 2 seconds, which is 45,000 times faster than numerical forecast methods. FourCastNet combines Fourier Neural Operator (FNO) to reduce computational complexity, achieve high-resolution prediction, and use the "pre-training + fine-tuning" paradigm to expand to downstream tasks (precipitation forecast). To achieve high-resolution prediction, FourCastNet uses the Adaptive Fourier Neural Operator (AFNO) model. The model network architecture is designed for high-resolution input, with ViT as the backbone network and combined with Fourier Neural Operator (FNO). The model learns mappings between function spaces to solve a series of nonlinear partial differential equations.

[0151] The Pangu Weather Model replaces the dynamic core and parameterization scheme in the traditional numerical model (NWP) with a purely data-driven AI method, which exceeds the traditional numerical model in forecasting effect and is 10,000 times faster. The Pangu Weather Model is the first AI method with higher accuracy than traditional numerical forecasting methods, and its 1-hour to 7-day forecast accuracy is higher than the European Meteorological Center's numerical model (ECMWF operational IFS). The forecast includes potential, humidity, wind speed, temperature, etc.; the horizontal spatial resolution is 0.25°×0.25°, the temporal resolution is 1 hour, covering 13 vertical layers, and can accurately predict fine-grained meteorological characteristics.

[0152] The Fuxi model uses AI algorithms to achieve global weather forecasts for the next 15 days. The forecast accuracy reaches the ensemble average of the European Centre for Medium-Range Weather Forecasts (ECMWF: Atmospheric Model Ensemble 15-dayforecast) for the first time in the evaluation set, significantly ahead of the European Centre for Deterministic Forecasts (ECMWF: Atmospheric ModelHigh resolution 10-day forecast), and the speed of generating results is also a thousand times faster than traditional models.

[0153] FengWu-GHR (Global High Resolution) is built based on multimodal and multi-task deep learning methods. The AI ​​large model "FengWu" has achieved for the first time an effective forecast of core atmospheric variables at high resolution for more than 10 days, and surpassed the model GraphCast released by DeepMind on 80% of the evaluation indicators. In addition, "FengWu" can generate high-precision global forecast results for the next 10 days in just 30 seconds, which is much more efficient than traditional models. On March 1, 2024, the global high-resolution AI weather forecast large model "FengWu-GHR" (Global High Resolution) was released, which for the first time used AI to achieve 10-kilometer-level modeling and forecasting of medium-term weather.

[0154] The prediction process of the Pangu meteorological model is as follows: both training and testing are carried out on the ECMWF fifth-generation reanalysis data (ERA5), which includes 43 years (1979-2021) of global real-time meteorological data. Among them, the data from 1979 to 2017 are used as training sets, the data from 2019 are used as validation sets, and the data from 2018, 2020, and 2021 are used as test sets.

[0155] The large model input data includes ground variables and high-altitude variables. Among them, the near-surface variables include: sea level pressure (MSLP), 10m zonal wind (U10), 10m meridional wind (V10) and 2m temperature (T2M), with a resolution of 0.25°; the high-altitude variables include: potential (Z), specific humidity (Q), temperature (T), zonal wind (U) and meridional wind (V), with a resolution of 0.25°, and are divided into 13 altitude layers, namely 1000hPa, 925hPa, 850hPa, 700hPa, 600hPa, 500hPa, 400hPa, 300hPa, 250hPa, 200hPa, 150hPa, 100hPa and 50hPa.

[0156] The output data of the large model have the same variables as the input data, namely ground variables and high-altitude variables, but the time is 6, 12, 18, ..., 72 hours in the future.

[0157] The input data includes near-surface data and high-altitude data, which are converted from the ERA5 data downloaded in step 1. The model code includes:

[0158] (1) input_surface.npy: near-surface variables (the variables are MSLP, U10, V10, T2M)

[0159] (2) input_upper.npy: upper altitude variables (variables are Z, Q, T, U and V in order)

[0160] (3) pangu_weather_6.onnx: A deep learning model file in ONNX format for weather forecasting. The model may predict weather conditions for the next 6 hours.

[0161] (4) pangu_weather_24.onnx: Model file in ONNX format, used for 24-hour weather forecast.

[0162] (5) pangu_weather_1.onnx: Model file in ONNX format, used for 1-hour short-term weather forecast.

[0163] (6) pangu_weather_3.onnx: Model file in ONNX format, used for 3-hour weather forecast.

[0164] (7) my_input_surface.npy: User-defined surface data file, but contains specific custom data or data at different time points.

[0165] (8) my_input_upper.npy: User-defined upper atmosphere data file, containing specific custom data or data at different time points.

[0166] (9) topography.npy: Contains topographic data, which is very important for calculating the effects of topography in meteorological models.

[0167] (10) land_mask.npy: Contains land mask data, which is used to distinguish between land and water bodies, helping the model to correctly handle different surface types during calculations.

[0168] (11) soil_type.npy: Contains soil type data, describing the soil types and their characteristics in different regions. This information can affect the hydrological and thermal processes on the surface.

[0169] like Figure 6As shown in the figure, the accuracy of the surface meteorological variables 2m temperature (T2M), 10m longitude and latitude wind speed (U10 and V10) is tested, which is consistently higher than the European Meteorological Center's operational IFS and NVIDIA's FourCastNet. Using the "forecast time gain" to measure, the accuracy improvement of the Pangu Meteorological Model relative to the operational IFS is equivalent to more than 18 hours of "forecast time gain". For example, for 2m temperature, the 3-day and 5-day operational IFS forecast root mean square errors are 1.34 and 1.75, the FourCastNet root mean square errors are 1.39 and 2.00, and the Pangu Meteorological Model reduces the error to 1.05 and 1.53. For the wind speed in the longitude direction of 10 meters, the 3-day and 5-day operational IFS forecast root mean square errors are 1.94 and 2.90, the root mean square errors of FourCastNet are 2.24 and 3.41, and the Pangu meteorological model reduces the errors to 1.61 and 2.53. The AI ​​meteorological model can accurately predict refined meteorological features, while increasing the prediction speed by 10,000 times, and can provide global weather forecasts in seconds.

[0170] Step 3: Calculate the typhoon path from the predicted meteorological element field.

[0171] The predicted meteorological element results do not include the typhoon path, and the typhoon eye position at different times needs to be extracted based on the meteorological element results. This step is based on the prediction results of four different meteorological models in step 2, and four predicted typhoon paths can be extracted.

[0172] A typhoon is a rotating structure with low pressure at the center, so the position of the typhoon eye can be calculated, determined and extracted from the two aspects of air pressure and rotation. If the above two conditions are met, it can be determined as the typhoon center.

[0173] The lowest point of sea level pressure: The center of a typhoon is usually located near the lowest point of sea level pressure. Therefore, the longitude and latitude corresponding to the lowest point of sea level pressure are extracted.

[0174] Calculate the maximum value of the 850 hPa curl: The typhoon center can also be determined by the maximum value of the positive curl of the 850 hPa layer. The curl ζ=(Ψv / Ψx)-(Ψu / Ψy), where u and v are the east-west and north-south components of the wind speed, respectively. Calculate ζmax.

[0175] Step 4: Based on the prediction results of different large meteorological models, an integrated forecasting method is used to obtain the best predicted typhoon path.

[0176] In step three, four different AI meteorological models were used to calculate and extract four typhoon path data sets. For different typhoon paths or at different times, the forecast accuracy of the typhoon paths in the four typhoon data sets is competitive. The main work of this step is to "select the best", that is, to adopt the idea of ​​​​integrated forecasting and select the point closest to the true value in the four predicted typhoon paths as the prediction result. This step uses the LightGBM machine learning model, and uses the four predicted typhoon path data sets (input variable X) and the CMA optimal path data set (output variable Y) as training data to further predict the typhoon path.

[0177] The input data structure includes: typhoon path information, which is used to identify different typhoon paths; the typhoon path prediction results of the FourCastNet large model, including the longitude and latitude of the typhoon path, the lowest sea surface pressure and the maximum wind speed at 10 m above the sea surface; the typhoon path prediction results of the Fuxi large model, including the longitude and latitude of the typhoon path, the lowest sea surface pressure and the maximum wind speed at 10 m above the sea surface; the typhoon path prediction results of the Pangu large model, including the longitude and latitude of the typhoon path, the lowest sea surface pressure and the maximum wind speed at 10 m above the sea surface; the typhoon path prediction results of the Fengwu large model, including the longitude and latitude of the typhoon path, the lowest sea surface pressure and the maximum wind speed at 10 m above the sea surface; the typhoon optimal path, that is, the true value of the typhoon path, including the longitude and latitude of the typhoon path, the lowest sea surface pressure and the maximum wind speed at 10 m above the sea surface; the typhoon path prediction results of the large model, including the longitude and latitude of the typhoon path, the lowest sea surface pressure and the maximum wind speed at 10 m above the sea surface. The output data is the best predicted typhoon path, including the longitude and latitude of the typhoon center, the lowest sea surface pressure and the maximum wind speed at 10m above the sea surface.

[0178] Step 5: Generate the best predicted typhoon wind field based on the best predicted typhoon path.

[0179] Based on the best predicted typhoon path in step 4, find the typhoon path sampling point with the smallest error among the predicted typhoon paths of the four large models, and extract the wind pressure field corresponding to the sampling point as the best predicted typhoon wind field.

[0180] Step 6: Combined wind field wind speed correction based on multi-station measured wind speed. Typhoons near the coast are affected by changes in the underlying surface, and wind speed and direction will change. This step corrects the typhoon-driven wind field based on the measured wind data near the coast.

[0181] Based on the wind speed measured at eight coastal meteorological stations, the inverse distance square method is used to correct the wind field within a radius of 500 kilometers (ISYE-R500) from the target area. The main methods are:

[0182] W cor (x,y,t)=k W (x, y, t)W(x, y, t);

[0183] ;

[0184] Where W cor (x, y, t) is the corrected wind speed at the point with coordinates (x, y) in the wind field at time t, W is the wind speed of the combined wind field before correction, k W is the wind speed correction factor. obs (x i ,y i , t) is the wind speed observation value of the i-th station among the n meteorological stations at time t, d i (x, y) is the distance between the point to be corrected and the i-th meteorological station, d i (x, y) = (x - x i ) 2 +(y- y i ) 2 The wind speed correction coefficient k at the meteorological station in the combined wind field W is the ratio of the measured wind speed at the site to the wind speed of the combined wind field. The wind speed correction coefficients at other locations are calculated by the inverse distance square interpolation method of the correction coefficients at n measured meteorological sites.

[0185] Step 7: Establish a wind-wave-current coupling mathematical model to calculate the hourly water level and effective wave height in the study area. Based on the topographic coastline data in step 1, the open source three-dimensional finite volume method ocean numerical model FVCOM is used to construct a wind-wave-current coupling mathematical model. Among them, the outer sea boundary is driven by a harmonic constant, and the external force input to the model is the best predicted typhoon wind pressure data of the AI ​​large model in the previous step.

[0186] The preferred embodiments of the present invention are described in detail above; however, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. A nearshore storm surge and typhoon wave forecasting method based on process and data dual drive, characterized in that: The steps include: S1. Obtain multi-source meteorological data, and process them using a multi-model adaptive integrated forecasting method to obtain high-precision meteorological forecast data; based on the high-precision meteorological forecast data, use a multi-dimensional feature typhoon path extraction and probabilistic forecasting method to process them to obtain probabilistic typhoon path forecast data; the multi-source meteorological data includes the European Center for Medium-Term Weather Forecasts data set, global coastline remote sensing data, seafloor topography data, nearshore chart data, and optimal typhoon path data; S2. Based on the probabilistic typhoon path forecast data, the physically constrained variational assimilation wind field correction method is used to obtain the corrected wind field data; S3. Based on the corrected wind field data, the multi-grid wind-wave-current coupling model solution method is used to obtain the nearshore storm surge and typhoon wave forecast data, including water level, flow velocity and wave height; S4. Based on the nearshore storm surge and typhoon wave forecast data, the integrated learning nearshore disaster forecast and uncertainty quantification method are used to calculate the final forecast results; Step S1 is further as follows: S11. Obtain the European Centre for Medium-Term Weather Forecasts dataset from the meteorological database, including near-surface variables and upper-air variables; obtain global coastline remote sensing data from the OpenStreetMap database; obtain seafloor topography data from the GEBCO database, extract water depth point data from the S57 nautical chart data, fuse the water depth point data with the seafloor topography data to form a fused data file; obtain the best typhoon path data from the Typhoon Data Center; store the European Centre for Medium-Term Weather Forecasts dataset, global coastline remote sensing data, fused data file and best typhoon path data as multi-source meteorological data; S12, inputting the multi-source meteorological data into the pre-trained Fourier prediction neural network, Pangu, Fuxi and Fengwu AI meteorological models respectively, to obtain four sets of future meteorological element forecast data; S13, based on the meteorological element forecast data, using the extended Kalman filter and introducing the time decay factor, dynamically estimate the forecast error covariance of each model; Based on the forecast error covariance, the weights of each model are adaptively adjusted to obtain comprehensive high-precision weather forecast data; S14. Based on high-precision meteorological forecast data, the main characteristic modes of the pressure field and wind field are extracted using an improved principal component analysis method; based on the main characteristic modes, the time series of the typhoon center position is constructed using the least squares spline fitting technique; based on the time series, the particle filter algorithm is used to simulate possible typhoon path samples; based on the pre-acquired observation data, the weights of the typhoon path samples are updated to obtain the probability distribution of the typhoon path and form probabilistic typhoon path forecast data; Step S2 is further as follows: S21. Based on the probabilistic typhoon path forecast data, combined with pre-stored terrain data and historical meteorological data, a four-dimensional variational data assimilation method is used to calculate the wind field analysis results; S22, obtain real-time observation data from the weather station network, including wind speed, wind direction and air pressure; Compare the real-time observation data with the wind field analysis results and calculate the innovation vector; S23, using the optimal interpolation method, interpolating the innovation vector to the entire preset calculation domain to obtain a spatially distributed correction field; S24, superimposing the correction field and the wind field analysis result to obtain the final corrected wind field data; Step S4 is further as follows: S41. Segment the nearshore storm surge and typhoon wave forecast data into time series to obtain segmented data, and divide the segmented data into a training set and a validation set; perform time-frequency analysis based on the nearshore storm surge and typhoon wave forecast data using wavelet transform to obtain wavelet coefficients of different scales; construct feature vectors based on the wavelet coefficients; S42, initializing hyperparameters, optimizing the hyperparameters by cross-validation based on the feature vectors, and obtaining optimal hyperparameters; training a preconfigured ensemble learning model based on the optimal hyperparameters and the training set, and obtaining a trained ensemble learning model; The sub-models of the integrated learning model include random forest model, gradient boosting tree model and support vector regression model; S43. Based on the trained ensemble learning model, the marginal contribution of each feature is calculated using the Shapley value method. Take a weighted average of the marginal contributions and calculate the feature importance of each feature; Based on the feature importance, the weight of the trained ensemble learning model is optimized to obtain an optimized ensemble learning model; S44. Based on the validation set, the optimized integrated learning model is used for prediction to obtain the prediction results of each sub-model; the prediction results of each sub-model are integrated using the Bayesian model averaging method to obtain the final forecast results, including the maximum water surge height, maximum wave height, maximum flow velocity and their temporal and spatial distribution; In step S14, based on high-precision meteorological forecast data, an improved principal component analysis method is used to extract the main characteristic modes of the pressure field and the wind field; based on the main characteristic modes, the least squares spline fitting technology is used to construct the time series of the typhoon center position further as follows: S141, performing wavelet transform processing on the high-precision weather forecast data to obtain time-frequency features and spatial scale features; combining the time-frequency features and the spatial scale features to construct a three-dimensional tensor; S142, decomposing the three-dimensional tensor into a core tensor and a factor matrix using the Tucker decomposition method; based on the core tensor and the factor matrix, applying a preset adaptive threshold to each pattern to perform feature screening, and obtaining a screened core tensor and factor matrix; Reconstruct the tensor using the filtered core tensor and factor matrix to obtain the feature representation after dimensionality reduction; S143, calculating the minimum air pressure and the maximum positive vorticity position at the sea level based on the feature representation after dimensionality reduction; Based on the minimum pressure and the maximum positive vorticity position, the candidate location of the typhoon center is determined; Based on the candidate location of the typhoon center, the time series of the typhoon center location is constructed using the least squares spline fitting method; Step S43 also includes: S43a, based on the optimized ensemble learning model, the Markov chain Monte Carlo method is used to sample the model weights to obtain the sampling results; S43b, based on the sampling results, calculate the probability distribution of the prediction; based on the probability distribution, estimate the standard deviation and quantile of the prediction, and quantify the uncertainty of the prediction; S43c, based on uncertainty, calculate the prediction interval at different risk levels; Step S21 is further as follows: S211, constructing a background field based on the probabilistic typhoon path forecast data in combination with pre-stored terrain data and historical meteorological data; S212, reading observation data from the observation database, and constructing a background error covariance matrix and an observation error covariance matrix based on the background field and the observation data to form an error function; S213. Using the conjugate gradient method to minimize the error function, obtain the wind field analysis result.

2. The nearshore storm surge and typhoon wave forecasting method based on process and data dual drive according to claim 1 is characterized in that: Step S3 is further as follows: S31. Obtain the terrain and shoreline data of the study area, and use the unstructured grid generation algorithm to divide the calculation domain; based on the calculation domain, use the adaptive grid refinement technology to increase the grid resolution and generate an adaptive grid; S32. Based on the adaptive grid, a gradually coarsening grid hierarchy is constructed, and restriction operators and extension operators between grids are defined; S33, based on the corrected wind field data, performing numerical iteration on the finest grid in the grid hierarchy to obtain updated state variables; Calculate fine grid residuals based on updated state variables and expected state variables; The fine grid residual is transferred to the coarse grid using a restriction operator to obtain the coarse grid residual; based on the coarse grid residual, an error equation is constructed and solved to obtain the coarse grid error correction value; The coarse grid error correction value is transferred to the fine grid using the continuation operator to obtain the fine grid error correction value; S34, based on the fine grid error correction value and the updated state variable, a smoothing algorithm is used to obtain a smoothed state variable; steps S33 to S34 are repeated until a preset convergence condition is reached, and the final state variable is output; S35. Based on the final state variables and the corrected wind field data, a separation solver is used to alternately solve the preconfigured wind field, wave and flow field sub-models to obtain nearshore storm surge and typhoon wave forecast data.

3. The method for forecasting nearshore storm surge and typhoon wave based on process and data dual drive according to claim 2 is characterized in that: The method also includes performing quality control on the real-time observation data of step S22, specifically: S22a, marking the data in the real-time observation data that exceeds the preset physical data range as abnormal data; S22b, calculating the first-order difference of the time series based on the real-time observation data; Based on the first-order difference, a time continuity check is performed to obtain data with abnormal jumps, which are marked as abnormal data; S22c, based on the real-time observation data, perform a spatial consistency check, obtain inconsistent data, and mark it as abnormal data; S22d. Remove the abnormal data from the real-time observation data and record it in the quality control log, and retain the real-time observation data that excludes the abnormal values.

4. The method for forecasting nearshore storm surge and typhoon wave based on process and data dual drive according to claim 2 is characterized in that: Step S3 may also be: S3a, read the coastline data and water depth data of the study area from the geographic information database; based on the coastline data and water depth data, use SMS software to generate the boundary and unstructured triangular mesh of the calculation domain; S3b, interpolating the corrected wind field data onto the unstructured triangular grid to generate a wind stress file; S3c. Based on the wind stress file, a finite volume community ocean model is constructed, including a wind-wave-current coupling module. The finite volume community ocean model is used to calculate the nearshore storm surge and typhoon wave forecast data, including water level, current velocity and wave height. S3d, based on the nearshore storm surge and typhoon wave forecast data, calculates the flooding range and depth and generates a visual chart.

Images