A coastal disaster forecasting method and system

Abstract

The application discloses a kind of coastal disaster forecasting method and system;Wherein the method obtains the data of target nearshore area and carries out data preprocessing and unstructured grid division;Coupling numerical model is used to generate "offshore driving-nearshore response" sample library;Through mutual information analysis, the dominant lag step number of each offshore factor on each node in the nearshore is extracted, and then the time rearrangement of key factor is carried out, and is spliced with static geographic features, to construct structured input tensor;With this, the graph attention spatio-temporal fusion model with non-structural grid as graph structure is trained;In real-time prediction, the implicit state assimilation mechanism based on observation error is introduced, and the model state is dynamically corrected to realize adaptive rolling prediction.The application realizes high-precision, high-efficiency nearshore disaster real-time early warning.

Description

Technical Field

[0001] This invention relates to nearshore disaster forecasting technology, and more particularly to a coastal disaster forecasting method and system. Background Technology

[0002] Refined disaster forecasting in nearshore waters is crucial for marine disaster prevention and mitigation. Currently, mainstream methods mainly rely on numerical simulation, which is computationally time-consuming and cannot meet the real-time rolling forecasting requirements under sudden disasters. Traditional data-driven methods are mostly based on regular grid construction, which is difficult to accurately fit the complex and winding coastlines and unstructured topography of nearshore areas, and fails to explicitly model the physical time delay of offshore driving signals, resulting in limited prediction accuracy and generalization ability.

[0003] Publication No. CN118940021A discloses a nearshore storm surge and typhoon wave forecasting method based on process and data dual-driven approaches. This method employs a "process and data dual-driven" architecture, using multi-model integration, wind field assimilation, and multi-grid coupled numerical models for forecasting. While the physical processes are complete, the computational cost is high, making minute-level rolling updates difficult. Publication No. CN119293745 discloses a data-driven nearshore wind and wave forecasting method and system. This method proposes a purely data-driven TCN-CNN hybrid model, utilizing cross-correlation (CCF) and dynamic time warping (DTW) to analyze time lag effects. However, its grid is the result of structured grid partitioning, which cannot accurately represent complex shoreline boundaries. Furthermore, its correction mechanism relies on periodic retraining of model parameters, making real-time state assimilation during the forecasting process impossible.

[0004] Therefore, there is an urgent need to develop a technical solution that can directly adapt to unstructured grids, establish an efficient and accurate mapping from offshore site-driven to nearshore full-field response, and achieve real-time assimilation and correction in rolling forecasts. Summary of the Invention

[0005] To address the shortcomings of existing technologies in the background section, this invention provides a coastal disaster forecasting method and system. The method can directly model on unstructured grids, utilize prior information from offshore stations, and update the implicit state of the model through real-time observation, thereby achieving high-resolution, high-precision, and high-efficiency rolling forecasting of nearshore disasters.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] In a first aspect, the present invention provides a method for forecasting coastal disasters, comprising:

[0008] S1: Acquire static geographic information, historical dynamic meteorological and oceanographic observation data, and multi-factor meteorological and hydrodynamic time-series data of offshore stations for the target nearshore area;

[0009] S2: Construct an unstructured grid covering the target nearshore area based on the static geographic information, construct a coupled numerical model of astronomical tide-storm surge-wave based on the nearshore unstructured grid, and then generate a reference dynamic sample library containing the correspondence between offshore driving and nearshore response.

[0010] S3: Time alignment, outlier removal and standardization are performed on the dynamic meteorological and oceanographic observation data in the reference dynamic sample library and the multi-factor meteorological and hydrodynamic time series data of offshore stations.

[0011] S4 employs a mutual information-based dependency analysis method to calculate the mutual information of the response of offshore site driving factors to each grid node in the target nearshore area under different time lag steps, and to determine the set of key driving factors and their dominant lag steps for each grid node.

[0012] S5: Based on the dominant lag step of each driving factor in the key driving factor set, the offshore driving factors are time-rearranged and spliced ​​with nearshore static geographic features to construct a structured input tensor that integrates spatiotemporal information.

[0013] S6: Using the structured input tensor as input, and the nearshore response data generated by the corresponding astronomical tide-storm surge-wave coupled numerical model in the reference dynamic sample library as labels (i.e., nearshore two-dimensional disaster field data as labels), supervise the training of the spatiotemporal fusion deep learning model based on graph attention network to obtain the coastal disaster prediction model.

[0014] S7: Acquire real-time meteorological and oceanographic observation data from nearshore stations and real-time multi-factor meteorological and hydrodynamic time-series data from offshore stations. Process the offshore station data into real-time structured input tensors using S3-S5 and input them into the pre-trained nearshore disaster forecasting model for rolling forecasts. Simultaneously, reconstruct the error field using the current real-time meteorological and oceanographic observation data from nearshore stations and map it as the correction increment of the model's hidden state to update the model state and serve as the starting point for subsequent forecasts.

[0015] Furthermore, in S1, the static geographic information includes water depth, shoreline vector, and seabed type; the meteorological and oceanographic observation data includes tide level, significant wave height, wave period, storm surge, and average wave direction collected by tide gauge stations, wave buoys, and meteorological observation stations in the target nearshore area; and the multi-factor time-series data of the offshore stations includes significant wave height, wave direction, wind speed, wind direction, air pressure, water level, air temperature and sea surface temperature, swell height, and spectral peak period.

[0016] Furthermore, the specific process of S2 is as follows:

[0017] S2.1: Based on the shoreline and water depth in the acquired static geographic information, the target nearshore area is divided into unstructured grids using SMS software;

[0018] S2.2: A coupled numerical model of astronomical tides, storm surges, and ocean waves is used for simulation calculations. Typical historical typhoon disaster events affecting the nearshore area of ​​the target region during a preset period are selected as simulation objects, and the optimal typhoon path and reanalysis meteorological field during the event are used as driving forces. Meteorological and oceanographic observation data (time series of measured tide levels at tide gauge stations and significant wave height data measured by wave buoys) collected in S1 are compared and verified with the corresponding time series simulated by the model. The key parameters of the model are iteratively adjusted until the error between the simulation results and the measured data is lower than the preset verification threshold.

[0019] S2.3: Based on the meteorological and oceanographic observation data collected in S1, statistical analysis is performed and the marginal probability distributions of each factor are fitted respectively; the joint probability distribution function of each factor is constructed using the Copula function; random sampling is performed from the joint distribution using the Monte Carlo method to generate multiple sets of virtual disaster parameter combinations; and each set of virtual parameters is transformed into a time-varying two-dimensional boundary for calculating the disaster field covering the target nearshore area, thereby driving the astronomical tide-storm surge-wave coupled numerical model to output a spatiotemporal continuous two-dimensional grid field of hydrodynamic factors (offshore station driving factors); and then integrating them to obtain a reference dynamic sample library of offshore driving-nearshore full-field response-mapping relationship.

[0020] Furthermore, the specific process of S4 is as follows:

[0021] S4.1: Hydrodynamic response sequence for unstructured grid node g in the target nearshore area and offshore site driving factors The lag time of the driving factors of offshore sites is discretized into lag steps τ, and time-shifted sample pairs are constructed. The Parzen window method is used to estimate the marginal probability density of the response sequence and the offshore site, as well as their joint probability density, and then the driving factor of the offshore site under the lag step τ is calculated. Mutual information of the response of unstructured grid nodes g in the nearshore area of ​​the target;

[0022] S4.2: The Monte Carlo scrambling method is used for testing, and within the preset lag range that passes the test, the peak value of the mutual information curve is searched, and the time corresponding to the peak value is determined as the offshore driving factor. The dominant hysteresis steps for unstructured mesh node g ;

[0023] S4.3: After determining the dominant lag steps of all offshore driving factors, the minimum redundancy maximum correlation criterion is used to select the set of key driving factors for each grid node g. .

[0024] Furthermore, offshore site driving factors Mutual information of the response of unstructured grid nodes g in the nearshore region of the target The calculation formula is as follows:

[0025] ;

[0026] in, This represents the current sequence of driving factors for offshore sites after the dominant lag shift. Represents the hydrodynamic time series of the response of an unstructured grid node g; and They represent and The marginal probability density; express and The joint probability density.

[0027] Furthermore, the process of selecting the key driving factor set is as follows: Initialize the key driving factor set S as an empty set; employ a forward search strategy; and in each iteration, calculate each unselected offshore driving factor x. j Overall rating The comprehensive scores are sorted in descending order, and the offshore driving factors that meet the preset threshold or preset number of comprehensive scores are included in the key driving factor set.

[0028] Overall score The calculation formula is as follows:

[0029] ;

[0030] in, yg(t) represents the current sequence of driving factors for offshore sites after dominant hysteresis translation; yg(t) represents the response sequence of nearshore grid nodes; S represents the set of key driving factors that have been selected before the current iteration step. S represents the sequence of driving factors for a certain offshore site that already exists in set S; |S| represents the number of driving factors for offshore sites that have been selected in set S; I represents the mutual information calculation function.

[0031] Furthermore, the specific process of S5 is as follows:

[0032] S5.1: For the current mesh node From the set of key driving factors Extracting the driving factors of each offshore site And obtain the dominant lag steps corresponding to the offshore driving factors. ;

[0033] S5.2: For each driving factor According to its dominant lag step number Shift the corresponding dominant lag step on the time axis. Obtain the aligned timing sequence This aligns the historical driving signals from the open sea with the current response signals from the nearshore area on a discrete index.

[0034] S5.3: Align the timing of each offshore driving factor With nodes Static geographic feature vectors Perform vector concatenation to generate the node at time [time]. Structured input tensors :

[0035] ;

[0036] Where [ ] denotes vector concatenation operation; This indicates traversing set S. g The driving factors of all offshore stations are taken from the discrete index. Observations at; This indicates the grid node. The static geographic feature vectors include water depth and bottom friction coefficient, etc.

[0037] Furthermore, the processing procedure for each training sample in the spatiotemporal fusion deep learning model of the graph attention network in S6 includes:

[0038] S6.1: Structured input tensors that incorporate spatiotemporal information Through a linear transformation layer and activation function, it is mapped to a high-dimensional node-driven feature vector. ;

[0039] S6.2: Using an unstructured mesh as the graph structure, compute the central nodes using the graph attention mechanism. Its first-order neighbor set Each neighbor node attention weights This leads to the understanding of the historical implicit states of neighboring nodes. Perform a weighted summation to generate the spatial context feature vector of the central node. ;

[0040] S6.3: Features of offshore driving Spatial context features The data is concatenated and input into a gated recurrent unit (GRU); the GRU then combines this data with the hidden state of the node from the previous time step. Calculate and output the new hidden state at the current time. ;

[0041] S6.4: Update the implicit state Through an output layer (linear transformation layer), the value of the predicted physical quantity (such as water level) of the node at the current moment is decoded; by traversing all grid nodes, the predicted nearshore two-dimensional disaster field at the current moment is obtained.

[0042] Furthermore, the specific process of S7 is as follows:

[0043] S7.1: Acquire real-time meteorological and oceanographic observation data of the target nearshore stations and real-time multi-factor meteorological and hydrodynamic time-series data of offshore stations. Process the offshore station data into real-time structured input tensors using S3-S5 and input them into the pre-trained nearshore disaster forecasting model for rolling forecasts. At the same time, use the real-time meteorological and oceanographic observation data of the target nearshore at the current moment to calculate the deviation vector between the measured value and the model's forecast value at the corresponding station location.

[0044] S7.2: If the norm of the deviation vector is less than a predetermined threshold, the near-shore two-dimensional disaster field at the current forecast time is directly output; if it is greater than a preset error threshold, the inverse distance weighted spatial interpolation algorithm is used to calculate the uncorrected increment Er of each grid node g in the entire field, centered on the observation station k, to generate an error field with all near-shore grid nodes; wherein, the formula for calculating the uncorrected increment Er is:

[0045] ;

[0046] in, Indicates the number of nearshore observation stations; Represents the measured values ​​at the observation station; This represents the predicted value indicating the location of the model's observation stations; This represents spatial weight; the greater the distance, the smaller the impact of the correction bias.

[0047] S7.3: Input the error field into a pre-trained feedback encoder network, which maps the spatial error field into a state correction increment vector of the same dimension as the model's latent states. This instantly updates the current hidden state of all nodes in the model;

[0048] S7.4: Update the implicit state Replace the original Starting from the updated state, the model decoder is re-executed or the calculation continues until the next forecast time. The corrected forecast field for the current time is output, and continuous rolling forecasts for subsequent times are carried out based on this.

[0049] Secondly, the present invention provides a coastal disaster forecasting system, the system being used to perform the method described above, including:

[0050] Data acquisition and preprocessing module: Acquires static geographic information of the target nearshore area, historical dynamic meteorological and oceanographic observation data, and multi-factor meteorological and hydrodynamic time series data of offshore stations; used to perform time alignment, outlier removal, and standardization processing on the dynamic meteorological and oceanographic observation data in the reference dynamic sample library and the multi-factor meteorological and hydrodynamic time series data of neighboring offshore stations.

[0051] Numerical model calculation module: used to construct an unstructured grid covering the target nearshore area based on the static geographic information, construct a coupled numerical model of astronomical tide-storm surge-wave based on the nearshore unstructured grid, and then generate a reference dynamic sample library containing the correspondence between offshore driving and nearshore response;

[0052] The lag factor identification module uses a dependency analysis method based on mutual information to calculate the mutual information of the response of the driving factors of offshore stations to each grid node in the target nearshore area under different time lag steps, and determines the set of key driving factors and its dominant lag step number corresponding to each unstructured grid node.

[0053] Structured Input Tensor Construction Module: Based on the dominant lag steps of each driving factor in the set of key driving factors, the offshore driving factors are time-rearranged and spliced ​​with nearshore static geographic features to construct a structured input tensor that integrates spatiotemporal information.

[0054] The graph neural network model inference and online assimilation and correction module is used to take a structured input tensor as input and the nearshore response data generated by the corresponding astronomical tide-storm surge-wave coupled numerical model as labels (i.e., nearshore two-dimensional disaster field data as labels) to supervise the training of a spatiotemporal fusion deep learning model based on graph attention network, thereby obtaining a coastal disaster forecasting model. It is also used to acquire real-time meteorological and oceanographic observation data from target nearshore stations and real-time multi-factor meteorological and hydrodynamic time-series data from offshore stations, process the offshore station data into a real-time structured input tensor, and input it into the trained nearshore disaster forecasting model for rolling forecasts. Simultaneously, it reconstructs the error field using the current real-time meteorological and oceanographic observation data of the target nearshore area and maps it to the correction increment of the model's hidden state, thereby updating the model state and serving as the starting point for subsequent forecasts.

[0055] This invention proposes a coastal disaster forecasting method and system. The method fundamentally eliminates boundary fitting errors associated with regular grids by directly fitting unstructured grids to complex coastlines. Physical time-delay alignment based on mutual information enables the data-driven model to possess physical interpretability. The trained graph neural network operates in forward inference mode, achieving a computational speed several orders of magnitude faster than traditional numerical models, enabling minute-level high-resolution forecasts. By constructing a reference dynamic sample library integrating historical and virtual extreme scenarios, and utilizing physical numerical models and Copula statistical methods, massive, physically consistent training samples are generated. This allows the deep learning model to learn disaster patterns beyond historical records, significantly improving its forecasting generalization ability and robustness for rare extreme events. Employing a "hidden state assimilation" mechanism, the model's internal state can be instantaneously corrected using sparse real-time observation data, achieving online error absorption and blocking, with a response speed far faster than traditional correction methods that require model retraining. This provides highly timely and reliable dynamic forecasting support for disaster prevention and mitigation decision-making, significantly improving the emergency response capabilities of nearshore areas. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 This is a basic flowchart of a coastal disaster forecasting method provided in an embodiment of the present invention;

[0058] Figure 2 This is a schematic diagram of a triangular mesh in a numerical model provided in an embodiment of the present invention;

[0059] Figure 3 These are the verification results of the numerical model provided in the embodiments of the present invention; wherein, (a) is a schematic diagram of the verification results of astronomical tide; (b) is a schematic diagram of the verification results of storm surge; and (c) is a schematic diagram of the verification results of typhoon wave.

[0060] Figure 4 This is a flowchart of the model structure of the graph attention network provided in the embodiment of the present invention;

[0061] Figure 5The following is a comparison chart of disaster field prediction results provided in an embodiment of the present invention: (a) is a schematic diagram of the actual value at 20:00 on 2024-09-05; (b) is a schematic diagram of the predicted value at 20:00 on 2024-09-05; (c) is a schematic diagram of the actual value at 3:00 on 2024-09-08; and (d) is a schematic diagram of the predicted value at 3:00 on 2024-09-08.

[0062] Figure 6 These are the comparison results of the actual values ​​and predicted values ​​of the selected test sites provided in the embodiments of the present invention; wherein, (a) is the comparison result of the actual values ​​and predicted values ​​of site 1; and (b) is the comparison result of the actual values ​​and predicted values ​​of site 2. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0064] Example 1

[0065] like Figure 1 As shown, the present invention provides a method for forecasting coastal disasters, including:

[0066] S1: Acquire static geographic information, historical dynamic meteorological and oceanographic observation data, and multi-factor meteorological and hydrodynamic time-series data of offshore stations for the target nearshore area.

[0067] Specifically, the static geographic information includes water depth, shoreline vector data, seabed geological type, nearshore vegetation density, etc.

[0068] The meteorological and oceanographic observation data includes tide levels, significant wave heights, wave periods, storm surges, and average wave directions collected by tide gauge stations, wave buoys, and meteorological observation stations in the nearshore area of ​​the target.

[0069] The multi-factor meteorological and hydrodynamic time-series data of the offshore stations include dynamic observation data closely related to storm surges and typhoon waves from neighboring offshore stations within a preset distance range of the target nearshore area. Specifically, these data include significant wave height, wave direction, wind speed, wind direction, air pressure, water level, air temperature and sea surface temperature, swell height, and spectral peak period.

[0070] S2: Construct an unstructured grid covering the target nearshore area based on the static geographic information, construct a coupled numerical model of astronomical tide-storm surge-ocean wave based on the nearshore unstructured grid, and then generate a reference dynamic sample library containing the correspondence between offshore driving and nearshore response.

[0071] S2.1: Based on the high-precision shoreline and water depth obtained from the static geographic information, the target nearshore area is divided into unstructured grids using SMS software. In the specific division, the grid division details can be adjusted according to actual needs, such as performing grid densification in nearshore harbors, islands and protective engineering areas, and setting the grid step size to 30-50m.

[0072] S2.2: A coupled numerical model of astronomical tides, storm surges, and ocean waves is used for calculations. (The model can be constructed according to actual needs and is not limited, such as a coupled system based on the finite volume method ocean model FVCOM and the nearshore wave model SWAN.) Typical historical typhoon disaster events affecting the target nearshore area in recent years are selected as simulation objects, driven by the optimal typhoon path and reanalysis meteorological field during the event. Meteorological and oceanographic observation data collected in S1 (time series of measured tide levels at tide gauge stations and significant wave height data measured by wave buoys) are compared and verified with the corresponding time series simulated by the model. Key parameters in the model, such as bottom friction coefficient and wind drag coefficient, are iteratively adjusted until the root mean square error (RMSE) between the simulation results and the measured data is lower than the preset verification threshold (e.g., tide level error <10cm, significant wave height error <15%).

[0073] S2.3: Based on the meteorological and oceanographic observation data collected in S1, statistical analysis is performed and the marginal probability distributions of key disaster-causing factors such as wind speed, central pressure, and typhoon radius are fitted. A joint probability distribution function for these multiple factors is constructed using the Copula function. A large-scale random sampling (e.g., 1000 times) is conducted from this joint distribution using the Monte Carlo method to generate multiple sets of virtual disaster parameter combinations. Since numerical models cannot directly identify discrete parameters, this embodiment uses the Holland parameterized wind field model to transform each set of virtual parameters into a time-varying two-dimensional boundary covering the disaster field calculation of the target nearshore area. A high-performance computing cluster (DPC) is used for computation, and the boundary-driven numerical model outputs a spatiotemporally continuous two-dimensional grid field containing hydrodynamic variables such as wave height, tide level, and average wave period. Therefore, by integrating the simulation results of historical events and virtual scenarios from the numerical model, a reference dynamic sample library covering both normal and extreme conditions and containing a complete mapping relationship between "offshore driving (boundary conditions) – nearshore full-field response" can be constructed for subsequent deep learning model training. The specific steps are as follows:

[0074] S2.3.1: Extract the key disaster-causing factors that have historically impacted the nearshore area of ​​the target region from the data collected in step S1. Fit the marginal cumulative distribution functions of these factors using extreme value distribution functions. During the fitting process, construct the log-likelihood function using the maximum likelihood estimation method (MLE). Calculate the optimal solutions for the shape, scale, and location parameters of each distribution function using an optimization algorithm, thereby determining the mathematical expression of each factor. Finally, evaluate the fitting results using the Kolmogorov-Smirnov test to ensure that the constructed marginal distributions accurately reflect the extreme value statistical patterns of historical data.

[0075] S2.3.2: Considering that there may be significant physical characteristics among the various disaster-causing factors, a Copula function is introduced to connect the marginal distributions in step S2.3.1, construct a multi-factor joint probability distribution function, and estimate the structural parameters of the Copula function by calculating the correlation coefficient of historical data to form a joint probability model describing the multi-dimensional characteristics of typhoons.

[0076] S2.3.3, based on the constructed joint probability model, randomized experiments are conducted using the Monte Carlo method. By extracting probability combinations from the joint distribution and using the inverse function of the marginal part for inverse mapping, multiple sets of virtual disaster parameter combinations are generated. Finally, a physical parameterization model is introduced to transform each of the above discrete parameter combinations into spatiotemporally continuous two-dimensional data, which serves as the driving boundary conditions for the numerical model, thus expanding the dynamic sample library of the numerical model. Finally, these are integrated to obtain a reference dynamic sample library containing the correspondence between offshore driving and nearshore response.

[0077] S3: Perform time alignment, outlier removal, and standardization on meteorological and oceanographic observation data and multivariate meteorological and hydrodynamic time series data from nearby offshore stations.

[0078] S3.1: Since the data sampling frequency of offshore observation stations often differs from the output frequency of nearshore numerical simulations, this example uses a standardized time step to reconstruct time series that do not meet the standard time step (i.e., time series with low or non-uniform sampling frequencies), aligning all factors on the time axis. The calculation process for cubic spline interpolation is as follows:

[0079] Suppose the known time series data points are (t0, y0), (t1, y1), ... (t... n ,y n ), at any two adjacent times (t i ,t i+1 Construct a cubic polynomial function S between ) i (t):

[0080] ;

[0081] Among them, y n This represents time series data of meteorological or oceanographic variables over time; i is the time index, i=0,1,...,n-1; coefficient a i b i c i d i It can be determined by solving a set of simultaneous derivatives.

[0082] S3.2: For observation data from offshore stations, firstly, set the threshold intervals (V) for each factor collected by the offshore stations. min V max If an observation at a certain moment does not belong to the set, it is determined to be an outlier and removed. Then, for the missing sequences after removal, a time-series moving average method is used to complete them. The calculation process of the time-series moving average method is as follows: assuming the time point of the missing data is t, select valid observation data with a window size of k before and after this point, and calculate their arithmetic mean as the completed value.

[0083] S3.3: For different factors such as wind speed, air pressure, tide level, and wave height at nearshore and offshore stations, the mean and standard deviation are calculated separately for each channel. The Z-Score standardization method is used to map the data to a dimensionless distribution with zero mean and unit variance. The calculation process is as follows:

[0084] First, for the set of all time step data of the c-th channel, X = {X1, X2, ... X} n} Calculate its mean μ and standard deviation σ, and then use the calculated μ and σ to transform each data point of the channel to obtain the standardized data m:

[0085] ;

[0086] Here, ε represents a very small number to prevent division by zero.

[0087] S4: Using a dependency analysis method based on mutual information, the mutual information of driving factors at offshore sites to the responses of each grid node in the target nearshore area is calculated at different time lag steps. This determines the set of key driving factors and their dominant lag steps for each unstructured grid node. The specific process is as follows:

[0088] S4.1: Since marine environmental data typically follows a non-Gaussian distribution and is highly nonlinear, this study aims to quantify the nonlinear driving mechanism of offshore meteorological and hydrodynamic conditions on different nearshore regions. The hydrodynamic response sequence y of unstructured grid nodes g within the target nearshore region is then analyzed. g (t) and offshore site driving factor x j(e.g., wind speed, air pressure, sea surface temperature), the lag time of the driving factors at offshore stations is discretized into lag steps τ, and time-shifted sample pairs (x j (t-τ), y g (t)) The Parzen window method is used to estimate the marginal probability density and joint probability density of the offshore site driving factor and the nearshore grid node g response sequence (i.e., hydrodynamic response sequence), and then the offshore site driving factor x under the lag step τ is calculated. j The mutual information of the responses of unstructured grid nodes g within the target nearshore area is determined. Specifically, the Parzen window method is used to estimate the marginal probability density of the driving factors and the response sequences of nearshore grid nodes g, as well as their joint probability density.

[0089] First, the marginal probability densities of the offshore site driving factor sequence and the nearshore response sequence are estimated using the Gaussian kernel function, where the offshore driving factor x j and the hydrodynamic response sequence y of the nearshore grid node g g The formulas for calculating the marginal probability density function of (t) are as follows:

[0090] ;

[0091] ;

[0092] in, and All are marginal probability densities; N is the length of the sample sequence; h x and h y The smoothness of the estimation curve is determined by the window width parameters corresponding to the driving factors and response variables of the offshore sites. The Gaussian kernel function;

[0093] Secondly, to capture the nonlinear coupling characteristics between driving factors and nearshore responses, a joint probability density function is constructed. By superimposing kernel functions in two-dimensional space, a nonparametric estimation of the synchronous evolution probability between offshore driving signals and nearshore hydrodynamic responses is achieved. Joint probability density function The calculation formula is:

[0094] ;

[0095] in, This is the joint probability density function; the other variables are the same as the marginal probability density functions, and will not be described in detail here.

[0096] Finally, the driving factor x of offshore sites j Mutual information of the response of unstructured grid nodes g in the nearshore region of the target The calculation formula is as follows:

[0097] ;

[0098] in, This represents the current sequence of driving factors for offshore sites after the dominant lag shift. Represents the hydrodynamic time series of the response of an unstructured grid node g; and They represent and The marginal probability density; express and The joint probability density.

[0099] S4.2: To screen for lag times with real physical meaning, a Monte Carlo scrambling method is used for testing. Within a preset lag range that passes the test, the peak value of the mutual information curve is searched, and the time corresponding to this peak value is determined as the driving factor x of the offshore station. j The dominant hysteresis steps for unstructured mesh node g The specific process of the Monte Carlo scrambling method test is as follows:

[0100] While maintaining y g Assuming (t) remains constant, randomly shuffle x j The time sequence of (t) is determined and the mutual information is recalculated. Multiple trials are conducted to construct a null hypothesis sample, and a significance threshold is set (in practice, the significance threshold is set to 95%). Within the lag range that passes the significance test, the peak value of the mutual information curve is searched. The time corresponding to this peak value is the dominant lag step of the offshore driving factor xj on the unstructured grid node g. .

[0101] S4.3: After determining the dominant lag steps of the driving factors for all offshore stations, in order to eliminate the high redundancy among the offshore station data, the minimum redundancy maximum correlation criterion (mRMR) is used to select the set of key driving factors S for each unstructured grid node g. g Specifically: Initialize the key driving factor set S as an empty set, and adopt a forward search strategy. In each iteration, calculate each unselected offshore driving factor x. j Overall rating The comprehensive scores are sorted in descending order, and the overseas driving factors that meet the preset threshold or preset number of comprehensive scores are included in the key driving factor set.

[0102] Among them, the overall score The calculation formula is as follows:

[0103] ;

[0104] in, This represents the current sequence of driving factors for offshore sites after dominant lag shift; y g (t) represents the response sequence of the nearshore grid nodes; S represents the set of key driving factors that have been selected before the current iteration step; S represents the sequence of driving factors for a certain offshore site that already exists in set S; |S| represents the number of driving factors for offshore sites that have been selected in set S; I represents the mutual information calculation function.

[0105] S5: Based on the dominant lag steps of each driving factor in the key driving factor set, the offshore driving factors are time-rearranged and spliced ​​with nearshore static geographical features to construct a structured input tensor that integrates spatiotemporal information.

[0106] S5.1: For the current mesh node From the set of key driving factors Extracting the driving factors of each offshore site And obtain the dominant lag steps corresponding to the offshore driving factors. ;

[0107] S5.2: For each driving factor According to its dominant lag step number Shift the corresponding dominant lag step on the time axis. Obtain the aligned timing sequence This aligns the historical driving signals from the open sea with the current response signals from the nearshore area on a discrete index.

[0108] S5.3: Time series of driving factors for each aligned offshore site With nodes Static geographic feature vectors Perform vector concatenation to generate the node at time [time]. Structured input tensors :

[0109] ;

[0110] Where [ ] denotes vector concatenation operation; This indicates iterating through all offshore site driving factors in set Sg and taking their values ​​at discrete indices. Observations at; This indicates the grid node. The static geographic feature vectors include water depth and bottom friction coefficient, etc.

[0111] S6: Using a structured input tensor that integrates spatiotemporal information as input, and the nearshore response data (i.e. nearshore two-dimensional disaster field data) generated by the corresponding astronomical tide-storm surge-wave coupled numerical model in the reference dynamic sample library as labels, the spatiotemporal fusion deep learning model based on graph attention network is trained under supervision to obtain the coastal disaster forecasting model.

[0112] The spatiotemporal fusion deep learning model of the graph attention network processes each training sample as follows:

[0113] S6.1: Structured input tensors that incorporate spatiotemporal information Through a linear transformation layer and activation function, it is mapped to a high-dimensional node-driven feature vector. .

[0114] More specifically, the input feature vector V is processed using the weight matrix W of the input layer. g Feature encoding is performed, physically mapping from the low-dimensional space to the high-dimensional feature space. A bias vector b is then added to adjust the activation threshold. Subsequently, the input is processed by a leaky rectified linear activation function (LeakyRELU), which preserves a smaller gradient when the input is negative. Finally, the output is a high-dimensional externally driven feature vector E with non-linear expressive power. g The calculation formula is as follows:

[0115] ;

[0116] Where W represents the learnable weight matrix of the input layer; LeakyRELU represents the leaky linear rectified activation function; and b represents the bias vector.

[0117] S6.2: Using an unstructured mesh as the graph structure, compute the central nodes using the graph attention mechanism. Its first-order neighbor set Each neighbor node attention weights This leads to the understanding of the historical implicit states of neighboring nodes. Perform a weighted summation to generate the spatial context feature vector of the central node. ;

[0118] More specifically, using the linear transformation matrix W s The latent states of the center node g and neighbor node k are transformed separately to obtain the center node features and neighbor node features, respectively. These features are then concatenated along the channel dimension to form node-pair features. Next, a dot product operation is performed on the node features using the learnable weight vector β to obtain the original attention coefficients. These coefficients are then processed by a normalized exponential function (softmax function) to transform them into propagation weights α in probability distribution form. Finally, the value projection matrix W is used...v The latent state of neighbor node k is transformed again. Then, the transformed features of all neighbor nodes are weighted and summed according to the calculated propagation weight α. The accumulated result is recorded as the comprehensive spatial context feature M of the surrounding water medium on the current node. g Among them, attention weights Spatial context feature vector of the central node The calculation formulas are as follows:

[0119] ;

[0120] ;

[0121] Among them, W s The linear feature transformation matrix used to map the hidden states of nodes to a high-dimensional attention space; β represents the learnable weight vector used to map the concatenated node pair features to scalar attention scores; W v represents the value projection matrix used to perform message transformation on the features of neighboring nodes; T represents the transpose operation; || represents the vector concatenation operation; softmax is the normalization exponential function.

[0122] S6.3: External driving features Spatial context features The data is concatenated and input into a gated recurrent unit (GRU); the GRU then combines this data with the hidden state of the node from the previous time step. Calculate and output the new hidden state at the current time. .

[0123] More specifically, the external driving feature vector E output by S6.1 g Spatial context features M g The features are concatenated to form the total momentum input vector at the current moment. This vector, along with the node's own historical latent state h(t-1), is then simultaneously input into the gated recurrent unit (GRU) to compute the two gated signals and update the gate z in parallel. t With Reset Gate r t , where z t As a balancing coefficient, it is used for weighted interpolation between "preserving the historical state from the previous time step" and "adopting the candidate state from the current time step"; r t It operates on historical states to determine how much accumulated information from the past is forgotten, and combines it with the current input, generating candidate hidden states through the Tanh activation function.

[0124] S6.4: Update the implicit state Through an output layer (linear transformation layer), the value of the predicted physical quantity (such as water level) of the node at the current moment is decoded; by traversing all grid nodes, the predicted nearshore two-dimensional disaster field at the current moment is obtained.

[0125] More specifically, through the aforementioned adjustment mechanism, the new hidden state of the node at the current moment is calculated. Then, a new computation process is executed in parallel: on the one hand, by outputting the weight matrix W of the mapping layer. out and bias b out A linear transformation is performed to decode it into a specific physical prediction scalar y, which is used to construct a two-dimensional disaster forecast field; on the other hand, it is directly passed as a recursive state to the next time step, serving as the historical input at time t+1. The overall calculation process is as follows:

[0126] ;

[0127] ;

[0128] Among them, z t and r t These represent the update gate and reset gate coefficient vectors generated by the Sigmoid function, respectively; ⊙ represents the Hadamard product operation; W h W represents the candidate state weight matrix; out b represents the output layer weights. out This represents the output layer bias vector.

[0129] S7: Acquire real-time meteorological and oceanographic observation data from nearshore stations and real-time multi-factor meteorological and hydrodynamic time-series data from offshore stations. Process the offshore station data into real-time structured input tensors using S3-S5 and input them into the pre-trained nearshore disaster forecasting model for rolling forecasts. Simultaneously, reconstruct the error field using the current real-time meteorological and oceanographic observation data from the nearshore area and map it as the correction increment of the model's hidden state, thereby updating the model state and serving as the starting point for subsequent forecasts. The specific process is as follows:

[0130] S7.1: Acquire real-time marine observation data from the target nearshore station and real-time multi-factor meteorological and hydrodynamic time-series data from adjacent offshore stations. Process the offshore station data into real-time structured input tensors using S3-S5 and input them into the pre-trained nearshore disaster forecasting model for rolling forecasts. Simultaneously, using the real-time meteorological and marine observation data of the target nearshore area at the current moment, calculate the deviation vector between the measured value and the model's forecast value at the corresponding station location.

[0131] S7.2: Calculate the norm of the deviation vector. If the norm of the deviation vector is less than a predetermined threshold, the original forecast field is directly published. If it is greater than the preset error threshold, the inverse distance weighted spatial interpolation algorithm is used to interpolate the local deviation with the observation station k as the center and extend it to all nodes of the unstructured computing grid according to its spatial distance weight, thereby reconstructing a global error field with spatial continuity.

[0132] ;

[0133] in, Represents the global error field; Indicates the number of nearshore observation stations; This represents the measured value at the observation station. This represents the predicted value indicating the location of the model's observation stations; This represents spatial weight; the greater the distance, the smaller the impact of the correction bias.

[0134] S7.3: Input the error field into a pre-trained feedback encoder network, which maps the spatial error field into a state correction increment vector of the same dimension as the model's latent states. This instantly updates the current hidden state of all nodes in the model. The updated formula is as follows:

[0135] ;

[0136] in, It is a relaxation factor; This is the updated implicit state.

[0137] S7.4: Update the implicit state Replace the original Starting from the updated assimilated state, the model decoder is re-executed or the calculation continues until the next forecast time. The corrected forecast field for the current time is output, and continuous rolling forecasts are carried out based on this, thereby effectively blocking the cumulative propagation of forecast errors on the time axis.

[0138] Example 2

[0139] Secondly, the present invention provides a coastal disaster forecasting system, the system being used to perform the method described above, including:

[0140] Data acquisition and preprocessing module: Acquires static geographic information of the target nearshore area, historical dynamic meteorological and oceanographic observation data, and multi-factor meteorological and hydrodynamic time series data of offshore stations; used to perform time alignment, outlier removal, and standardization processing on the dynamic meteorological and oceanographic observation data in the reference dynamic sample library and the multi-factor meteorological and hydrodynamic time series data of offshore stations.

[0141] Numerical model calculation module: used to construct an unstructured grid covering the target nearshore area based on the static geographic information, construct a coupled numerical model of astronomical tide-storm surge-wave based on the nearshore unstructured grid, and then generate a reference dynamic sample library containing the correspondence between offshore driving and nearshore response;

[0142] Lag factor identification module: Using a dependency analysis method based on mutual information, the mutual information of the driving factors of offshore stations to the response of each grid node in the target nearshore area is calculated at different time lag steps, and the set of key driving factors and their dominant lag steps corresponding to each unstructured grid node are determined.

[0143] Structured Input Tensor Construction Module: Based on the dominant lag steps of each driving factor in the set of key driving factors, the offshore driving factors are time-rearranged and spliced ​​with nearshore static geographic features to construct a structured input tensor that integrates spatiotemporal information.

[0144] The graph neural network model inference and online assimilation and correction module is used to take a structured input tensor as input and use the nearshore response (nearshore two-dimensional disaster field data) generated by the corresponding astronomical tide-storm surge-wave coupled numerical model and numerical model simulation in the reference dynamic sample library as a label to supervise the training of the spatiotemporal fusion deep learning model based on graph attention network, so as to obtain the coastal disaster forecasting model. It is used to acquire real-time meteorological and oceanographic observation data of the target nearshore station and real-time multi-factor meteorological and hydrodynamic time series data of offshore station, process the offshore station data into real-time structured input tensor and input it into the trained nearshore disaster forecasting model for rolling forecasts. At the same time, it reconstructs the error field using the real-time meteorological and oceanographic observation data of the target nearshore at the current moment and maps it into the correction increment of the model's hidden state, thereby updating the model state and serving as the starting point for subsequent forecasts.

[0145] Example 3

[0146] This embodiment provides a readable storage medium storing a computer program that, when invoked by a processor, performs the steps of the method described above.

[0147] Example 4

[0148] This embodiment provides an electronic terminal, including a processor and a memory, wherein the memory stores a computer program, and the processor calls the computer program to perform the steps of the method described above.

[0149] It should be understood that, in the embodiments of the present invention, the processor may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor. The memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of the memory may also include non-volatile random access memory. For example, the memory may also store device type information.

[0150] The readable storage medium is a computer-readable storage medium, which can be an internal storage unit of the controller described in any of the foregoing embodiments, such as the controller's hard drive or memory. The readable storage medium can also be an external storage device of the controller, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the controller. Further, the readable storage medium can include both the controller's internal storage unit and external storage devices. The readable storage medium is used to store the computer program and other programs and data required by the controller. The readable storage medium can also be used to temporarily store data that has been output or will be output.

[0151] Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned readable storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0152] It is understood that the same or similar parts in the above embodiments can be referred to each other, and the contents not described in detail in some embodiments can be referred to the same or similar contents in other embodiments.

[0153] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

[0154] Experimental verification

[0155] The following are practical applications of the present invention, which can prove the effectiveness of the invention.

[0156] This embodiment selects a specific location as the target area. The numerical model adopts a coupled model of FVCOM and SWAN. The example uses Super Typhoon Rammasun, which occurred in this region from 00:00 on July 10, 2014 to 06:00 on July 20, 2014 (UTC). An unstructured triangular mesh is constructed using SMS software, and the resolution is refined to 200m within the strait and port areas, as shown in the attached figure. Figure 2 As shown. Three tide gauge stations and wave stations were selected in the open sea to verify astronomical tides, storm surges, and significant wave heights. The results are as follows. Figure 3 As shown, this indicates that the model can reflect astronomical tides (such as...) very well. Figure 3 (a) Storm surge (e.g.) Figure 3 (b) and typhoon waves (such as Figure 3 (c)) The variation characteristics. In addition, this experiment also used the Copula-Monte Carlo method to generate 800 virtual extreme typhoon paths, and used the Delft3D-FLOW-WAVE coupled model with the Holland parametric wind field model as the boundary condition for calculation to establish a sufficient database of dynamic responses to disasters within the bay.

[0157] Offshore stations selected grid nodes from a meteorological center's reanalysis dataset as driving forces to capture early prior information before a typhoon enters a specific area. A uniform 1-hour resampling and Z-score standardization process was performed on the collected offshore buoy data and nearshore disaster field grid data. Subsequently, for each nearshore grid node, the dominant lag steps of each driving factor (wind speed, wind direction, wave period, air pressure, etc.) of the offshore buoy station were calculated, followed by physical time translation and concatenation with the static water depth attributes of each grid node to construct a structured input tensor. This tensor was then input into the model for training, and the model structure is as follows. Figure 4As shown, the AdamW optimizer was used during training, with an initial learning rate of 1e-3 and a batch size of 32. Training was conducted for 500 epochs on a single NVIDIA 4090 GPU. The final model evaluation used root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). 2 An evaluation will be conducted.

[0158] To demonstrate the model's prediction results, the entire process of a typhoon entering the site (14:00 on September 1, 2024 to 14:00 on September 8, 2024) was selected for verification. The overall disaster field prediction results are attached. Figure 5 As shown in the figure, the two-dimensional prediction field of the model of this invention has high consistency in spatial distribution. Furthermore, due to the topological structure of the unstructured grid, the prediction field exhibits good prediction results at the tortuous shoreline boundary, with low RMSE, MAE, and R... 2 The values ​​were 0.187m, 0.134m, and 0.983, respectively. Furthermore, to further demonstrate the model's early warning capability in nearshore disaster areas, the predicted time series data from two nearshore stations located in the grid-densified area were extracted and compared with the actual values. The verification results are attached. Figure 6 As shown, where, Figure 6 (a) Comparison of actual and predicted values ​​for site 1 Figure 6 (b) Comparison of actual and predicted values ​​at station 2. As can be seen from the figure, the model has good prediction results in different regions of the high-resolution grid. Even when the wave height is greater than 4m, the maximum error does not exceed 0.30m.

Claims

1. A method for forecasting coastal disasters, characterized in that, include: S1: Acquire static geographic information, historical dynamic meteorological and oceanographic observation data, and multi-factor meteorological and hydrodynamic time-series data of offshore stations for the target nearshore area; S2: Construct an unstructured grid of the coverage area based on static geographic information, construct a coupled numerical model of astronomical tide-storm surge-wave based on the nearshore unstructured grid, and generate a sample library containing the correspondence between offshore driving and nearshore response. S3: Preprocess the meteorological and oceanographic observation data in the sample library with the multi-factor meteorological and hydrodynamic time series data from offshore stations; S4: Using a dependency analysis method based on mutual information, calculate the mutual information of the response of the driving factors of offshore stations to each grid node in the region under different time lag steps, and determine the set of key driving factors and their dominant lag steps for each grid node. S4.1: Response sequence for unstructured grid node g in the target nearshore area and offshore site driving factors The lag time of the driving factors of offshore sites is discretized into lag steps τ, and time-shifted sample pairs are constructed. The Parzen window method is used to estimate the marginal probability density of the response sequence and the driving factors of offshore sites, as well as their joint probability density. Then, the driving factors of offshore sites under the lag step τ are calculated. Mutual information of the response of unstructured grid nodes g in the nearshore area of ​​the target; S4.2: The Monte Carlo scrambling method is used for testing, and within the preset lag range that passes the test, the peak value of the mutual information curve is searched, and the time corresponding to the peak value is determined as the driving factor of the offshore station. The dominant hysteresis steps for unstructured mesh node g ; S4.3: After determining the dominant lag steps of the driving factors for all offshore sites, the minimum redundancy maximum correlation criterion is used to select the set of key driving factors for each grid node g. ; S5: Based on the dominant lag step of each driving factor in the key driving factor set, the driving factors of offshore stations are time-rearranged and spliced ​​with nearshore static geographical features to construct a structured input tensor that integrates spatiotemporal information. S6: Using the structured input tensor as input and the nearshore response data generated by the corresponding numerical model simulation in the sample library as labels, supervise the training of the spatiotemporal fusion deep learning model based on graph attention network to obtain the coastal disaster prediction model. S7: Acquire real-time meteorological and oceanographic observation data from regional stations and multi-factor meteorological and hydrodynamic time-series data from offshore stations. Process the offshore station data into real-time structured input tensors and input them into the pre-trained nearshore disaster forecasting model for rolling forecasts to obtain the nearshore two-dimensional disaster field at the current forecast time. At the same time, reconstruct the error field using the real-time meteorological and oceanographic observation data of the region at the current time and map it into the correction increment of the model's hidden state to update the model state and serve as the starting point for subsequent forecasts.

2. The method according to claim 1, characterized in that, In S1, the static geographic information includes water depth, shoreline vector, and seabed type; the meteorological and oceanographic observation data includes tide level, significant wave height, wave period, storm surge, and average wave direction collected by tide gauge stations, wave buoys, and meteorological observation stations in the target nearshore area; the multi-factor time-series data of the offshore stations includes significant wave height, wave direction, wind speed, wind direction, air pressure, water level, air temperature and sea surface temperature, swell height, and spectral peak period.

3. The method according to claim 1, characterized in that, The specific process of S2 is as follows: S2.1: Based on the shoreline and water depth in the acquired static geographic information, the target nearshore area is divided into unstructured grids using SMS software; S2.2: A coupled numerical model of astronomical tide, storm surge, and ocean waves is used for simulation calculation. Typical historical typhoon disaster events affecting the nearshore area of ​​the target within a preset period are selected as simulation objects. The optimal typhoon path and reanalysis meteorological field during the event are used as driving forces. The meteorological and oceanographic observation data collected in S1 are compared and verified with the corresponding time series simulated by the model. The key parameters of the model are adjusted iteratively until the error between the simulation results and the measured data is lower than the preset verification threshold. S2.3: Based on the meteorological and oceanographic observation data collected in S1, statistical analysis is performed and the marginal probability distributions of each factor are fitted separately; the joint probability distribution function of each factor is constructed using the Copula function; random sampling is performed from the joint distribution using the Monte Carlo method to generate multiple sets of virtual disaster parameter combinations; and each set of virtual parameters is transformed into a time-varying two-dimensional boundary for calculating the disaster field covering the target nearshore area, thereby driving the astronomical tide-storm surge-wave coupled numerical model to output the spatiotemporal continuous two-dimensional grid field of the driving factors of the offshore station; and then the reference dynamic sample library of offshore driving-nearshore full-field response-mapping relationship is obtained.

4. The method according to claim 1, characterized in that, Offshore site driving factors Mutual information of the response of unstructured grid nodes g in the nearshore region of the target The calculation formula is as follows: ； in, This represents the current sequence of driving factors for offshore sites after the dominant lag shift. Represents the response sequence of an unstructured mesh node g; and They represent and The marginal probability density; express and The joint probability density.

5. The method according to claim 1, characterized in that, The process of selecting the set of key driving factors is as follows: Initialize the set of key driving factors. Given an empty set, a forward search strategy is employed. In each iteration, the driving factor for each offshore site that has not yet been selected is calculated. Overall rating The comprehensive scores are sorted in descending order, and the driving factors of offshore stations that meet the preset threshold or preset number of comprehensive scores are included in the key driving factor set. Overall score The calculation formula is as follows: ； in, This represents the current sequence of driving factors for offshore sites after dominant lag shift; y g (t) represents the response sequence of the nearshore grid node; This represents the set of key driving factors that have been selected before the current iteration step; Represents a set A sequence of driving factors for a certain offshore site already existing in China; | | represents a set The number of driving factors for offshore sites already selected; I represents the mutual information calculation function.

6. The method according to claim 1, characterized in that, The specific process of S5 is as follows: S5.1: For the current mesh node From the set of key driving factors Extracting the driving factors of each offshore site And obtain the dominant lag steps corresponding to the driving factors of offshore sites. ; S5.2: For each driving factor According to its dominant lag step number Shift the corresponding dominant lag step on the time axis. Obtain the aligned timing sequence This aligns the historical driving signals from the open sea with the current response signals from the nearshore area on a discrete index. S5.3: Time series of driving factors for each aligned offshore site With nodes Static geographic feature vectors Perform vector concatenation to generate the node at time [time]. Structured input tensors : ； Where [ ] denotes vector concatenation operation; Indicates traversing a collection The driving factors of all offshore stations are taken from the discrete index. The observation value at that location; cg represents the grid node. The static geographic feature vector.

7. The method according to claim 1, characterized in that, The spatiotemporal fusion deep learning model of the S6 graph attention network processes each training sample as follows: S6.1: Structured input tensors that incorporate spatiotemporal information Through a linear transformation layer and activation function, it is mapped to a high-dimensional node-driven feature vector. ; S6.2: Using an unstructured mesh as the graph structure, compute the central nodes using the graph attention mechanism. Its first-order neighbor set Each neighbor node attention weights This leads to the understanding of the historical implicit states of neighboring nodes. Perform a weighted summation to generate the spatial context feature vector of the central node. ; S6.3: Features of offshore driving Spatial context features The data is concatenated and input into the gated loop unit; the gated loop unit combines the implicit state of the node at the previous time step. Calculate and output the new hidden state at the current time. ; S6.4: Update the implicit state Through an output layer, the predicted physical quantity scalar value of the node at the current moment is decoded; by traversing all grid nodes, the predicted nearshore two-dimensional disaster field at the current moment is obtained.

8. The method according to claim 1, characterized in that, The specific process of S7 is as follows: S7.1: Acquire real-time meteorological and oceanographic observation data from nearshore stations and real-time multi-factor meteorological and hydrodynamic time-series data from offshore stations. Process the offshore station data into real-time structured input tensors using S3-S5 and input them into the pre-trained nearshore disaster forecasting model for rolling forecasting to obtain the nearshore two-dimensional disaster field at the current forecast time. At the same time, using the real-time meteorological and oceanographic observation data of the target nearshore area at the current time, calculate the deviation vector between the measured value and the model's forecast value at the corresponding station location. S7.2: If the norm of the deviation vector is less than a predetermined threshold, the near-shore two-dimensional disaster field at the current forecast time is directly output; if it is greater than a preset error threshold, the inverse distance weighted spatial interpolation algorithm is used to calculate the uncorrected increment Er of each grid node g in the entire field, centered on the observation station k, to generate an error field with all near-shore grid nodes; wherein, the formula for calculating the uncorrected increment Er is: ； in, Indicates the number of nearshore observation stations; Represents the measured values ​​at the observation station; This represents the predicted value indicating the location of the model's observation stations; Indicates spatial weights; S7.3: Input the error field into a pre-trained feedback encoder network, which maps the spatial error field into a state correction increment vector of the same dimension as the model's latent states. This instantly updates the current hidden state of all nodes in the model; S7.4: Update the implicit state Replace the original Starting from the updated state, the model decoder is re-executed or the calculation continues until the next forecast time, outputting the corrected forecast field for the current time, and then continuous rolling forecasts are carried out for subsequent times.

9. A coastal disaster forecasting system, said system being used to perform the method as described in any one of claims 1-8, characterized in that, include: Data acquisition and preprocessing module: Acquires static geographic information of the target nearshore area, historical dynamic meteorological and oceanographic observation data, and multi-factor meteorological and hydrodynamic time series data of offshore stations; used to perform time alignment, outlier removal, and standardization processing on the dynamic meteorological and oceanographic observation data in the reference dynamic sample library and the multi-factor meteorological and hydrodynamic time series data of offshore stations. Numerical model calculation module: used to construct an unstructured grid covering the target nearshore area based on the static geographic information, construct a coupled numerical model of astronomical tide-storm surge-wave based on the nearshore unstructured grid, and then generate a reference dynamic sample library containing the correspondence between offshore driving and nearshore response; The lag factor identification module uses a dependency analysis method based on mutual information to calculate the mutual information of the response of the driving factors of offshore stations to each grid node in the target nearshore area under different time lag steps, and determines the set of key driving factors and its dominant lag step for each grid node. Structured Input Tensor Construction Module: Based on the dominant lag step of each driving factor in the set of key driving factors, the driving factors of offshore stations are time-rearranged and spliced ​​with nearshore static geographical features to construct a structured input tensor that integrates spatiotemporal information. The graph neural network model inference and online assimilation and correction module is used to take the structured input tensor as input and the nearshore response data generated by the corresponding astronomical tide-storm surge-wave coupled numerical model in the reference dynamic sample library as labels to supervise the training of the spatiotemporal fusion deep learning model based on graph attention network, so as to obtain the coastal disaster forecast model. This method is used to acquire real-time meteorological and oceanographic observation data from nearshore stations and real-time multi-factor meteorological and hydrodynamic time-series data from offshore stations. The offshore station data is processed into real-time structured input tensors and input into a pre-trained nearshore disaster forecasting model for rolling forecasts, resulting in a nearshore two-dimensional disaster field at the current forecast time. Simultaneously, the error field is reconstructed using the real-time meteorological and oceanographic observation data of the target nearshore area at the current time, and mapped as a correction increment of the model's implicit state, thereby updating the model state and serving as the starting point for subsequent forecasts.

Images